Nanophotonics 2020; 9(13): 4075–4095

Review

Sajjad Abdollahramezani, Omid Hemmatyar and Ali Adibi*

Meta-optics for spatial optical analog computing

https://doi.org/10.1515/nanoph-2020-0285 1 Introduction
Received May 13, 2020; accepted August 25, 2020; published online
September 7, 2020
For decades, optics has proved to be an unrivaled means of
communication between two points as close as two chip-
Abstract: Rapidly growing demands for high-performance
scale modules or as far as two intercontinental data cen-
computing, powerful data processing, and big data necessi-
ters. To a similar extent, optical computing has been
tate theadvent ofnoveloptical devicesto perform demanding
overshadowed by powerful digital computers empowered
computing processes effectively. Due to its unprecedented
by highly integrable transistors as mature electronic
growth in the past two decades, the field of meta-optics offers
switches. However, when it comes to intricate problems
a viable solution for spatially, spectrally, and/or even
such as optimization dealing with very large data sets,
temporally sculpting amplitude, phase, polarization, and/or
even breakthrough application-specific integrated circuit
dispersion of optical wavefronts. In this review, we discuss
technologies such as a tensor processing unit have
state-of-the-art developments, as well as emerging trends, in
shown substantial limitations in practice [1]. All-optical
computational metastructures as disruptive platforms for
computing holds the promise to solve challenging prob-
spatial optical analog computation. Two fundamental ap-
lems such as large-scale combinatorial optimizations [2, 3]
proaches based on general concepts of spatial Fourier trans-
and probabilistic graphical models [4], which have found
formation and Green’s function (GF) are discussed in detail.
many practical applications in artificial intelligence, image
Moreover, numerical investigations and experimental dem-
processing, and social networks.
onstrations of computational optical surfaces and meta-
Optical computing generally refers to the numerical
structures for solving a diverse set of mathematical problems
computation of one-dimensional or multidimensional data
(e.g., integrodifferentiation and convolution equations)
using photons as the primary carrier medium. Optical
necessary for on-demand information processing (e.g., edge
computing has been an active field of research for over 60
detection) are reviewed. Finally, we explore the current
years with impressive achievements in analog/digital and
challenges and the potential resolutions in computational
classic/quantum information processing that leverage
meta-optics followed by our perspective on future research
various computational approaches including Turing ma-
directions and possible developments in this promising area.
chines, brain-inspired neuromorphic architectures, and
Keywords: edge sensing; Fourier transformation; Green’s metaphoric systems. Optical computing as a serious rival
function; metasurfaces; optical analog computing. has experienced ups and downs in different time periods [5–
7]. This stems from the indisputable superiority of optical
technology in some important aspects including inherent
parallel processing, low cross talk, passive components
with zero static energy consumption, and high space and
time bandwidth products that potentially alleviate the
inherent shortcomings of digital electronics, namely speed,
heat generation, and power hungriness [7–9]. Although
Sajjad Abdollahramezani and Omid Hemmatyar contributed equally early attempts have been devoted to digital optical
to this article.
computing by mimicking the principles of electronic com-
*Corresponding author: Ali Adibi, School of Electrical and Computer puters, people found the eventual goal out of reach due to
Engineering, Georgia Institute of Technology, 778 Atlantic Drive NW, the lack of small-footprint and energy-efficient nonlinear
Atlanta, GA 30332-0250, USA, E-mail: [email protected] elements working as functional optical switches [10]. In this
Sajjad Abdollahramezani and Omid Hemmatyar, School of Electrical regard, optical analog computing has gained more attrac-
and Computer Engineering, Georgia Institute of Technology, 778
tion leading to more striking developments in recent years.
Atlantic Drive NW, Atlanta, GA 30332-0250, USA,
E-mail: [email protected] (S. Abdollahramezani), Generally, analog optical computing can be classified into
[email protected] (O. Hemmatyar) (i) linear platforms enabling space-invariant operations

Open Access. © 2020 Sajjad Abdollahramezani et al., published by De Gruyter. This work is licensed under the Creative Commons Attribution
4.0 International License.
4076 S. Abdollahramezani et al.: Meta-optics for spatial optical analog computing

(such as correlation and convolution), as well as space- miniaturized configurations. As the most interesting
variant operations (such as Hough or coordinates trans- paradigm, metastructures (including both metamaterials
form), and (ii) nonlinear systems enabling pivotal opera- and metasurfaces [12–33]) hold great promise to imprint
tions such as logarithm transformation, as well as the desired transformations in amplitude, phase, and
thresholding [5, 11]. Due to the vast diversity of achieve- polarization of the impinging light thanks to the (sub)
ments in each field, we limit the scope of this review to the wavelength-scale scatterers with optimized size, shape,
spatial optical analog computing within the linear shift- orientation, and composition.
invariant (LSI) systems. Motivated by such technological developments,
Over the past six decades, spatial analog optical Engheta et al. recently introduced the concept of
computing has been arguably indebted to the progress in “computational metamaterials” [34]. These structures
game-changing technologies such as diffractive and perform mathematical operations by direct manipulation
refractive optical elements, holograms, spatial light of propagating light waves through judiciously designed
modulators (SLMs), and microelectromechanical mirrors metamaterial and metasurface platforms. This work
[11]. However, computing systems established based on inspired an ever-increasing interest to propose and
those components are complex, challenging to align, and demonstrate viable solutions to expand the horizon of
bulky in size, hindering integration with compact nano- small-footprint computational systems for on-demand
photonic circuits. Recent extraordinary advances in applications ranging from complex mathematical opera-
nanofabrication have pushed the thickness limit of tions to real-time edge detection and large-scale image
various optical elements deep into the nanometric scale, processing to machine vision. Two fundamental ap-
enabling stronger interaction of light–matter necessary proaches have been pursued so far to realize computa-
for full control over the properties of incident light in tional meta-optics: (i) spatial Fourier transfer approach,

A Spatial Fourier Transfer Approach B Green’s Function Approach

Metalens
ℱ

Metasurface Metasurface
H(kx) H(kx, ky)

Metalens
ℱ
First-order spatial differentiator Second-order spatial differentiator

Figure 1: Computational meta-optics. Conceptual representation of spatial analog computing based on general concepts of (A) spatial Fourier
transfer and (B) Green’s function (GF) approach. The former leverages a 4f correlator system formed by two series lenses with equal focal
lengths and an intermediate complex-valued spatial filter located at the shared focal plane. The first metalens transforms the incident light
with a spatially variant profile (Gaussian in this example) to the Fourier domain (i.e., k-space), while the second metalens transfers back the
light processed by the intermediate metasurface to the spatial domain. The transfer function (i.e., H(kx)) associated with the kernel of the
operator of choice, one-dimensional (1D) first-order differentiation in this example, is encoded into the metasurface. The latter takes
advantage of the nonlocal response of an engineered single-layer optical surface or a multilayered slab that performs the desired kernel of
choice. The computational system in (B) performs the second-order spatial derivation for two-dimensional (2D) edge detection.
 S. Abdollahramezani et al.: Meta-optics for spatial optical analog computing 4077

in which a 4f system incorporating a metascreen for powerful concept can be applied to a metaprocessor given
realization of the desired transfer function is employed, that f (x, y) and g (x, y) are the field profiles of the incident
and (ii) Green’s function (GF) approach, in which a and transmitted/reflected optical beams, respectively. In
resonant or nonresonant metascreen is utilized to realize this regard, the transfer function of the system, i.e., H (kx,
the desired spatial impulse response associated with the ky), associated with the desired mathematical operator of
mathematical operator of choice (see conceptual images choice can be implemented by employing a spatially
in Figure 1). In the following sections, we first elaborate variant subblock with position-dependent transmission/
basic principles of each approach and then discuss rele- reflection coefficient. It is noteworthy that the realization
vant recent developments. These new paradigms offer of an inverse Fourier transform subblock with natural
real-time spatial wave-based processing mechanisms materials is not practically feasible. Based on the well-
through miniaturized all-optical computing machines or known relation F {F {A(x, y)}} ∝ A(−x, −y), a Fourier lens
potentially integratable hardware accelerators. can be used instead to perform inverse Fourier transform
operation at the expense of image mirroring of the desired
output [34].
Many scientific phenomena and engineering problems
2 Spatial Fourier transfer approach
such as heat transfer, electromagnetic radiation, Kirchh-
This approach has been investigated for decades in 4f off’s second law in circuit analysis can be described
correlators using bulky optical components [11]. Inspired through a system of partial differential, integral, and
by conventional 4f systems, the generic architecture of an integrodifferential equations. As such, finding the exact
optical metaprocessor is formed by the integration of a solution to a given set has been the main focus of many
planar metamaterial or cascaded transmit/reflectarray of works in the field of analog computing. To this end, we
suitably structured metasurfaces with the Fourier trans- centralize the rest of discussion around the solution of
form subblocks (such as metalenses, thin lenses, or these types of problems. Based on Fourier transform prin-
graded-index [GRIN] media). The conceptual representa- ciples, the nth derivative of a 2D wavefunction, to which
tion of a metaprocessor comprising an intermediate an arbitrary function is decomposed, is linked to the cor-
metasurface that realizes the desired transfer function responding wavenumber and the first spatial Fourier
associated with the mathematical operator of choice (here transform according to the following equation [11]:

∇n f
x, y
F −1 (ik)n F f
x, y

first-order spatial differentiation) and a pair of metalenses
(3)
performing the exact Fourier transform of the one-
√ 
dimensional (1D) field profile at their back focal planes is in which i
−1 and k
k 2x + k 2y . Given the continuum of
shown in Figure 1A. Such a compelling framework can wavenumbers forming the basis of field expansion, the
significantly shrink the overall size of traditional bulky 4f transfer function corresponding to the nth derivation fol-
systems. lows a parabola profile with the order of n. Without loss of
Considering an LSI system, for any given input func- generality, we limit the problem to more practical optical
tion f (x, y), the output function of the system g (x, y) is computing problems dealing with one transverse variable.
calculated according to the convolution relation defined as We assume the continuum of wavenumbers is called x that
follows [34]: represents wavenumber, and not the spatial variable,
g
x, y
h
x, y∗f
x, y
along the x-axis. Thus, the desirable transfer function is in
the form of H(x) ∝ (ix)n . Since the transmission coefficient

∬h
x − x′, y − y′f
x′, y′dx ′ dy ′ , (1) cannot surpass unity in the passive (gainless) media, the
in which the two-dimensional (2D) spatial impulse transfer function must be normalized to the lateral size of
response of the system is h(x, y) and ∗ stands for the linear the structure D, i.e., H(x) ∝ (ix/(D/2))n . For the integration
convolution operation. Equation (1) can be represented in operation, the required transfer function is defined as
the spatial Fourier domain as follows [34]: H(x) ∝ (d/ix) in the spatial Fourier domain with a singu-
larity at x = 0, which has to be handled. It has been proven
g
x, y
F −1
H
k x , k y F f
x, y, (2)
that imposing constant value of unity on H(x) in the small
where (kx, ky) denotes the 2D spatial frequency variables in region d (typically an order of magnitude less than D) near
the Fourier space, H (kx, ky) is the spatial Fourier transform the origin mitigates well any need for gain requirements
of h (x, y), and (F −1 { ⋅ })F { ⋅ } represents the (inverse) [34]. Upon calculation of the transfer function of the
Fourier transform operator. Without loss of generality, this desired operator of choice, the rest is dealing with the
4078 S. Abdollahramezani et al.: Meta-optics for spatial optical analog computing

Fourier transform (GRIN(+)) and inverse Fourier trans-

form (GRIN(−)) subblocks, 2D GRIN media characterized
by length Lg, unit relative permeability (i.e., μr = ±1), and
parabolic-shaped permittivity, defined as ϵy
±ϵc [1 −
(π/2Lg )2 y2 ], with positive and negative sign for Fourier
and inverse Fourier transform, respectively, are consid-
ered. For the intermediate subblock (i.e., spatial Fourier
filter), either a planar thin metamaterial with trans-
versally inhomogeneous optical properties ϵmm (y)/ϵ0

μmm (y)/μ0 or a transmitarray of metasurfaces, which
consists of composite plasmonic–dielectric nanobricks
[35], can be used. The transmitarray of metasurfaces is
wisely engineered to encode the desired phase and
amplitude patterns into the transmitted wave while
minimizing disturbing reflections, which is necessary for
highly efficient implementation of any desired local
transfer function. Different mathematical operations
were explored including first- and second-order differ-
entiation, integration, and convolution. For the sake of
brevity, we limit the discussion to the first-order spatial
derivation.
Considering the thickness of the planar metamaterial
as Δ < λ0, where λ0 is the free-space wavelength of the
infrared light, the relative permittivity and permeability
associated with the normalized transfer function,
Figure 2: Metamaterials for spatial analog computing [34]. i.e., H(y) ∝ iy/(W/2), is described as follows [34]:
ϵmm
y μmm
y
(A) Cascaded properly designed GRIN(+)/metastructure/GRIN(−)
λ0 −iW
system to perform mathematical operations in the spatial Fourier

i ln . (4)
domain. Two designs are proposed for the middle metastructure ϵ0 μ0 2πΔ 2y
including (i) a specific-purpose thin (Δ = λ0/3) single-layer
metasurface with the prescribed permittivity and permeability (left Figure 2B illustrates the field distribution evolution
inset) and (ii) a general-purpose transmitarray of metasurfaces (so- (snapshot in time) as the wave propagates through the
called metasurface transmitarray or MTA) comprising three computational system. Comparison of the simulated elec-
metasurfaces formed by an array of hybrid plasmonic–dielectric tric field distribution at the output plane (2Lg + Δ) and the
nanocubes that provide a rich set of transmission amplitude and
first-order spatial derivative calculated analytically, given
phase profiles. A snapshot on the evolution of the z-component of
the electric field through the system operating as the first-order
that the input function is f (y) ∝ exp( −y2 ), corroborates
differentiator formed based on (B) the first scenario (i.e., i) and the performance of this approach. As a more practically
(C) second scenario (i.e., ii). Simulation results (both real and feasible approach, the authors leveraged a transmitarray of
imaginary parts) are compared with the analytical solutions in the metasurfaces cascaded with two GRIN(+) media to imple-
insets. GRIN, graded-index. ment the first-order spatial derivation. Figure 2C represents
the comparison assessment between numerical simula-
proper discretization in finite steps of the phase and tions and analytical results.
amplitude of the transfer function and assigning the To achieve strongly independent phase and amplitude
appropriate elements of the metasurface to encode the modulation, Pors et al. [36] utilized the rich nature of gap
transmission or reflection of the incident field. In what surface plasmon mode in a plasmonic metareflectarray. As
follows, different frameworks for hardware implementa- shown in Figure 3A, their proposed structure consists of an
tion of the desired transfer function are discussed in detail. array of gold (Au) nanobricks on a metal substrate sepa-
In a seminal work, Silva et al. [34] proposed to use a rated by an oxide spacer layer. Such a confined mode
system comprising three cascaded subblocks of Fourier excited in the gap between the top metallic nanoresonator
transform, an inhomogeneous spatial Fourier filter, and and the substrate propagates back and forth between the
inverse Fourier transform. As shown in Figure 2A, for the two ends of the nanoresonator leading to the energy
 S. Abdollahramezani et al.: Meta-optics for spatial optical analog computing 4079

generation highly miniaturized computational circuits.

Motivated by that, Abdollahramezani et al. [37] recently
introduced the concept of “metalines” that endows full
control over the amplitude and phase profiles of the
propagating graphene plasmons in a dynamic fashion. As
shown in Figure 4A, a transmitarray of such metalines,
whose surface conductivity can be locally and indepen-
dently controlled through height encoding of the substrate
beneath the graphene layer (or equivalently an external
bias), enables realization of any transfer function in an
ultrathin, integrable, and truly planar platform. Due to the
high confinement of graphene plasmons that empowers
the integration of the Fourier transform subblock in a
cascaded structure, the presented 2D configuration is
orders of magnitude smaller than the traditional
metasurface-based counterparts. To facilitate the calcula-
tion of surface conductivity–dependent transmission/
reflection coefficients (Figure 4B) of the piecewise constant
Figure 3: Reflective plasmonic metasurfaces for spatial analog elements of the dynamic metaline, the authors leveraged a
signal processing [36].
representative simplified analytical treatment on the scat-
(A) Color map of the calculated reflection coefficient as a function of
tering of graphene plasmons from inhomogeneous
the width of nanobricks for a reflective metasurface comprising an
array of Au nanobricks on top of an Au substrate separated with a thin lateral hetrostructures. Figure 4C represents a comparison
oxide layer. Green curves indicate two contours of reflection phase assessment of the computational system (i.e., GRIN lens/
with π-phase difference. (B) Position-dependent reflection amplitude/ transmitarray of metalines/GRIN lens) designed to perform
phase profile (top panel) and the corresponding nanobrick width the second-order spatial derivation upon excitation with a
(bottom panel) for a 50-μm-wide differentiator and integrator. (C)
beam with Sinc profile.
Bright-field images and the measured average normalized reflectivity
from the metasurface realizing the transfer function of differentiation To harness the full potential of plasmonic and
(top panel) and integration (bottom panel). (D) Reflected far-field dielectric modes within a single structure, Chizari et al.
intensity (top panel) and intensity profile averaged along the central [38] proposed to use a metasurface comprising an engi-
part in the y-direction as a function of the x-coordinate (bottom panel) neered array of size-variant nanoresonators on top of
for the differentiator (left panel) and the integrator (right panel)
system. Here, the excitation is through a bulky optical lens acting
a silver (Ag) backreflector separated with an oxide inter-
as Fourier and inverse Fourier subblocks simultaneously. mediate layer (see Figure 5A). The combination of low-
aspect ratio rotating silicon (Si) nanobricks, which sup-
port the low-loss Mie resonance, and the lossy back-
leakage to free space. Strong modulation granted by such a
reflector, which facilitates multiple reflections within the
Fabry–Perot-like resonance within the engineered inclu-
low index spacer layer, grants unique agility on the
sion facilitates manipulation of the scattered near-infrared
required phase delay and amplitude modulation of the
light both spectrally and spatially. As shown in Figure 3B
cross-polarized light. As shown in Figure 5B, the designed
and C, it is evident that the fabricated computational
unit cell can impart full 2π phase coverage corresponding
metasurfaces fairly follow the features of their designed
with the desired amplitude in the range of (0, 1) on the
transfer functions while discrepancies exist in the spatial
reflected near-infrared light. For the full operation, such a
variation of the reflectance due to mainly fabrication tol-
compact design only needs one coupled lens serving as
erances, uncertainty in the material optical properties, and
interference of high-spatial frequencies reflected from the both Fourier and inverse Fourier transform subblocks
bare substrate. Hence, the proof-of-concept demonstration simultaneously. Figure 5C shows that a good agreement is
of the differentiator and integrator is not in very good achieved between the numerical simulation results and
agreement with the numerical calculations as depicted in the designed first-order differentiator and integrator us-
Figure 3D. However, this work is the first experimental ing such a complementary metal-oxide-semiconductor
implementation in the field of computational meta-optics. (CMOS)-compatible metasurface.
Shrinking the conventional bulky 4f correlator to on- While the reflective plasmonic metasurfaces generally
chip integrable configurations is highly desirable for next- provide larger phase span thanks to the added Fabry–Perot
4080 S. Abdollahramezani et al.: Meta-optics for spatial optical analog computing

demonstrated that an all-dielectric metasurface based on an

array of anisotropic Si nanoresonators can fully tailor the
transmission amplitude and phase responses locally and
almost independently (see Figure 6A), which can facilitate
the realization of mathematical operators. Such a unique
feature is granted by the magnetic dipole moment on ac-
count of the circular displacement current excitation and
the electric dipole moment due to considerable charge
accumulation at the corner of each resonator (see
Figure 6B). Thanks to the the interplay between the first and
second Mie resonant modes (i.e., magnetic and electric
dipole moment) achieved by changing the major-to-minor
axis aspect ratio of the nanoresonator, fairly any desired
transformation of the amplitude and phase can be imparted
on the transmitted light (see Figure 6A). The authors lever-
aged the full potential of a cascaded, compact platform of
metalens/computational metasurface/metalens to perform
all-optical signal processing including constant coefficient
integrodifferential equation (see Figure 6C and D), which
has not been shown previously. Taking advantage of the
incorporated metalenses as the functional Fourier and in-
verse Fourier transform subblocks, rather than the bulky
conventional lenses, paves the way for highly integrable on-
demand computational systems.
Considering well-established foundry-based silicon
photonics, implementation of on-chip mathematical op-
erators promises the next generation of small footprint,
low-power consumption, and multipurpose computational
photonic integrated circuits. More recently, Wang et al. [40]
demonstrated parallel signal processing by leveraging a 1D
Figure 4: Linear transversely invariant graphene-based on-chip high-contrast transmitarray of metasurfaces (see
metastructure to perform mathematical operations [37].
Figure 7A). By judiciously adjusting the width and length of
(A) Sketch of a two-dimensional (2D) graphene-based computing
the void slots in the Si-on-insulator substrate, complete
system consisting of two cascaded graded-index (GRIN) lenses
incorporating a transmitarray of three symmetric stacked metalines control of the transmitted amplitude and phase profile can
with total length of D and width of W (inset: basic unit cell of the be achieved over a high bandwidth (see Figure 7B).
metaline). (B) Transmission phase and amplitude of the metaline as Figure 7C illustrates experimental results of the fabricated
a function of the chemical potential associated with the surface three-layer system necessary for successful implementa-
conductivity. (C) Snapshots of the z-component of the electric field
tion of on-chip mathematical operators. While the first and
distribution through the GRIN lens/transmitarray of metalines/GRIN
lens system as a second-order differentiator given that the input second metalenses perform the Fourier and inverse Fourier
wave is a Sinc function. transform, the spatially varying transmission coefficient is
encoded into the middle mask layer to perform the required
1D transfer function (see inset in Figure 7C). The calculated
channel, transmittive metasurfaces grant easy access to the analytical, numerically simulated, and measured spatial
processed field profile at their output, which is a must for spectra are in good alignment for the spatial differentiator
cascadable optical systems. However, optical metasurfaces as shown in the bottom panel of Figure 7C. Exploiting an
supporting plasmonic resonances exhibit intrinsic non- integrated metalens (shown in Figure 7D) significantly
radiative losses and limited scattering cross sections. miniaturizes the overall size of the on-chip photonic
To mitigate these challenges, high-contrast metasurfaces processor.
have recently garnered significant attention in the field of More recently, Zhou et al. [41] have shown experi-
nanophotonics. Abdollahramezani et al. [39] recently mental realization of first-order spatial differentiation and
 S. Abdollahramezani et al.: Meta-optics for spatial optical analog computing 4081

implemented right after the metasurface), only two shaded

areas are left at the output that indicate the edges of the
input image. The fabricated metasurface with 8 mm × 8 mm
pattern area embedded inside a 3-mm-thick glass substrate
is shown in Figure 8F. The metasurface pattern is written by
a femtosecond pulsed laser beam focused 50 μm beneath
the surface of the glass. Interestingly, this edge detection
technique enables tunable resolution at the resultant edges
by varying the PB phase gradient period. Moreover, as
shown in Figure 8G, this edge detection technique is sen-
sitive to the orientation of the phase gradient metasurface
(i.e., the gradient direction of the 1D metasurface indicated
by white arrows). The use of thick glass-based meta-
structure in their implementation makes the metasurface
highly efficient in terms of transmitted power (around
90%) which is the main advantage over the plasmonic
metasurfaces. Additionally, since the operational band-
Figure 5: Analog optical computing using hybrid dielectric–
width of this approach is not limited by the critical plas-
plasmonic metasurfaces [38].
(A) Schematic representation of the dielectric graded-index (GRIN)
monic coupling condition [42], realization of broadband
media/hybrid metasurface to perform mathematical operations transfer functions is guaranteed. An interesting extent of
including spatial differentiation and integration. The metasurface is this work can be implementation of a broadband highly
composed of an arrangement of silicon (Si) nanobricks on an oxide efficient 2D edge detector.
layer deposited on a silver (Ag) substrate. (B) Simulated amplitude
In addition to so far demonstrated platforms enabling
and corresponding phase profiles of the reflected cross-polarized
light considering the rotation angle of 45° with respect to the x-axis mathematical operators, several other theoretical works
for the constituent nanoresonator. (C) Comparison of the simulated exist in the literature using spatial Fourier transformation
reflected cross-polarized field and the exact analytical solution for concept to realize first-order differentiation using bilayered
the first-order differentiator (top panel) and first-order integrator metasurfaces [43], differential and integral operations
(bottom panel).
using Ag dendritic metasurfaces [44, 45], and multiway
parallel mathematical operations based on discrete meta-
demonstrated broadband 1D edge detection by leveraging
materials [46].
the spin–orbit interaction of light and a properly designed
Pancharatnam–Berry (PB) phase metasurface sandwiched
between two orthogonally aligned laser-written linear Glan
polarizers. The PB phase gradient metasurface splits the 3 Green’s function approach
incident linearly polarized (LP) beam into left-handed
circularly polarized (LCP) and right-handed circularly In the GF approach, the desired optical transfer function is
polarized (RCP) beams propagating in opposite directions directly implemented into the wave vector domain (Fourier
(see Figure 8A). To explore the effectiveness of the pro- or k-space) by using the nonlocal (angular dependent)
posed approach, the edges of a squared-shaped object (slit) response of a suitably designed metamaterial or metasur-
are detected in an analog fashion. The real-space image face as shown in Figure 1B [47–51]. Due to the direct
(i.e., the electric field distribution of the object in the spatial implementation of the desired transfer function in this
domain) and the Fourier space spectrum of the object are method, performing Fourier/inverse Fourier transforms is
shown in Figure 8C and B, respectively. By introducing a no longer needed, which in turn leads to the reduction in
PB phase gradient metasurface at the Fourier plane (see the overall structure size [34]. On the other hand, it can be
Figure 8D), the LCP and RCP components of the output very difficult to realize a desired but arbitrary complex
electrical field at the image plane obtain opposite phase transfer function in the GF approach. Here, we present a
gradients. This, in turn, results in a slight shift of LCP and general 2D description of the GF approach which can be
RCP images in opposite directions as shown in Figure 8E. simplified into a 1D description as described later in this
When this output image passes through the analyzer section. First, assume that a field component of a mono-
(i.e., the orthogonal Glan polarizer along y-direction chromatic beam propagating in the +z direction is
4082 S. Abdollahramezani et al.: Meta-optics for spatial optical analog computing

can be represented as an optical transfer function (OTF)

denoted by the following equation [53]:

H TE−TE
k x , k y  H TE−TM
k x , k y 
H
k x , k y 

H TM−TE
k x , k y  H TM−TM
k x , k y 
. (5)

The OTF tensor H(k x , k y ) can transform the incident field

profile to a reflected or transmitted field profile. The manip-
ulation of the geometry of the structure representing the
H(k x , k y ) allows for designing the on-diagonal elements
(couplingofparallelpolarizations)andoff-diagonalelements
(coupling of orthogonal polarizations) in Eq. (5). It should be
notedthat inthis formulation, only the zeroth-order reflection
and refraction are considered to occur. The elements of the
OTF of a certain structure can be calculated using numerical
simulation methods such as the finite element method of
finite difference time domain [53]. By simply multiplying the
incident field profiles with the OTF of a structure, one can
obtain the transmitted or reflected field profile of the corre-
sponding output plane waves as follows [53]:

GTE
k x , k y  F TE
k x , k y 

H
k x , k y 
GTM
k x , k y  F TM
k x , k y 
. (6)
Figure 6: All-dielectric metasystems as equation solvers [39].
(A) Numerically calculated phase and amplitude of transmitted light
from the computational metasurface as a function of minor and
To design the tailored structures for realization of the
major axis dimension. (B) Simulated electric and magnetic lines in desired OTFs, different optical phenomena have been
the cross section of a silicon (Si) nanodisk at the electric and used. Here, we classify them into two main categories as
magnetic resonances fulfilled at structural parameters associated resonance- and nonresonance-based optical phenomena.
with the dots shown in (A). (C) Snapshot of the y-component of the
As a leading work, Silva et al. [34] investigated optimized
simulated electric field propagating through the computational
metasurface realizing the kernel for solving the
multilayered metamaterial slabs, which are homogeneous in
integrodifferentiation equation. the transverse directions while inhomogeneous in the lon-
(D) Comparison assessment between the normalized electric field at gitudinal direction, to realize an appropriate GF associated
the output of the proposed metalens/computational metasurface/ with the second-order spatial derivation. A fast synthesis
metalens system and analytical results for integrodifferential
method to calculate the corresponding optical constant and
equation solving.
thickness of parallel subwavelength layers was developed.
The ten-layer nonmagnetic metamaterial in Figure 9A is
designed to manipulate the transmission coefficient to match
modulated by an arbitrary 2D signal profile f(x, y). This field the second-order derivative kernel for any incidence angle.
profile can be decomposed into transverse electric (TE) and Figure 9B and C demonstrate the evolution of magnetic field
magnetic (TM) linearly polarized plane waves when rep- distribution and the simulation results at the output of the
resented in the wave vector space. Through Fourier trans- multilayered slab upon excitation with a nonregular poly-
form, one can obtain the complex amplitude of each of nomial function and a city skyline border, respectively.
these plane wave components, i.e. F TE (k x , k y ) or Figure 9D shows the simulation results for a five-layer GF
F TM
k x , k y . It is noteworthy that F TE (k x , k y ) and F TM
k x , k y  slab characterized by a rectangular spatial kernel performing
can be obtained only by knowing the transverse compo- convolution operation of a rectangular function.
nents of the electric field. One can calculate other field
components using Maxwell’s equations as described in the
study by Bykov et al. [52]. The next step is to judiciously 3.1 Resonance-based GF approach
design an optical structure that properly manipulates the
amplitude and phase of each of the decomposed TE and TM Different types of optical resonances such as Fano reso-
plane wave of the incident field profile. The application of nance [16, 54, 55], surface plasmon resonance, and guided
such a desired amplitude and phase on these plane waves mode resonance (GMR) have been utilized for
 S. Abdollahramezani et al.: Meta-optics for spatial optical analog computing 4083

theory, the authors in the study by Zhu et al. [42] showed

that the OTF of this structure around kx = 0 and under
critical coupling condition (achieved by simply controlling
the thickness of the metal film) can be approximated as
follows:

HTM (k x ) ≈ eiϕ ik x B (7)

in which ϕ corresponds to the phase change during the

direct reflection at the glass–metal interface, and B is
expressed in terms of the radiative leakage rate of the SPP
and the intrinsic material loss rate, showing that the
structure in Figure 10A can perform first-order derivative
for off-normal incidence (i.e., θ0 ≠ 0) as shown in
Figure 10B. Given that a first-order differentiator is able to
map the abrupt changes in the spectrum (either amplitude
or phase) of a input signal to some sharp peaks in an output
signal, the authors in the study by Zhu et al. [42] experi-
mentally demonstrated 1D edge detection of the input
Stanford logo image as shown in Figure 10C.
To broaden the operational spatial bandwidth for on-
demand high-resolution edge detection, Kwon et al. [50]
Figure 7: On-chip analog computing based on a Si-on-insulator proposed to use a high-efficiency transmittive metasurface
(SOI) platform [40]. consisting of a periodic array of split-ring resonators (SRRs)
(A) Schematic representation of an on-chip high-contrast
as demonstrated in Figure 11A. They theoretically showed
transmitarray defined on a SOI substrate to manipulate the
transmittance and phase response of the in-plane propagating that by introducing a periodic (sinusoidal) permittivity
wave. (B) Simulated amplitude (top panel) and phase (bottom panel) modulation in SRR gaps, the nonlocal response of the
spectra of the transmitted field as functions of the slot width (w) and metasurface can be engineered to achieve desired OTFs for
height (h) for an incident wavelength of 1550 nm. Contours with 2π the second-order differentiator. The interaction of the
phase shift are depicted by dotted lines. (C) The input incident wave
surface GMRs with the leaky-wave resonances can lead to
profile (top panel) and analytical, simulated, and measured
distribution at the output (bottom panel) of the on-chip some Fano resonances in the transmission response of this
differentiator. The conceptual design and scanning electron metasurface as shown in Figure 11B. Working at the
microscopy (SEM) of the fabricated device are shown in the inset. (D) transmission zero frequency of the Fano resonance at
SEM of the fabricated on-chip metalens located in front of a mode normal incidence, the transmission response of the
expander and coupled to 11 single-mode waveguides to facilitate
nonlocal metasurface changes from zero to unity for obli-
characterization of the spatial distribution of focused light.
que incidence (see Figure 11C). This functionality can be
implementing OTFs to perform mathematical operations in utilized for performing wide bandwidth second-order dif-
wave vector domain. ferentiation as shown in Figure 11D. They also modified this
In 2017, Zhu et al. [42] experimentally demonstrated a structure by adding a misplaced array of metallic wires to
surface plasmon–based spatial differentiator using a sim- break the vertical and horizontal mirror symmetry (see
ple metal–dielectric plasmonic structure in the reflective Figure 11E). Such a suitably engineered configuration en-
Kretchmann prism configuration as shown in Figure 10A. ables realization of the first-derivative kernel as shown in
They showed that when the parallel component of the Figure 11F. Figure 11G represents the input signal and the
TM-polarized incident wave to the interface matches with corresponding simulated response of this mathematical
the wave vector of the surface plasmon polariton (SPP) at operator in comparison to the numerically calculated
the metal–air interface, a strong SPP is excited and prop- result. In a further study, by integration of two identical 1D
agates along the metal surface. The interaction between operation metasurfaces relatively rotated by 90°, which
the radiation of the SPP leakage with the direct reflection implies 90° rotational symmetry, such a configuration
from the glass–air interface determines the amplitude of enabled realization of an identical second-order derivation
the reflected beam and, in turn, the OTFs of the structure. for TM-polarized waves along the x and y axes. This plat-
Considering this mechanism and using coupled mode form can be leveraged for more practical 2D edge detection
4084 S. Abdollahramezani et al.: Meta-optics for spatial optical analog computing

beam. GMRs have also used for implementing optical in-

tegrators. The authors in the study by Zangeneh-Nejad and
Khavasi [60] utilized a prism coupling configuration to
implement the OTF of a 1D optical integrator in the trans-
mission response of a multilayer dielectric slab structure. It
is shown that the dielectric slabs can be replaced by gra-
phene sheets to enable miniaturized optical integrator.
In another work, Guo et al. proposed to use a photonic
crystal slab (see Figure 12A) in the transmission mode to
realize the Laplacian operation on a 2D input, i.e.
∇2
∂2x + ∂2y . The ∇2 operation in real space is equivalent to
the following OTF [56]:

T TE−TE
k x , k y 
H
k x , k y 

0
T TM−TM
k x , k y 
. (8)
0

Figure 8: Edge detection using Pancharatnam–Berry (PB) phase To realize this ideal 2D Laplacian operation, the au-
gradient metasurfaces [41]. thors in the study by Guo et al. [56] utilized the guided
(A) The concept of PB phase. Upon illumination of an incident
modes of the photonic crystal slab near the Γ point in the
linearly polarized (LP) beam, the PB phase metasurface splits the
light into left-handed circularly polarized (LCP) and right-handed
Brillouin zone, and the authors in the study by Guo et al.
circularly polarized (RCP) beams with opposite directions. The [62] used isotropic image filters to have identical responses
Fourier space spectrum (B) and real-space image (C) of a square to both polarization and make the off-diagonal elements
object. When a PB phase gradient metasurface is added at the equal to zero as required by Eq. (8). It is worth mentioning
Fourier plane, (B) and (C) change to (D) and (E), respectively. (F) The that working in the transmission mode is more compatible
fabricated metasurface with a pattern area of 8 × 8 mm embedded in
for image processing applications which is an advantage of
a 3-mm-thick glass substrate (scale bar is 5 mm). The polariscope
optical image of the area marked in the left panel is shown in the the work in the study by Guo et al. [56] over other ap-
right panel (scale bar is 25 μm). The constituent elements in one proaches for Laplace operator implementation [52]. The
period of the metasurface are indicated by the red bars in the right numerical demonstration of the proposed structure shown
panel, and the inset shows the SEM of the fabricated device (scale in Figure 12A is demonstrated in Figure 12B, where the
bar is 500 nm). (G) The top left panel shows the original image, while
transmitted image of an unpolarized incident beam of the
other three panels show detected images at the output of the
system. Different orientation of the intermediate metasurface, Stanford emblem, as well as some slot patterns, is calcu-
which is indicated by the white arrows, resolves different portion of lated. Later, Bezus et al. [63] employed the resonance of a
edges. dielectric ridge on a slab waveguide to implement both
optical differentiation and integration in the spatial
domain.
scenarios where unpolarized light for the illumination of Zhou et al. [61], more recently, demonstrated the
the image is utilized as depicted in Figure 11H where the applicability of flat optics for direct image differentiation
edges of the City University of New York (CUNY) logo are using a compact all-dielectric metastructure. In their work,
well resolved. first a metasurface-based differentiator was used in com-
The authors in the study by Saba et al. [57] employed the bination with the conventional optical components and a
GMR of a periodic array of dielectric resonators buried in a camera sensor for high-speed edge detection (see
dielectric slab to perform 2D edge detection. The interaction Figure 13A). Such a metasurface affords a improved
between the GMRs with the leaky waves of the structure transmission amplitude profile as shown in Figure 13B. In a
results in some Fano resonances in the transmission second approach, a Si-based photonic crystal in combi-
response [58]. Working at the zero transmission corre- nation of a metalens was employed to realize the second-
sponding to the normally incident light, the authors order derivation to directly discriminate edges in an image
numerically demonstrated a second-order differentiator (see Figure 13C). Such a compact and monolotic image
owing to the even symmetry of the structure. In another processing system paves the way for real-time computer
work, the authors in the study by Bykov et al. [59] experi- vision tasks, as depicted in Figure 13D.
mentally demonstrated a first-order differentiator by In a follow-up work to the study by Kwon et al. [50],
employing GMRs in the case of an oblique incident Gaussian Cordaro et al. [49] designed and implemented an all-
 S. Abdollahramezani et al.: Meta-optics for spatial optical analog computing 4085

Figure 10: (A) The structure of a surface plasmon-based first-order

spatial differentiator consisting of a Ag film on top of the glass in the
Kretschmann configuration [42]. (B) The first-order spatial differ-
entiator optical transfer function (OTF) obtained by the experimental
measurement (dotted lines) and the numerical fitting (solid lines).
(C) Edge detection of the Stanford tree logo and the letters as the
incident image (left panel) using the plasmonic structure shown in
(A). The reflected intensity image is shown in the right panel.

Several other proposals and demonstrations using

Figure 9: Spatial analog computing using metamaterial slabs based
on Green’s function approach [34].
spin Hall effect (SHE) of light [65], prism coupling config-
(A) Designed multilayered slab to realize the desired Green’s uration [66], reflective hybrid plasmonic–dielectric meta-
function (GF) kernel corresponding to the selected mathematical surfaces [48], periodic plasmonic metasurfaces covered by
operation. Evolution of electromagnetic wave in the ten-layered graphene [67], multiinput–multioutput computational
nonmagnetic GF slab designed to perform second-order metasurfaces [68], ultrathin bianisotropic metasurfaces
differentiation on a quadratic polynomial function (B) and a city
[69], polarization-insensitive structured surfaces with
skyline (C). In (C), the simulated output result is directly compared
with the input function to highlight the performance of the system in tailored nonlocality [70], and engineering the spatial
detecting sharp edges. (D) GF slab with five layers designed to dispersion of the electric dipole resonance in dielectric
realize convolution with a rectangular kernel. metasurfaces [71] to perform mathematical operators based
on GF approach have been reported in the literature.

dielectric metasurface formed by a 1D array of Si nano- 3.2 Nonresonance-based GF approach

beams to locally tailor the optical properties of the incident
S-polarized light. Due to engineered spatial dispersion, the The authors in the study by Youssefi et al. [64] employed
interference of the broad Fabry–Perot resonance of struc- the Brewster effect, which is associated with the zero
ture with the quasi-guided mode along the surface results reflection of a TM-polarized incident beam from an inter-
in an asymmetric Fano resonance in the transmission mode face between two dielectric medium, to realize first-order
shown in Figure 14A. The sharp response in frequency is differentiation. The schematic of their proposed structure is
associated with strong nonlocality governing the angular shown in Figure 15A in which the symmetry of the system is
sensitivity of the transmission response. The strong broken by applying oblique incident beam. Figure 15B
amplitude modulation correlated to the incoming k-vector shows the exact GF and its approximation around ky = 0
is the basis for implementing a spatial filter associated with which can be used for implementation of first-order de-
the mathematical operator of choice. The simulated rivative. To study the performance of the proposed
transmission amplitude and phase as a function of the configuration, the proposed structure is illuminated with a
normalized in-plane wave vector for the second-order dif- Sinc function beam profile with the bandwidth of
ferential kernel at the operational wavelength of 633 nm is W = 0.09k0 at the Brewster angle. The calculated first-order
shown in Figure 14B. Figure 14C shows the results of the derivative of the input field is compared with the exact
simulated metasurface; a rectangular input profile results solution as shown in Figure 15C. Later, the same group
in doublet signals at the edges of the rectangle, and the demonstrated that using a simple half-wavelength dielec-
sinusoidal input function results in the sinusoid becoming tric slab in the reflection mode, first-order differentiation
flipped. can be realized [72]. Using a transmission-line approach,
4086 S. Abdollahramezani et al.: Meta-optics for spatial optical analog computing

or refracted at a single optical planar interface under par-

axial approximation. They employed SHE of light, i.e. the
polarization-dependent transverse shift of an optical beam
totally reflected from a planar interface [73], to compute the
spatial differentiation of the incident beams. As Figure 16C
shows, when the obliquely incident paraxial beam has an
electric field distribution of f(x, y), the output field distri-
bution associates with df (x, y)/dy. To validate their
finding, the authors in Zhu et al. [65] performed experi-
mental measurements of the OTF of a glass–air interface as
shown in Figure 16C. Moreover, they applied their method
to a Gaussian incident beam profile and performed the first-
order spatial differentiation as the first-order Hermite–
Gaussian profile of the reflected beam in Figure 16C shows.
Finally, they applied their spin-optical method to perform
an edge detection operation. Figure 16D demonstrates the
incident image field of a Chinese character encoded in the
x-component of the field based on the amplitude modula-
tion. The measured y-component of the reflected field ex-
hibits the resolved outlines of the character (see Figure 16E)
which are more visible in the x direction. The edges parallel
to the x direction are more visible since the differentiation
is performed along the y direction.
More recently, Estakhri et al. [74] in a disruptive
approach demonstrated a computational metastructure
platform that solves linear integral equations by tailoring
the complex-valued electromagnetic wave propagating
through the specially designed recursive paths. The con-
ceptual representation of the metastructure is shown in
Figure 17A where the solution to the Fredholm integral
equation of the second kind is obtained. While the given
Figure 11: Nonlocal metasurface. integral operator with kernel K(u, v) is implemented in the
(A) The structure of a metasurface consisting of an array of split-ring metamaterial block, the input signal Iin(u) is introduced
resonators (SSRs) under TM illumination [50]. (B) The spectrum of to the equation through a set of coupling elements
the transmission response of the metasurface when SSRs are
along the feedback waveguides. The overall relation
periodically modulated. (C) The optical transfer function (OTF) if the
metasurface fitted to a second-order differentiator. (D) Output of an describing the behavior of the equation solver follows
b
ideal second-order differentiator and of the metasurface from the g(u)
I in (u) + ∫a K(u, v)g(v)v, in which g(u) is the un-
input signal shown in the top panel. (E) Broken-symmetry
metasurface enabling an asymmetric response with respect to
known solution of the integral equation. To realize the
positive and negative kx to realize first-order differentiation. (F) The equivalent N × N matrix equation of the abovementioned
transmission profile as a function of the incidence angle for a governing equation, N feedback waveguides are exploited
symmetry-broken metasurface enabling first-derivative operation. to sample the input of the metamaterial block to its output.
(G) Output of an ideal first-order differentiator working as an edge
To showcase the performance of this approach in a phys-
detector. (H) Detected edges of the CUNY logo for illumination with
unpolarized light from the normal direction [50].
ical system, the authors leveraged a judiciously patterned
metastructure included in a feedback loop, which calcu-
lated the inverse of a known N × N matrix. Figure 17B shows
they showed that employing impedance matching condi- the numerical simulation results for the distribution of the
tion, one can perform first-order differentiation on oblique electric field in the assembled system consisting of the
incident input signals. prescribed metastructure in combination with the five
Recently, Zhu et al. [65] experimentally demonstrated feedback waveguides and directional couplers. The com-
spatial differentiation of the incident beam when reflected parison assessment between the theoretical and
 S. Abdollahramezani et al.: Meta-optics for spatial optical analog computing 4087

simulation results corroborate the fidelity of the proposed solution, as well as spatial frequency filters. Despite the
approach. In a proof-of-concept demonstration at the mi- existing promising platforms and design approaches
crowave regime, the feedback process is internally per- enabling computational metasystems, there is still room
formed in a reflective system, as shown in Figure 17C. The for improvement and progress in this rather infant field. In
topology optimization method was leveraged to inversely what follows, we elaborate some of unexplored aspects,
design an optimized metastructure realizing the kernel key challenges, and possible opportunities in the general
associated with the operator of choice. The simulated field of computational meta-optics.
electric field and comparison between the numerical and In general, the architecture of a conventional optical
experimental results are represented in Figure 17C. processor is relied on cascaded input, processing, and
output planes. To leverage the profound potential of meta-
optics for realization of flat optical processors, miniaturiza-
tion and possible stacking of planes are indispensable. The
4 Summary and outlook first (or input) plane, where mostly an electrical to optical
conversion of raw data is performed at the video rate using
Spatial analog computing platforms facilitate wave-based, an SLM, is considered to be the bottleneck of most practical
real-time, high-throughput, and large-scale information computational systems. More recently, several tuning
processing with low-energy consumption. Fourier trans- mechanisms including electrical, thermal, mechanical, and
formation and GF have been pursued as two powerful optical have been introduced [75] enabling reconfigurable
concepts for realization of mathematical operations. While nanoscale SLMs and phased array antennas. Tuning the
the former is relied on 4f systems with two Fourier trans- refractive index of Si based on thermooptic effects [76],
form lenses and one intermediate spatial frequency filter, manipulating the electrooptic characteristics of indium tin
the latter leverages the nonlocal response of judiciously oxide [77, 78] relying on free carrier effects, controlling the
designed optical components to implement the desired effective physical state of liquid crystals [79] exploiting the
kernel. Due to their unprecedented capabilities in sculpt- ultrafast transport dynamics of hot electrons in a hybrid
ing the scattered light, metastructures enable miniaturi- crystal [80], and phase conversion of phase change mate-
zation of traditional bulky optical systems to the integrable rials between the amorphous and crystalline states [81–89]
ultracompact frameworks. The intersection of the afore-
mentioned fundamental theoretical concepts and func-
tionalized metamaterials and metasurfaces promise
realization of computational meta-optics. We discussed
recent advances in spatial analog optical computing de-
vices enabling mathematical operations including inte-
gration, differentiation, and integrodifferential equation

Figure 13: Two-dimensional (2D) image differentiation using

Figure 12: (A) The schematic of the photonic crystal structure for nanophotonic materials [61].
Laplacian differentiation consisting of a photonic crystal slab (A) Schematic of a photonic crystal slab acting as a Laplacian
separated from a uniform dielectric slab by an air gap [56]. operator that transforms an image, Ein, into its second-order
(B) Incident Stanford emblem (top left panel) and calculated derivative, E out ∝ ∇2 E in . (B) Simulated color-coded transmission
transmitted image with unpolarized light (top right panel). The coefficient amplitude as a function of frequency and incident angle
spatial resolution of the differentiator is shown in the bottom panels along the Γ − X direction (ϕ
0°) for p polarization. (C) SEM of the
(incident slot patterns in the left and calculated transmitted images fabricated Si photonic crystal. (D) Imaging results for the target
with unpolarized light in the right). without (top row) and with (bottom row) the differentiator.
4088 S. Abdollahramezani et al.: Meta-optics for spatial optical analog computing

focusing [93], computational imaging [94], and multiplane

holography [95] promise the emergence of these platforms as
new paradigms in computational systems. Moreover, recent
development of multiplexed and multifunctional meta-
surfaces through integration of different information chan-
nels into a single or multilayered metasurfaces with
segmented or interleaved metaatoms [96, 97] could have
deep impact on the future optical computation technology.
Finally, integration of resonant metasurfaces with III–V
compound semiconductors with strong adjustability helps
shrinking the overall size of the photodetector arrays or even
cameras in the last (or output) layer of an optical computing
system.
Feature selection and object detection are key steps in
image processing applications which heavily rely on edge
detection of input data. So far, most demonstrations have
suffered from the narrow operational spatial bandwidth
associated with the low resolution of processed images. In a
Figure 14: Optical computing based on Fano-resonant theoretical study, Karimi et al. [98] discussed the funda-
metasurface [49]. mental gain-resolution limit that applies to optical analog
(A) Simulated transmission spectra of a metasurface consisting of an edge detectors. Considering the Rayleigh criterion, the
array of dielectric nanobeams as the incident angle is changed from
derived formula reveals a linear relation between the gain of
0 (blue line) to 0.3 rad (yellow line) in 15 steps. The red dashed line
indicates the wavelength of operation (633 nm). (B) Transmission the spatial filter and the achievable image resolution. The
amplitude (solid blue line) and phase (solid orange line) for the authors demonstrated that a simple dielectric slab wave-
metasurface optimized for second-derivative operation (sketched in the guide can theoretically outperform the available flat optical
inset) at 633 nm. The simulated transfer function is compared to the differentiator. The calculated physical bound for a generic
ideal case (dashed lines). The transmission reference plane is set such
spatial differentiator has provided a useful metric to
that the transmission phase at normal incidence equals −π. (C)
Rectangular and sinusoidal input functions and 2D image that are used
compare and assess the performance of existing edge
to numerically test the metasurface operation. The signal is discretized detection devices, which needs to be taken care of for future
into 1000 pixels with individual pixel size set such that the Nyquist improvements. In a recent work, the effect of a time modu-
range matches the operational range in k-space of the metasurface (left lated incident light on the response of a plasmonic spatial
column). Metasurface output for the input in part b. For the 2D image, differentiator based on a prism coupling configuration is
differentiation is performed line by line along the x-axis (right column).
investigated [66]. The authors in the study by Zhang et al.
The size of each input pixel is set so that the Nyquist range matches the
operational bandwidth of the metasurface in the k-space. [66] have shown that the plasmonic spatial differentiator can
operate with an estimated speed of 1013 frame/s. This high
time bandwidth of the plasmonic spatial differentiator in
conjunction with its high space bandwidth provides the
have shown interesting capabilities. Such approaches ulti- ability for realization of high-throughput real-time image
mately can grant moderate- to high-speed addressability at processing applications.
the pixel-level necessary for high-resolution on-demand To improve the robustness of computational meta-
imaging and computing applications. Furthermore, by surfaces against defects or discontinuities at the interface,
combination of an arrangement of high quality factor reso- Zhang and Zhang [99] proposed to use precise tuning of the
nant nanoantennas with quantum dots, strong spontaneous balance between the asymmetric leak rate and the intrinsic
emission occurs leading to light-emitting metasurfaces. The loss rate of the unidirectional SPP leaky mode in a nonre-
judiciously designed array of such metasurfaces can effec- ciprocal plasmonic platform. To do so, they leveraged the
tively manipulate the radiation pattern that paves the way nontrivial topological properties of a gyrotropic material
toward creation of on-chip flat sources with tailored light whose time reversal symmetry is broken under a static
fields [90–92]. The second (or processing) plane which magnetic field. The investigated double-layer flat structure
generally accommodates lenses, holograms, or nonlinear enables ideal realization of first-order spatial differentia-
elements can also take advantage of best-of-breed all- tion for fairly insensitive edge detection in the terahertz
dielectric metasurfaces. Recent advances in achromatic regime.
 S. Abdollahramezani et al.: Meta-optics for spatial optical analog computing 4089

technique has been relied on high-resolution electron-

beam lithography. Despite successful experimental dem-
onstrations, none of these structures have been broadly
adopted to the industry mainly due to small features in
their structures, which are challenging to be resolved with
available industry-standard photolithography technolo-
gies [101]. To address this issue, fabrication constraints
have to be incorporated in inverse design algorithms. One
way to do so is simply subdividing the design region into
rectangular pixels which are larger than the minimum
allowable feature size. However, this pixel-based approach
is not optimal for most practical optical devices with
smooth curves in their structures. Although one approach
to consider these smooth curves in the design process is to
use a convolution filter followed by thresholding [117], this
can result in artifacts smaller than achievable feature sizes.
Figure 15: Example of nonresonance-based GF approach.
(A) The realization of the first-order differentiation when an incident
As a solution to this problem, Piggott [118] proposed to
beam impinging and reflected at Brewster angle [64]. (B) The exact impose curvature constraints on the device boundary in the
and approximated GF distribution corresponding to the first-order algorithm. In this boundary parameterized optimization, a
differentiation. (C) Comparison of the exact solution and the minimum radius of curvature is enforced to avoid features
simulated first-order derivation for an input Sinc function with smaller than an acceptable threshold, and, simulta-
W = 0.09k0.
neously, any gaps or bridges narrower than a chosen
threshold are periodically eliminated by cutting them in
So far, implementation of computational metadevices half. Although these techniques have shown improve-
has been majorly relied on intuition-based approaches ments in terms of fabrication tolerances when tested
where metastructures utilize simple geometries governed by through experiments for electron-beam lithography and
well-understood physics. When it comes to mathematical through simulations for photolithography, their robust-
operators governed by complex multifunctional kernels, ness against systematic/random errors in photolithog-
traditional design techniques (such as brute force) suffer raphy such as defocusing and dosage errors is still
from considerable drawbacks in translating the well- questionable. Therefore, it is essential to incorporate the
formulated transfer function to the real physical device robustness against these fabrication process variations
platform. This is mainly due to hyperdimensional nature of into the optimization algorithm in order to successfully
the optimization problem associated with sophisticated deal with the industry-based fabrication challenges.
input to output mapping where parametric sweeps exhibit Leveraging highly confined SPPs can facilitate the ul-
inefficiency to be employed as an ideal optimizer. Recently, tracompact computational devices based on 2D planar
inverse design approaches based on local and global opti- plasmonic configurations. Kou et al. [119] demonstrated
mization techniques have attracted significant attention in the complex operation of Fourier Transform carried out at a
enabling nontrivial high-performance meta-optic configu- velocity close to the speed of light based on highly confined
rations targeted to a wide range of applications. Among plasmon waves. Thanks to the reduced dimensionality,
several inverse design approaches, global step-by-step their SPP-based device enables ultracompact computa-
searching algorithms (such as genetic or particle swarm), tional devices with high accessible spatial resolution way
adjoint-based topology optimization implementations, and beyond the diffraction limit of the incident light.
neural network–assisted optimization approaches have Light–matter interaction in 2D materials and transition
proven to be compelling candidates to push forward high- metal dichalcogenides in the form of plasmon polaritons or
performance nanophotonic devices [100–116]. Coupled to exciton–polaritons has attracted intense interests recently
recent advances in nanofabrication technologies, physical [120–122]. Thanks to their atomic-scale size, a great deal of
modeling, and computational power, such single and mul- opportunities can be seen for analog computing applica-
tiobjective optimization approaches can benefit next- tions by harnessing these types of dynamics. In particular,
generation computational metaprocessors. easy-to-fabricate, wafer-scale exfoliated graphene can be
Thus far, fabrication of functional nanophotonic de- served as a real deep subwavelength metasurface platform
vices inverse designed by a topology optimization with profound addressability of gate-tunable metaatoms.
4090 S. Abdollahramezani et al.: Meta-optics for spatial optical analog computing

platforms [125, 126]. In this regard, leveraging powerful,

elegant techniques such as dispersive Fourier transformation
governing dispersive medium with high group delay disper-
sion benefiting from time lens for chirp modulation would be
useful [127, 128]. Optical neural networks based on sub-
wavelength diffractive optics have gained significant atten-
tion lately. Colburn et al. [129] proposed an optical frontend
for a convolutional neural network to alleviate the substantial
challenges with the energy consumption and latency of
existing electronic-based networks. They leveraged meta-
optics to implement a dense architecture of 4f corrolator that
incorporates a general complex-valued transparent mask
which is a physical equivalent of the Fourier transform of the
desired kernel. The optical frontend that carries the pre-
dominant burden of computation is coupled to software-
based implementations of subsequent layers (including
nonlinearity) to compete with the fully electronic networks
dealing with large date sets. A useful comparison assessment
on the classification accuracy of Modified National Institute
of Standards and Technology (MNIST) data as well as
insightful discussions on the fundamental information ca-
pacity of the system are also provided. The introduced plat-
form can be leveraged for practical implementation of hybrid
photonic–electronic neural networks [130–132].
Extending the applicability of mathematical operations
to the microwave wavelengths could take advantage of the
ease of fabrication and simpler characterization of artificial
structures [133]. To showcase the effectiveness of quantum
Figure 16: (A) Schematic of spatial differentiation from the spin Hall analog behaviors in wave-based signal processing, micro-
effect (SHE) of light on an optical planar interface between two wave metamaterials were utilized as quantum searching
isotropic materials, e.g., an air–glass interface [65]. (B) simulators [134]. The explored dielectric structure that relies
Measurement of the spatial spectral transfer function on an air–
on a 3D-printing technology performs Grover’s search algo-
glass interface. (C) Spatial differentiation demonstration for a
Gaussian illumination with an incident angle θ0 = 45°. (D) Incident
rithm, as a fast solver paradigm with quadratic speedup over
image comprises a Chinese character for encoded with amplitude classical methods and comparable to well-known quantum
modulation on E in x . (E) Measured cross-polarized reflected light computing. The hardware implementation of other quantum
(i.e., E in
y ) form the structure illuminated with the incident filed in (D). algorithms such as Deutsch–Jozsa algorithm has been also
experimentally demonstrated more recently [135]. Such
compelling implementations reveal the potential of
Although the expanded family of these materials provides computational metastructures in realization of beyond-
a rich set of metasurface frameworks working at opera- classical mathematical operations. In a distinct work, us-
tional wavelengths down to the visible range, due to their ing a single programmable digital metasurface architecture
ultrathin nature, achieving high-performance optical ele- with spatiotemporal response, the authors in the study by
ments with good phase agility is still questionable [123]. Rajabalipanah et al. [136] realized multiple analog signal
More recently, metasurfaces have shown great potentials processing functions at microwave frequencies with specific
for optical pulse shaping applications by finely tailoring the application to detect the edges for sharp changes in the
temporal profile of a near-infrared femtosecond pulse [124]. In incident field.
analogy to the spatial analog computing approaches such Beyond the optical and microwave analog computing,
compelling architectures could be exploited to realize tem- recently, acoustic signal shows strong potential to enable
poral analog signal processing including differentiation, complex computing applications and signal processing.
integration, and Hilbert transformation which have been Leveraging crystals of interconnecting pipes [137], topo-
already demonstrated in photonic integrated circuit logical insulators [138], layered labyrinthine metamaterials
 S. Abdollahramezani et al.: Meta-optics for spatial optical analog computing 4091

the first three drawbacks to analog systems, there are still

several fundamental physical constraints (imposed by
quantum mechanics and thermodynamics) and practical
limitations (such as lengths of communication lines and the
corresponding signal transit times) with them.

Acknowledgment: The authors thank Tyler Brown and

Kirsten Masselink for helpful comments.
Author contribution: All the authors have accepted
responsibility for the entire content of this submitted
manuscript and approved submission.
Research funding: This work was supported by the Office of
Naval Research (ONR) (N00014-18-1-2055, Dr. B. Bennett).
Conflict of interest statements: The authors declare no
conflicts of interest regarding this article.

References
[1] N. P. Jouppi, C. Young, N. Patil, et al., “In-datacenter
performance analysis of a tensor processing unit,” in
Proceedings of the 44th Annual International Symposium on
Computer Architecture, 2017, pp. 1–12.
[2] A. Marandi, Z. Wang, K. Takata, R. L. Byer, and Y. Yamamoto,
Figure 17: Solving integral equation with waves in inverse-designed “Network of time-multiplexed optical parametric oscillators as a
metastructures [74]. coherent ising machine,” Nat. Photonics, vol. 8, p. 937, 2014.
(A) Left panel: the schematic of a closed-loop network consisting of a [3] P. L. McMahon, A. Marandi, Y. Haribara, et al., “A fully
properly designed kernel operator (i.e., the metamaterial block) that programmable 100-spin coherent ising machine with all-to-all
performs the desired integral operation on the input wave, N connections,” Science, vol. 354, p. 614, 2016.
feedback waveguides for realization of the recursive paths, and [4] M. Babaeian, P.-A. Blanche, R. A. Norwood, et al., “Nonlinear
directional couplers to feed/probe the input/output wave. Right optical components for all-optical probabilistic graphical
panel: complex density plot of the first complex-valued kernel K1(u, model,” Nat. Commun., vol. 9, p. 1, 2018.
v) in the (u, v) plane in Equation (2) in the study by Estakhri et al. [74]. [5] P. Ambs, “Optical computing: a 60-year adventure,” Adv. Opt.
(B) Left panel: time snapshot of the simulated out-of-plane Technol., vol. 2010, pp. 1–15, 2010.
component of the electric field distribution in the closed-loop [6] R. Athale and D. Psaltis, “Optical computing: past and future,”
network when excited at port 1, with the other four ports unexcited. Optic Photonics News, vol. 27, p. 32, 2016.
Right panel: inhomogeneous metamaterial kernel designed for [7] J. Touch, A.-H. Badawy, and V. J. Sorger “Optical computing,”
performing integral operation enclosed with absorbing layers on top Nanophotonics, vol. 6, p. 503, 2017.
and bottom and perfect electric conducting walls on all other sides. [8] H. J. Caulfield and S. Dolev, “Why future supercomputing
(C) Left panel: an image of the fabricated metastructures in which requires optics,” Nat. Photonics, vol. 4, p. 261, 2010.
white, gray, and black areas show polystyrene, air, and absorbing [9] D. R. Solli and B. Jalali, “Analog optical computing,” Nat.
materials, respectively. The outer electric conducting walls are made Photonics, vol. 9, p. 704, 2015.
of aluminum. Right panel: time snapshot of the simulated out-of- [10] R. S. Tucker, “The role of optics in computing,” Nat. Photonics,
plane component of the electric field distribution in the closed-loop vol. 4, p. 405, 2010.
network when excited at port 3. [11] J. W. Goodman, Introduction to Fourier Optics, USA, Roberts and
Company Publishers, 2005.
[12] N. Engheta and R. W. Ziolkowski, Metamaterials: Physics and
[139, 140], people have shown several interesting results in Engineering Explorations, USA, John Wiley & Sons, 2006.
this promising field. [13] N. Yu, P. Genevet, M. A. Kats, et al., “Light propagation with
Historically, most existing optical computing works are phase discontinuities: generalized laws of reflection and
refraction,” Science, vol. 334, no. 6054, p. 1210713, 2011.
in the analog, continuous domain. However, there are some
[14] N. Yu and F. Capasso, “Flat optics with designer metasurfaces,”
significant drawbacks to optical analog processing including
Nat. Mater., vol. 13, p. 139, 2014.
limited flexibility, noise susceptibility, approximation error, [15] A. E. Miroshnichenko, S. Flach, and Y. S. Kivshar, “Fano
and input–output device limitations [141]. Although digital resonances in nanoscale structures,” Rev. Mod. Phys., vol. 82,
optical processors have been firstly introduced to circumvent p. 2257, 2010.
4092 S. Abdollahramezani et al.: Meta-optics for spatial optical analog computing

[16] B. Luk’yanchuk, N. I. Zheludev, S. A. Maier, et al., “The fano [36] A. Pors, M. G. Nielsen, and S. I. Bozhevolnyi, “Analog computing
resonance in plasmonic nanostructures and metamaterials,” using reflective plasmonic metasurfaces,” Nano Lett., vol. 15,
Nat. Mater., vol. 9, p. 707, 2010. p. 791, 2015.
[17] H. A. Atwater and A. Polman, “Plasmonics for improved [37] S. Abdollahramezani, K. Arik, A. Khavasi, and Z. Kavehvash,
photovoltaic devices,” in Materials for Sustainable Energy: A “Analog computing using graphene-based metalines,” Opt.
Collection of Peer-Reviewed Research and Review Articles from Lett., vol. 40, p. 5239, 2015.
Nature Publishing Group, Singapore, World Scientific, 2011, [38] A. Chizari, S. Abdollahramezani, M. V. Jamali, and J. A. Salehi,
pp. 1–11. “Analog optical computing based on a dielectric meta-reflect
[18] Y. Li, S. Kita, P. Muñoz, et al., “On-chip zero-index array,” Opt. Lett., vol. 41, p. 3451, 2016.
metamaterials,” Nat. Photonics, vol. 9, p. 738, 2015. [39] S. Abdollahramezani, A. Chizari, A. E. Dorche, M. V. Jamali, and
[19] A. Arbabi, Y. Horie, M. Bagheri, and A. Faraon, “Dielectric J. A. Salehi, “Dielectric metasurfaces solve differential and
metasurfaces for complete control of phase and polarization integro-differential equations,” Opt. Lett., vol. 42, p. 1197, 2017.
with subwavelength spatial resolution and high transmission,” [40] Z. Wang, T. Li, A. Soman, D. Mao, T. Kananen, and T. Gu, “On-chip
Nat. Nanotechnol., vol. 10, p. 937, 2015. wavefront shaping with dielectric metasurface,” Nat. Commun.,
[20] A. I. Kuznetsov, A. E. Miroshnichenko, M. L. Brongersma, vol. 10, p. 1, 2019.
Y. S. Kivshar, and B. Luk’yanchuk, “Optically resonant dielectric [41] J. Zhou, H. Qian, C.-F. Chen, et al., “Optical edge detection based
nanostructures,” Science, vol. 354, pp. 846–854, 2016. on high-efficiency dielectric metasurface,” Proc. Natl. Acad. Sci.
[21] S. Jahani and Z. Jacob, “All-dielectric metamaterials,” Nat. U. S. A., vol. 116, p. 11137, 2019.
Nanotechnol., vol. 11, no. 1, p. 23, 2016. [42] T. Zhu, Y. Zhou, Y. Lou, et al., “Plasmonic computing of spatial
[22] I. Staude and J. Schilling, “Metamaterial-inspired silicon differentiation,” Nat. Commun., vol. 8, p. 15391, 2017.
nanophotonics,” Nat. Photonics, vol. 11, p. 274, 2017. [43] M. Farmahini-Farahani, J. Cheng, and H. Mosallaei,
[23] P. Genevet, F. Capasso, F. Aieta, M. Khorasaninejad, and “Metasurfaces nanoantennas for light processing,” JOSA B, vol.
R. Devlin, “Recent advances in planar optics: from plasmonic to 30, p. 2365, 2013.
dielectric metasurfaces,” Optica, vol. 4, p. 139, 2017. [44] H. Chen, D. An, Z. Li, and X. Zhao, “Performing differential
[24] O. Hemmatyar, B. Rahmani, A. Bagheri, and A. Khavasi, “Phase operation with a silver dendritic metasurface at visible
resonance tuning and multi-band absorption via graphene- wavelengths,” Optic Express, vol. 25, p. 26417, 2017.
covered compound metallic gratings,” IEEE J. Quantum Electron., [45] H. Chen, D. An, and X. Zhao, “Quasi-periodic dendritic
vol. 53, p. 1, 2017. metasurface for integral operation in visible light,” Molecules,
[25] S. Wang, P. C. Wu, V.-C. Su, et al., “A broadband achromatic vol. 25, p. 1664, 2020.
metalens in the visible,” Nat. Nanotechnol., vol. 13, p. 227, 2018. [46] Y. Wu, Z. Zhuang, L. Deng, Y. Liu, Q. Xue, and Z. Ghassemlooy,
[26] O. Hemmatyar, S. Abdollahramezani, Y. Kiarashinejad, “Arbitrary multi-way parallel mathematical operations based on
M. Zandehshahvar, and A. Adibi, “Full color generation with planar discrete metamaterials,” Plasmonics, vol. 13, p. 599, 2018.
fano-type resonant HfO2 nanopillars designed by a deep- [47] T. Zhu, J. Huang, and Z. Ruan, “Optical phase mining by adjustable
learning approach,” Nanoscale, vol. 11, p. 21266, 2019. spatial differentiator,” Adv. Photonics, vol. 2, p. 016001, 2020.
[27] A. Krasnok, D. Baranov, H. Li, M.-A. Miri, F. Monticone, and A. Alú, [48] Y. Zhou, W. Wu, R. Chen, W. Chen, R. Chen, and Y. Ma, “Analog
“Anomalies in light scattering,” Adv. Opt. Photonics, vol. 11, optical spatial differentiators based on dielectric metasurfaces,”
p. 892, 2019. Adv. Opt. Mater., vol. 8, p. 1901523, 2020.
[28] O. Reshef, I. De Leon, M. Z. Alam, and R. W. Boyd, “Nonlinear [49] A. Cordaro, H. Kwon, D. Sounas, A. F. Koenderink, A. Alu, and
optical effects in epsilon-near-zero media,” Nat. Rev. Mater., vol. A. Polman, “High-index dielectric metasurfaces performing
4, p. 535, 2019. mathematical operations,” Nano Lett., vol. 19, p. 8418, 2019.
[29] C. U. Hail, A.-K. U. Michel, D. Poulikakos, and H. Eghlidi, “Optical [50] H. Kwon, D. Sounas, A. Cordaro, A. Polman, and A. Alù, “Nonlocal
metasurfaces: evolving from passive to adaptive,” Adv. Opt. metasurfaces for optical signal processing,” Phys. Rev. Lett., vol.
Mater., p. 1801786. https://doi.org/10.1002/adom.201801786. 121, p. 173004, 2018.
[30] P. Yu, L. V. Besteiro, Y. Huang, et al., “Broadband metamaterial [51] Y. Fang, Y. Lou, and Z. Ruan, “On-grating graphene surface
absorbers,” Adv. Opt. Mater., vol. 7, p. 1800995, 2019. plasmons enabling spatial differentiation in the terahertz
[31] Z. Wu, M. Zhou, E. Khoram, B. Liu, and Z. Yu “Neuromorphic region,” Opt. Lett., vol. 42, p. 3840, 2017.
metasurface,” Photonics Res., vol. 8, p. 46, 2020. [52] D. A. Bykov, L. L. Doskolovich, E. A. Bezus, and V. A. Soifer,
[32] O. Hemmatyar, M. A. Abbassi, B. Rahmani, M. Memarian, and “Optical computation of the laplace operator using phase-
K. Mehrany, “Wide-band/angle blazed dual mode metallic shifted Bragg grating,” Optic Express, vol. 22, p. 25084, 2014.
groove gratings,” IEEE Trans. Antenn. Propag., vol. 68, p. 1, [53] A. Roberts, D. E. Gómez, and T. J. Davis, “Optical image
2020. processing with metasurface dark modes,” JOSA A, vol. 35,
[33] K. Arik, O. Hemmatyar, and Z. Kavehvash, “Beam manipulation p. 1575, 2018.
by hybrid plasmonic-dielectric metasurfaces,” Plasmonics, [54] M. F. Limonov, M. V. Rybin, A. N. Poddubny, and Y. S. Kivshar,
vol. 15, no. 3, pp. 639–645, 2019. “Fano resonances in photonics,” Nat. Photonics, vol. 11, p. 543,
[34] A. Silva, F. Monticone, G. Castaldi, V. Galdi, A. Alù, and 2017.
N. Engheta, “Performing mathematical operations with [55] Y. Xu, S. Li, B. Hou, and H. Chen, “Fano resonances from
metamaterials,” Science, vol. 343, p. 160, 2014. gradient-index metamaterials,” Sci. Rep., vol. 6, p. 19927, 2016.
[35] F. Monticone, N. M. Estakhri, and A. Alu, “Full control of [56] C. Guo, M. Xiao, M. Minkov, Y. Shi, and S. Fan, “Photonic crystal
nanoscale optical transmission with a composite metascreen,” slab Laplace operator for image differentiation,” Optica, vol. 5,
Phys. Rev. Lett., vol. 110, p. 203903, 2013. p. 251, 2018.
 S. Abdollahramezani et al.: Meta-optics for spatial optical analog computing 4093

[57] A. Saba, M. R. Tavakol, P. Karimi-Khoozani, and A. Khavasi, [76] L. Wang, M. Eliceiri, Y. Deng, et al., “Fast reversible phase change
“Two-dimensional edge detection by guided mode resonant silicon for visible active photonics,” Adv. Funct. Mater. vol. 30,
metasurface,” IEEE Photonics Technol. Lett., vol. 30, p. 853, no. 17, p. 1910784, 2020.
2018. [77] G. Kafaie Shirmanesh, R. Sokhoyan, R. A. Pala, and H. A. Atwater,
[58] S. Peng and G. M. Morris, “Resonant scattering from two- “Dual-gated active metasurface at 1550 nm with wide (> 300°)
dimensional gratings,” JOSA A, vol. 13, p. 993, 1996. phase tunability,” Nano Lett., vol. 18, p. 2957, 2018.
[59] D. A. Bykov, L. L. Doskolovich, A. A. Morozov, et al., “First-order [78] G. K. Shirmanesh, R. Sokhoyan, P. C. Wu, and H. A. Atwater,
optical spatial differentiator based on a guided-mode resonant Electro-Optically Tunable Universal Metasurfaces, 2019, arXiv
grating,” Optic Express, vol. 26, p. 10997, 2018. preprint arXiv:1910.02069.
[60] F. Zangeneh-Nejad and A. Khavasi, “Spatial integration by a [79] S.-Q. Li, X. Xu, R. M. Veetil, V. Valuckas, R. Paniagua-Domínguez,
dielectric slab and its planar graphene-based counterpart,” Opt. and A. I. Kuznetsov, “Phase-only transmissive spatial light
Lett., vol. 42, p. 1954, 2017. modulator based on tunable dielectric metasurface,” Science,
[61] Y. Zhou, H. Zheng, I. I. Kravchenko, and J. Valentine, “Flat optics vol. 364, p. 1087, 2019.
for image differentiation,” Nat. Photonics, vol. 14, p. 316, 2020. [80] M. Taghinejad, H. Taghinejad, Z. Xu, et al., “Ultrafast control of
[62] C. Guo, M. Xiao, M. Minkov, Y. Shi, and S. Fan, “Isotropic phase and polarization of light expedited by hot-electron
wavevector domain image filters by a photonic crystal slab transfer,” Nano Lett., vol. 18, p. 5544, 2018.
device,” JOSA A, vol. 35, p. 1685, 2018. [81] A. Tittl, A.-K. U. Michel, M. Schäferling, et al., “A
[63] E. A. Bezus, L. L. Doskolovich, D. A. Bykov, and switchable mid-infrared plasmonic perfect absorber with
V. A. Soifer, “Spatial integration and differentiation of multispectral thermal imaging capability,” Adv. Mater., vol.
optical beams in a slab waveguide by a dielectric ridge 27, p. 4597, 2015.
supporting high-q resonances,” Optic Express, vol. 26, [82] P. Li, X. Yang, T. W. Maß, et al., “Reversible optical switching of
p. 25156, 2018. highly confined phonon–polaritons with an ultrathin phase-
[64] A. Youssefi, F. Zangeneh-Nejad, S. Abdollahramezani, and change material,” Nat. Mater., vol. 15, p. 870, 2016.
A. Khavasi, “Analog computing by brewster effect,” Opt. Lett., [83] S. Abdollahramezani, H. Taghinejad, T. Fan, Y. Kiarashinejad, A.
vol. 41, p. 3467, 2016. A. Eftekhar, and A. Adibi, Reconfigurable Multifunctional
[65] T. Zhu, Y. Lou, Y. Zhou, et al., “Generalized spatial differentiation Metasurfaces Employing Hybrid Phase-Change Plasmonic
from the spin hall effect of light and its application in image Architecture, 2018, arXiv preprint arXiv:1809.08907.
processing of edge detection,” Phys. Rev. Appl., vol. 11, [84] M. Wuttig, H. Bhaskaran, and T. Taubner, “Phase-change
p. 034043, 2019. materials for non-volatile photonic applications,” Nat.
[66] J. Zhang, Q. Ying, and Z. Ruan, “Time response of plasmonic Photonics, vol. 11, p. 465, 2017.
spatial differentiators,” Opt. Lett., vol. 44, p. 4511, 2019. [85] F. Ding, Y. Yang, and S. I. Bozhevolnyi, “Dynamic metasurfaces
[67] Y. Lou, Y. Fang, and Z. Ruan, Optical Computation of Divergence using phase-change chalcogenides,” Adv. Opt. Mater., vol. 7,
Operation for Vector Field, 2020, p. 10649, arXiv preprint arXiv: no. 14, p. 1801709, 2019.
2003. [86] S. Abdollahramezani, O. Hemmatyar, H. Taghinejad, et al.,
[68] A. Babaee, A. Momeni, A. Abdolali, and R. Fleury, Parallel Optical “Tunable nanophotonics enabled by chalcogenide phase-
Computing Based on Mimo Metasurface Processors with change materials,” Nanophotonics, vol. 9, p. 1189, 2020.
Asymmetric Optical Response, 2020, arXiv preprint arXiv: [87] H. Taghinejad, S. Abdollahramezani, A. A. Eftekhar, et al.,
2004.02948. Ito-based Mu-Heaters for Multi-Stage Switching of Phase-
[69] A. Abdolali, A. Momeni, H. Rajabalipanah, and K. Achouri, Change Materials: Towards Beyond-Binary Reconfigurable
“Parallel integro-differential equation solving via multi-channel Integrated Photonics, 2020, arXiv preprint arXiv:
reciprocal bianisotropic metasurface augmented by normal 2003.04097.
susceptibilities,” New J. Phys., vol. 21, p. 113048, 2019. [88] Y. Zhang, C. Fowler, J. Liang, et al., Electrically Reconfigurable
[70] H. Kwon, A. Cordaro, D. Sounas, A. Polman, and A. Alu, “Dual- Nonvolatile Metasurface Using Low-Loss Optical Phase Change
polarization analog 2D image processing with nonlocal Material, 2020, arXiv preprint arXiv:2008.06659.
metasurfaces,” ACS Photonics, vol. 7, pp. 1799–1805, 2020. [89] S. Abdollahramezani, O. Hemmatyar, M. Taghinejad, et al.,
[71] L. Wan, D. Pan, S. Yang, et al., “Optical analog computing of Dynamic Hybrid Metasurfaces, 2020, arXiv preprint arXiv:
spatial differentiation and edge detection with dielectric 2008.03905.
metasurfaces,” Optic Lett., vol. 45, p. 2070, 2020. [90] S. T. Ha, Y. H. Fu, N. K. Emani, et al., “Directional lasing in
[72] F. Zangeneh-Nejad, A. Khavasi, and B. Rejaei, “Analog optical resonant semiconductor nanoantenna arrays,” Nat.
computing by half-wavelength slabs,” Optic Commun., vol. 407, Nanotechnol., vol. 13, p. 1042, 2018.
p. 338, 2018. [91] R.-M. Ma and R. F. Oulton, “Applications of nanolasers,” Nat.
[73] N. Hermosa, A. Nugrowati, A. Aiello, and J. Woerdman, “Spin hall Nanotechnol., vol. 14, p. 12, 2019.
effect of light in metallic reflection,” Opt. Lett., vol. 36, p. 3200, [92] A. Vaskin, R. Kolkowski, A. F. Koenderink, and I. Staude, “Light-
2011. emitting metasurfaces,” Nanophotonics, vol. 8, p. 1151, 2019.
[74] N. M. Estakhri, B. Edwards, and N. Engheta, “Inverse-designed [93] W. T. Chen, A. Y. Zhu, and F. Capasso, “Flat optics with
meta-structures that solve equations,” Science, vol. 363, dispersion-engineered metasurfaces,” Nat. Rev. Mater., vol. 1,
p. 1333, 2019. pp. 1–17, 2020b.
[75] M. Taghinejad and W. Cai, “All-optical control of light in [94] S. Colburn, A. Zhan, and A. Majumdar, “Metasurface optics for
micro-and nanophotonics,” ACS Photonics, vol. 6, p. 1082, full-color computational imaging,” Science Adv., vol. 4,
2019. p. eaar2114, 2018.
4094 S. Abdollahramezani et al.: Meta-optics for spatial optical analog computing

[95] Q. Jiang, G. Jin, and L. Cao, “When metasurface meets Optimization of Photonic Devices, 2020, arXiv preprint
hologram: principle and advances,” Adv. Optic Photonics, vol. arXiv:2007.02205.
11, p. 518, 2019. [114] S. An, B. Zheng, M. Y. Shalaginov, et al., A Freeform Dielectric
[96] S. Chen, W. Liu, Z. Li, H. Cheng, and J. Tian, “Metasurface- Metasurface Modeling Approach Based on Deep Neural
empowered optical multiplexing and multifunction,” Adv. Networks, 2020, arXiv preprint arXiv:2001.00121.
Mater., vol. 32, p. 1805912, 2020. [115] D. Sell, J. Yang, S. Doshay, and J. A. Fan, “Periodic dielectric
[97] S. Chen, Z. Li, W. Liu, H. Cheng, and J. Tian, “From single- metasurfaces with high-efficiency, multiwavelength
dimensional to multidimensional manipulation of optical functionalities,” Adv. Opt. Mater., vol. 5, p. 1700645, 2017.
waves with metasurfaces,” Adv. Mater., vol. 31, p. 1802458, [116] D. Zhu, Z. Liu, L. Raju, A. S. Kim, and W. Cai, Multifunctional
2019. Meta-Optic Systems: Inversely Designed with Artificial
[98] P. Karimi, A. Khavasi, and S. S. M. Khaleghi, “Fundamental limit Intelligence, 2020, arXiv preprint arXiv:2007.00130.
for gain and resolution in analog optical edge detection,” [117] Y. Elesin, B. S. Lazarov, J. S. Jensen, and O. Sigmund, “Design of
Optics Express, vol. 28, p. 898, 2020. robust and efficient photonic switches using topology
[99] W. Zhang and X. Zhang, “Backscattering-immune computing of optimization,” Photonics Nanostruct. Fund. Appl., vol. 10,
spatial differentiation by nonreciprocal plasmonics,” Phys. Rev. p. 153, 2012.
Appl., vol. 11, p. 054033, 2019. [118] A. Y. Piggott, Automated Design of Photonic Devices. Stanford
[100] J. S. Jensen and O. Sigmund, “Topology optimization for nano- University, USA: Alexander Yukio Piggott, 2018.
photonics,” Laser Photonics Rev., vol. 5, p. 308, 2011. [119] S. S. Kou, G. Yuan, Q. Wang, et al., “On-chip photonic Fourier
[101] S. Molesky, Z. Lin, A. Y. Piggott, W. Jin, J. Vucković, and transform with surface plasmon polaritons,” Light Sci. Appl.,
A. W. Rodriguez, “Inverse design in nanophotonics,” Nat. vol. 5, p. e16034, 2016.
Photonics, vol. 12, p. 659, 2018. [120] Q. H. Wang, K. Kalantar-Zadeh, A. Kis, J. N. Coleman, and
[102] S. D. Campbell, D. Sell, R. P. Jenkins, E. B. Whiting, M. S. Strano, “Electronics and optoelectronics of two-
J. A. Fan, and D. H. Werner, “Review of numerical dimensional transition metal dichalcogenides,” Nat.
optimization techniques for meta-device design,” Opt. Nanotechnol., vol. 7, p. 699, 2012.
Mater. Express, vol. 9, p. 1842, 2019. [121] K. Novoselov, A. Mishchenko, A. Carvalho, and A. C. Neto, “2D
[103] D. Liu, Y. Tan, E. Khoram, and Z. Yu, “Training deep neural materials and van der waals heterostructures,” Science, vol.
networks for the inverse design of nanophotonic structures,” 353, 2016, https://doi.org/10.1126/science.aac9439.
ACS Photonics, vol. 5, p. 1365, 2018. [122] M. Taghinejad, Z. Xu, H. Wang, et al., “Photocarrier-induced
[104] X. Lin, Y. Rivenson, N. T. Yardimci, et al., “All-optical machine active control of second-order optical nonlinearity in monolayer
learning using diffractive deep neural networks,” Science, vol. MOS2,” Small, vol. 16, p. 1906347, 2020.
361, p. 1004, 2018. [123] J. van de Groep, J.-H. Song, U. Celano, Q. Li, P. G. Kik, and M. L.
[105] Z. Liu, D. Zhu, S. P. Rodrigues, K.-T. Lee, and W. Cai, “Generative Brongersma, “Exciton resonance tuning of an atomically thin
model for the inverse design of metasurfaces,” Nano Lett., vol. lens,” Nat. Photonics, vol. 1, 2020. https://doi.org/10.1038/
18, p. 6570, 2018. s41566-020-0624-y.
[106] A. Zhan, R. Gibson, J. Whitehead, E. Smith, J. R. Hendrickson, [124] S. Divitt, W. Zhu, C. Zhang, H. J. Lezec, and A. Agrawal, “Ultrafast
and A. Majumdar, “Controlling three-dimensional optical optical pulse shaping using dielectric metasurfaces,” Science,
fields via inverse mie scattering,” Sci. Adv., vol. 5, vol. 364, p. 890, 2019.
p. eaax4769, 2019. [125] W. Liu, M. Li, R. S. Guzzon, et al., “A fully reconfigurable
[107] J. Jiang, D. Sell, S. Hoyer, J. Hickey, J. Yang, and J. A. Fan, “Free- photonic integrated signal processor,” Nat. Photonics, vol. 10,
form diffractive metagrating design based on generative p. 190, 2016.
adversarial networks,” ACS Nano, vol. 13, p. 8872, 2019. [126] R. Slavík, Y. Park, N. Ayotte, et al., “Photonic temporal
[108] Y. Kiarashinejad, S. Abdollahramezani, M. Zandehshahvar, integrator for all-optical computing,” Optic Express, vol. 16,
O. Hemmatyar, and A. Adibi, “Deep learning reveals underlying p. 18202, 2008.
physics of light-matter interactions in nanophotonic devices,” [127] K. Goda and B. Jalali, “Dispersive Fourier transformation for fast
Adv. Theory Simul., vol. 2, p. 1900088, 2019. continuous single-shot measurements,” Nat. Photonics, vol. 7,
[109] E. Khoram, A. Chen, D. Liu, et al., “Nanophotonic media for p. 102, 2013.
artificial neural inference,” Photonics Res., vol. 7, p. 823, 2019. [128] M. A. Foster, R. Salem, D. F. Geraghty, A. C. Turner-Foster,
[110] Y. Kiarashinejad, M. Zandehshahvar, S. Abdollahramezani, M. Lipson, and A. L. Gaeta, “Silicon-chip-based ultrafast optical
O. Hemmatyar, R. Pourabolghasem, and A. Adibi, “Knowledge oscilloscope,” Nature, vol. 456, p. 81, 2008.
discovery in nanophotonics using geometric deep learning,” Adv. [129] S. Colburn, Y. Chu, E. Shilzerman, and A. Majumdar, “Optical
Intell. Syst., vol. 2, p. 1900132, 2019. frontend for a convolutional neural network,” Appl. Optics, vol.
[111] J. Jiang, M. Chen, and J. A. Fan, Deep Neural Networks for the 58, p. 3179, 2019.
Evaluation and Design of Photonic Devices, 2020, arXiv preprint [130] X. Sui, Q. Wu, J. Liu, Q. Chen, and G. Gu, “A review of optical
arXiv:2007.00084. neural networks,” IEEE Access, vol. 8, p. 70773, 2020.
[112] Y. Kiarashinejad, S. Abdollahramezani, and A. Adibi, “Deep [131] J. Moughames, X. Porte, M. Thiel, et al., “Three-dimensional
learning approach based on dimensionality reduction for waveguide interconnects for scalable integration of photonic
designing electromagnetic nanostructures,” NPJ Comput. neural networks,” Optica, vol. 7, p. 640, 2020.
Mater., vol. 6, p. 1, 2020. [132] Q. Zhang, H. Yu, M. Barbiero, B. Wang, and M. Gu, “Artificial
[113] Z. A. Kudyshev, A. V. Kildishev, V. M. Shalaev, and A. neural networks enabled by nanophotonics,” Light Sci. Appl.,
Boltasseva, Machine Learning Assisted Global vol. 8, p. 1, 2019.
 S. Abdollahramezani et al.: Meta-optics for spatial optical analog computing 4095

[133] X. Ding, Z. Wang, G. Hu, et al., “Metasurface holographic image [137] F. Zangeneh-Nejad and R. Fleury, “Performing mathematical
projection based on mathematical properties of fourier operations using high-index acoustic metamaterials,” New J.
transform,” PhotoniX, vol. 1, p. 1, 2020. Phys., vol. 20, p. 073001, 2018.
[134] W. Zhang, K. Cheng, C. Wu, Y. Wang, H. Li, and X. Zhang, [138] F. Zangeneh-Nejad and R. Fleury, “Topological analog signal
“Implementing quantum search algorithm with processing,” Nat. Commun., vol. 10, p. 1, 2019.
metamaterials,” Adv. Mater., vol. 30, p. 1703986, 2018. [139] S.-Y. Zuo, Q. Wei, Y. Cheng, and X.-J. Liu, “Mathematical
[135] K. Cheng, W. Zhang, Z. Wei, et al., “Simulate Deutsch–Jozsa operations for acoustic signals based on layered labyrinthine
algorithm with metamaterials,” Optics Express, vol. 28, metasurfaces,” Appl. Phys. Lett., vol. 110, p. 011904, 2017.
p. 16230, 2020. [140] S. Zuo, Q. Wei, Y. Tian, Y. Cheng, and X. Liu, “Acoustic analog
[136] H. Rajabalipanah, A. Abdolali, S. Iqbal, L. Zhang, and T. J. Cui, computing system based on labyrinthine metasurfaces,” Sci.
“How do space-time digital metasurfaces serve to perform Rep., vol. 8, p. 1, 2018.
analog signal processing?,” 2020, arXiv preprint arXiv: [141] A. A. Sawchuk and T. C. Strand, “Digital optical computing,”
2002.06773. Proc. IEEE, vol. 72, p. 758, 1984.