REVIEWS

Analogue computing with

metamaterials
Farzad Zangeneh-​Nejad   1, Dimitrios L. Sounas2, Andrea Alù   3 and Romain Fleury   1 ✉
Abstract | Despite their widespread use for performing advanced computational tasks, digital
signal processors suffer from several restrictions, including low speed, high power consumption
and complexity, caused by costly analogue-​to-​digital converters. For this reason, there has
recently been a surge of interest in performing wave-​based analogue computations that avoid
analogue-​to-​digital conversion and allow massively parallel operation. In particular, novel
schemes for wave-​based analogue computing have been proposed based on artificially
engineered photonic structures, that is, metamaterials. Such kinds of computing systems,
referred to as computational metamaterials, can be as fast as the speed of light and as small as
its wavelength, yet, impart complex mathematical operations on an incoming wave packet or
even provide solutions to integro-​differential equations. These much-​sought features promise to
enable a new generation of ultra-​fast, compact and efficient processing and computing hardware
based on light-​wave propagation. In this Review, we discuss recent advances in the field of
computational metamaterials, surveying the state-​of-​the-​art metastructures proposed to
perform analogue computation. We further describe some of the most exciting applications
suggested for these computing systems, including image processing, edge detection, equation
solving and machine learning. Finally, we provide an outlook for the possible directions and the
key problems for future research.

Nowadays, digital signal processors (DSPs) are ubiqui- to perform discretization properly because the signals
tously used to carry out a wide variety of computational vary too rapidly. These limitations have led to a surge
tasks, from relatively simple to highly complex ones1. of interest in revisiting the idea of analogue comput-
DSPs generally consist of three basic components: an ing, which may be traced back to several decades ago,
analogue-​to-​digital converter (A/D), a processing unit in which signal processing happens in the analogue
and, often, a digital-​to-​analogue converter (D/A). The domain and, hence, the analogue-​to-​digital conversion
first sub-​block (A/D) takes an analogue signal (repre- is not needed.
senting, for instance, an image or voice) and discretizes it An analogue computer is a device that takes advan-
into a series of bits. The second sub-​block, the processing tage of continuous variations in a given physical phe-
1
Laboratory of Wave unit, manipulates the discretized version of the analogue nomenon to perform a certain computational or
Engineering, School of signal according to the desired computational operation. processing task. The first electronic or mechanical ana-
Electrical Engineering, The resulting digital stream can then be converted back logue computers, originally preferred to digital versions,
Swiss Federal Institute of
into the analogue domain using the D/A sub-​block. were based on continuously varying quantities, such as
Technology in Lausanne
(EPFL), Lausanne,
Despite their versatile functionality, DSPs have electric current or mechanical motion3 (Box 1). Despite
Switzerland. several drawbacks, most of which arise from the una- being free of A/D and D/A converters, such kinds of
2
Department of Electrical voidable analogue-​to-​digital conversion. In particular, computers turned out to be very slow and bulky, hin-
Engineering, Wayne State the A/D and D/A sub-​blocks of DSPs tend to consume dering their applicability in modern systems, in which
University, Detroit, MI, USA. significant power. This renders DSPs cost-​inefficient, high speed and miniaturization are sought. In addition,
3
Photonics Initiative, especially when it comes to performing simple com- small errors triggered by noise were found to propagate
Advanced Science Research putational tasks such as differentiation or integration2. and be amplified, as the signals were processed in series.
Center, City University of
New York, New York, NY, USA.
The A/D and D/A converters also restrict the speed Digital computers could overcome these challenges, and
✉e-​mail: romain.fleury@ of processing, given that the discretization process is took over the scene.
epfl.ch time-​consuming and cannot be performed in a mas- Recently, the interest in analogue computing was
https://doi.org/10.1038/ sively parallel manner. Moreover, at high frequencies revived in the context of metamaterials research, as
s41578-020-00243-2 (above the GHz range), the A/D and D/A converters fail it was shown that subwavelength structures could

NaTure RevIewS | MAteRiAls volume 6 | March 2021 | 207

Reviews

implement computing functionalities by leverag- platform is indeed based on the enhanced interaction
ing light propagation in suitably engineered artificial of optical fields with artificially created structures, the
photo­nic materials4. This solution enables ultra-​fast metamaterials5–18. As opposed to conventional electronic
speeds, low loss, subwavelength form factors and mas- and mechanical computers, computational photo­nic
sively parallel operations, holding the promise to over- metamaterials can be very fast. This is because they
come the aforementioned challenges. This computing operate at the speed of light and, more importantly, are

Box 1 | History of analogue computing

The earliest known analogue a b
computer, invented in Greece
between 150 and 100 BC,
is the Antikythera
mechanism135,136, intended for
specific astronomical
applications (see the figure,
panel a). The device was
composed of nearly 40 gears
and wheels, and was devised
to model the position of the
Moon and the Sun in their
orbits, providing the possibility
of predicting eclipses.
Similar mechanical analogue
computers were later
developed for astronomical c d
purposes. A prominent, more R
recent example of such a
device is the astronomical
clock137 that appeared in the C
Vin
14th century, which was Vout
capable of analysing complex
astronomical phenomena,
such as the relative positions
of the Earth, Sun, Moon and
planets.
Mechanical analogue
computers have also been C
used for computational R
purposes other than
Vin
astronomical calculations. Vout
One of the mechanical
computing devices commonly
used since the 17th century
were side rulers138, performing
multiplications and divisions139 (see the figure, panel b). Such kinds of computers, developed shortly after the description of logarithms, worked based
on the fact that the multiplication (or division) of two real numbers can be expressed as the addition (or subtraction) of their logarithms.
Before the introduction of portable electronic computers, many other mechanical computers were developed. Examples include: the planimeter, used
to calculate the area within a closed 2D shape139; the tide predictor140 (developed in 1878), a mechanical analogue computer predicting the behaviour
of sea tides; the Dumaresq135, an analogue computer developed in 1902 to relate the parameters of fire-​control systems to the ones of a moving target
object; the nomograph141 (invented in 1918), a special mechanical equation solver solving a specific class of equations having the form f(x,y,z) = 0;
and the differential analyser135 (invented in 1930), which could solve second-​order and higher-​order differential equations.
Despite their simple principle of operation, mechanical analogue computers suffer from several restrictions, including their large size and high
production cost. Driven by advances in electronics, electronic analogue computers emerged. The underlying principle of electronic analogue
computers is that, by engineering the connectivity between the circuit elements, the transfer function of a circuit can be tailored to follow that
of a mathematical operation. The top part of the figure, panel c, for example, represents an electronic circuit realizing the operator of first-​order
differentiation. By employing simple circuit-​analysis techniques, the transfer function of the circuit can be found as Vout/Vin = iωRC (where Vout is the
output voltage, Vin the input voltage, ω the frequency, R the resistance and C the capacitance). In the time domain, the relation between Vout and Vin is
expressed as Vout = −RCdVin/dt, indicating that the circuit acts as a first-​order differentiator. The bottom part of panel c illustrates another electronic
circuit, which acts as an analogue integrator.
The first generation of electronic analogue computers was developed in the middle of 20th century. For example, the computer shown in the figure,
panel d, called AKAT-1, was composed of several operational amplifiers, transistors and regular passive elements. The system, developed in 1959,
was designed to address complex dynamic processes, such as heat transfer, by analysing the associated differential equations. Compared with their
mechanical ancestors, electronic analogue computers were much faster and smaller, and could be adapted to a wider range of analogue functionalities.
Their early adoption was abandoned in favour of digital computers, which were more robust to noise.
Figure, panel a, image from CPA Media Pte Ltd/Alamy Stock Photo; panel b, image from Anton Starikov/Alamy Stock Photo; panel d, image courtesy
Tablet magazine.

208 | March 2021 | volume 6 www.nature.com/natrevmats

 Reviews

capable of performing a large number of operations in the edges of the image and keeping its slower variations.
parallel. At the same time, the very small wavelength of This optical system, known as a 4F correlator19 (the
optical waves enables miniaturization and integration. name comes from the fact that the system is four focal
These features, hence, create the ideal conditions for car- lengths long), can, therefore, apply a large number of lin-
rying out specific-​purpose signal-​processing tasks with ear operations to the Fourier spectrum of an image. The
ultra-​fast speed and massive parallelization, at scales science of Fourier optics, indeed, consists in tailoring
potentially smaller than the wavelength. the local transmission amplitude and phases of the
In this Review, we discuss recent advances in this Fourier mask to achieve various advanced functions.
thriving area of research. We start by describing dif- Compared with conventional electronic analogue
ferent approaches proposed for wave-​based analogue computers (Box 1), analogue signal processors based
computing, including the Fourier optics approach19,20, on Fourier optics are much faster, because the speed of
the Green’s function metamaterial approach4 and the light is much larger than the drift velocity of electrons.
metasurface approach4. We then outline the designs Yet, their bulky structures, involving at least four focal
and unusual properties of state-​of-​the-​art computational lengths, broadly hinder their miniaturization. Alignment
metamaterials, proposed in various areas of wave phys- issues and aberrations caused by the realistic features of
ics, such as photonics and phononics. Afterwards, we the lenses further complicate the picture. Despite all
discuss recent developments and applications of compu- of these challenges, Fourier optics is a well-​established
tational metamaterials in modern engineering, such as field of science and technology19.
equation solving, machine learning and topological ana- Electromagnetic metamaterials, compound mate-
logue signal processing. In the last part of the Review, we rials made of artificial scatterers designed to achieve
provide an outlook for possible future directions, includ- desired macroscopic properties, and their 2D versions,
ing multifunctional computational metamaterials, intel- electromagnetic metasurfaces, can shrink dramatically
ligent metamaterial computing systems and wave-​based the size of these processing systems and avoid the need
analogue signal processors based on disorder. for Fourier-​transforming the image twice. Like analogue
signal processors based on Fourier lenses, metamaterial
Analogue computing principles computing systems can be superfast (because they are
Fourier optics based on propagating waves). However, as opposed to
Principle of operation. The idea of leveraging waves to conventional Fourier optics signal processors, compu-
perform analogue computing dates back to several dec- tational metamaterials and metasurfaces can be made
ades before the development of computational metama- smaller than the operation wavelength, because they
terials. In fact, a simple (convex) lens acts as a Fourier usually rely on subwavelength resonant scatterers. Two
transformer on an image placed in its focal plane19–22, approaches have been introduced to computational
transforming a constant illumination, in the form of a metamaterials: the Green’s function approach and the
uniform, monochromatic plane-​wave incident field with metasurface approach, as detailed in the following.
planes of constant phase perpendicular to the lens axis,
into a single dot at its focal length, which approaches Green’s function approach
a delta Kronecker function in the limit of a lens with Principle. The Green’s function method4,23–38 has guided
infinite aperture. Conversely, the emission from a point-​ several designs of metamaterial computing systems. As
like source is ideally converted into a uniform plane wave the name of the method denotes, in the Green’s function
by the lens. The ability of lenses to take the Fourier trans- approach, the Green’s function of the operator of choice
form of incident fields is at the basis of Fourier optics, is directly realized in real space, without transforming
which is the simplest example of a platform for using back and forth from the spatial to the spectral domain,
light to perform signal-​processing tasks. gaining in compactness and avoiding possible challenges
The general principle of Fourier optics is to con- in error propagation and alignment issues. Let us con-
vert a signal to the Fourier space, perform the signal-​ sider a hypothetical linear system (Fig. 1a), which acts on
processing operation on the Fourier-​transformed signal an input signal f and turns it in into the output g. The
and then convert the output back to the regular space. signals f and g could be, for example, the optical fields
As an example, let us suppose that an arbitrary image that are incident on and transmitted through a metama-
is placed in the front focal plane of an optical lens. At terial. For simplicity, we assume here that the fields only
the back focal plane of the lens, the 2D spatial Fourier vary with respect to a single spatial coordinate y. Because
transform of the corresponding image is generated. our system should apply an operation on this incoming
At this plane, the spectral features of the image can be 1D function independent of its specific y variations, we
manipulated using a mask plate with a specific trans- require that the metamaterial properties are invariant
verse transparency pattern. If, for instance, we cover the upon translations along y, whereas the input and out-
centre of the Fourier plane with an opaque mask plate, put signals, f(y) and g(y), explicitly depend on y. From a
and then use a second lens to inverse-​Fourier-​transform system theory point of view, f(y) and g(y) are related to
the image, the low-​order Fourier components will be each other through the Green’s function of the metama-
suppressed, but the information associated with the terial, defined in the Fourier space as H(ky) = FT[g(y)]/
higher-​order harmonics will be maintained, enhancing FT[f(y)], in which FT stands for the Fourier transform.
the edges. Similarly, a pinhole mask placed at the centre By engineering the metamaterial properties, we can
of the Fourier plane attenuates the information associ- actually tailor the associated Green’s function H(ky)
ated with the high-​order Fourier components, blurring to match the transfer function of the desired operator.

NaTure RevIewS | MAteRiAls volume 6 | March 2021 | 209

Reviews

This design can be performed either by direct optimiza- composed by a stack of subwavelength metamaterial
tion or by inverse design, using, for instance, generalized slabs. Each layer has a specific thickness (di), permittiv-
sheet transition conditions39 in the case of metasurface ity (εi) and permeability (μi). By running an optimiza-
design. To provide a concrete example, let us consider the tion on the parameters di, εi and μi, it is possible to make
irregularly shaped transfer function H0(ky), whose ampli- the Green’s function of the multilayered structure to be
tude is shown in Fig. 1b. The transfer function is associ- approximately equal to the transfer function of the oper-
ated with an arbitrarily chosen mathematical operation. ator of choice, that is, H0(ky). In this case, in real space,
The Green’s function method can be employed to imple- the optimized metamaterial acts as an analogue signal
ment this transfer function in a composite metamaterial, processor, applying the desired mathematical operation
for instance, the multilayer structure shown in Fig. 1c, in the spatial domain to any incident signal f(y).

a b 1 c f1(y)

f(y) g(y) H0(ky) df1(y)

Linear operator H(ky)
f2(y) dy

df2(y)
dy
0
–1 ky/k0 1

d e 1 f 1 f(y)
x H(ky)
Ideal
f(y)
x˝
0
H0(ky)

z df(y)
y θ dy 1 g(y) Ideal
n
θ
y˝ BW 0
z˝
–1 –1
–1 1 –40 40
ky/k0 y/λ0

g h i
1 1
Prism Prism H(ky) f(y)
Ideal 0
g(y) = ∫f(y)
H0(ky)

–1
1 g(y)
Ideal
θ
f(y)
n2 n1 n2 n1 n2
0 0
–0.2 0.2 –40 40
ky/k0 y/λ0

Fig. 1 | Wave-based analogue computing based on the Green’s function region, the Green’s function can be approximated with a linear function (the
approach. a | A hypothetical analogue computing system characterized by dashed line), following the transfer function of the ideal differentiator.
the Green’s function H(ky). b | Transfer function (amplitude) of an arbitrary f | Corresponding differentiated reflected field (bottom panel) when the
linear operator to be realized using the Green’s function method. c | Example interface is illuminated with a Gaussian incident field (top panel). g | An
of a computational metamaterial based on the Green’s function method. The analogue spatial integrator, based on resonant tunnelling through a
metamaterial consists of multilayered dielectric slabs. By optimizing dielectric slab waveguide. n1 and n2 are the refractive indices of the core and
the permittivity, permeability and thickness of each slab, it is possible to cladding layers, respectively. h | At a specific incident angle, the momentum
engineer the Green’s function of the structure such that it matches the of the incident field matches the one of the guided mode, leading to a
desired transfer function, H0(ky). d | Analogue spatial differentiator based on resonant tunnelling peak in the transmission spectrum of the structure.
the Green’s function method. The differentiator consists only of an interface The spectral line shape of this resonance peak is the same as the one of
between a dielectric and free space, on which a transverse-​magnetic-​ an ideal integrator (the dashed line). i | Demonstration of the operation
polarized incident beam impinges at the Brewster angle. e | Green’s function of the integrator. Panel c is adapted with permission from ref.4, AAAS.
(reflection spectrum) of the Brewster differentiator near the Brewster angle. Panels d–f adapted with permission from ref.40, © The Optical Society.
For sufficiently wide incident signals with bandwidth (BW) within the yellow Panels g and h adapted with permission from ref.42, © The Optical Society.

210 | March 2021 | volume 6 www.nature.com/natrevmats

 Reviews

Implementation. The Green’s function method has been in Fig. 1g, the well-​known dielectric slab waveguide
used to perform specific-​purpose computing operations geometry, whose core and cladding layers have refrac-
in recent years. One of the most popular mathematical tive indices of n2 = 3.4 and n1 = 1.5, respectively. The
operations implemented using this method, which is structure is excited from the far field using a prism cou-
commonly used to determine sharp variations, edges pler at incident angle θ. At some specific θ, the trans-
and extremal values of signals, is differentiation. A sim- verse momentum of the incoming beam f(y) becomes
ple optical structure that can be used to calculate the equal to the one of the guided mode of the slab wave-
derivative of incident waves under certain conditions40 guide, leading to a resonance peak in the transmission
is shown in Fig. 1d. This basic differentiator consists spectrum (Fig.  1h) . Around this resonance, we can
of an interface between two dielectrics with different approximate the transmission coefficient of the system
refractive indices (n1 = 1 and n2 = 3.4). A transverse-​ with T(ky) = A/ky, in which A is a constant depending
magnetic (TM)-​polarized incident field impinges on on the quality factor of the resonance. This approxi-
the interface. The TM Fresnel reflection coefficient mation is acceptable only for signals that satisfy two
of the interface for the incident angle θ is expressed as conditions: first, the incident field must have a suffi-
ciently small spectral bandwidth, because, away from
n 1 1 − (n 1/n 2sin θ )2 − n 2 cos θ the transparency condition, the transmission is zero
R(θ ) = . (1) and no output is generated. Second, the incident field
n 1 1 − (n 1/n 2sin θ )2 + n 2 cos θ must only possess non-​zero spatial-​f requency com-
ponents. Indeed, integrating a signal with a non-​zero
At the incident angle θB = tan−1(n2/n1), known as the spatial-​frequency component requires amplifying the
Brewster angle41, the TM reflection coefficient of the struc- output, because the transfer function of the ideal inte-
ture vanishes: this is the case represented in Fig. 1d. It is grator T(ky) = A/ky blows up at ky = 0. Such operation
easy to verify that the relation between the Brewster angle, is, therefore, impossible in a passive device for which
the angular parameter θ and the wavenumber ky is the transmission is always below one, unless one only
works with signals without any component at ky = 0.
k y = k 0 sin(θ − θ B ) (2) Under these conditions, T(ky) is similar, up to a constant
factor, to the transfer function of the ideal integrator
in which θ varies between 0 and π/2. Note that the (G(ky) = 1/iky). Hence, such a simple structure func-
coordinate system for the signal is not aligned with tions like an analogue integrator. This is illustrated in
the inter­face, as represented in Fig.1d. Equation 2 defines Fig. 1i, which shows the transmitted field (bottom panel)
a one-​to-​one mapping between the angular parameter θ corresponding to a Gaussian derivative incident signal
and the wavenumber ky, based on which the Fresnel (which, indeed, has no zero-​spatial-​frequency com-
reflection coefficient of Eq. 1 can be transformed into ponent, top panel). As expected, the transmitted field
the spatial Fourier domain. The corresponding spa- has a Gaussian profile. We note that the structures in
tial Fourier spectrum of the Fresnel reflection coef- Fig. 1d,g are not metamaterials per se and, as such, can-
ficient around the Brewster angle is shown in Fig. 1e. not be used for realizing arbitrarily complex comput-
As observed, the reflection spectrum of the interface ing operations. Yet, these examples provide an intuitive
becomes equal to zero at ky = 0 (note that this wave- idea of the underlying principle of the Green’s function
number corresponds to θ = θB, Eq. 2). Near this zero, the method, which is directly applicable to metamaterial
reflection spectrum can be approximated with a linear platforms.
function of the form R(ky) ≈ Aky (Taylor expansion), in The described analogue differentiator and integrator
which A = −(n 2/2 − 1/2n 23)/k 0 (the slope of the dashed can be employed as building blocks to construct more
line in Fig. 1e). Interestingly, R(ky) is very similar to complex analogue operations. Suppose, for instance, that
the Green’s function of the ideal spatial differentiator, we want to realize an analogue second-​order differenti-
namely, H(ky) = iky, other than a proportionality coef- ator. This can be readily accomplished by cascading two
ficient. This implies that, for signals impinging on the realizations of first-​order differentiators, each of which
interface at the Brewster angle with a sufficiently narrow differentiates the incident signal one time. Similarly,
spatial spectrum around that angle (the yellow-​shaded a second-​order integrator can be realized by cascading
region in Fig. 1e), the reflected field is the derivative two first-​order integrators. We stress that these two
of the incident field. This is demonstrated in Fig. 1f, in practical examples are limited in the extent of images
which the reflected field (g(y), bottom panel) corres­ that can be processed (only images with slow varia-
ponding to a Gaussian incident signal ( f(y), top panel) tions impinging from specific angles) and in the overall
is shown. g(y) has a Gaussian derivative profile, evidenc- efficiency of the output. Yet, the general platform used
ing the proper operation of a differentiator, assuming for the Green’s function method, as in Fig. 1c, in which
that the image does not have too large spectral features. optimized multilayers can arbitrarily tailor the Green’s
It should be noted that, because the differentiator works function output H(ky), hold the promise to implement
near a zero of reflection, the amplitude of the derivative more sophisticated and efficient analogue processing
signal (the reflected field) is generally small, leading metamaterials.
to relatively low signal-​to-​noise ratio. This property is
inherent to the derivative operation. Examples and applications. The applicability of the
A similar approach can be used to perform ana- Green’s function method for performing wave-​based
logue integration42: consider the configuration shown analogue computing has been experimentally verified

NaTure RevIewS | MAteRiAls volume 6 | March 2021 | 211

Reviews

in a series of proposals. In a first example, somewhat includes a thin layer of metallic film and a prism cou-
analogous to the implementation in Fig. 1d, an analogue pler. At the incident angle at which the momentum of
spatial differentiator based on the resonance behaviour the excitation field matches the one of the SPP mode
of surface plasmon polaritons (SPPs) was experimen- (phase-​matching condition), the reflection spectrum
tally demonstrated43. The differentiator, shown in Fig. 2a, drops to zero (Fig. 2b). Around this dip, the reflection

a SPP b c
1
Z

f(y)
|H(ky)|
θ0 Ideal
Fit
Measured

g(y)
y 0
–0.04 0 0.04
f(y) ky/k0
y g(y) = df/dy

d e f
1 1.5π
Metasurface
Ideal

f(y)
Arg{T}
|H(ky)|

H
g(y)
E

g(y)
f(y) θ

0 –1.5π
–0.63 0 0.63
ky/k0

g h i
1
Metasurface
Ideal
f(y)
|H(ky)|

g(y)

400 nm

0
0 0.3
ky/k0
Si

400 nm
Al2O3

Fig. 2 | Analogue computing systems based on the Green’s function (solid line, left vertical axis) and phase (dashed line, right vertical axis) of the
method. a | A spatial analogue differentiator based on a thin metallic film Green’s function of the structure in panel d, agreeing well with the one of
supporting surface plasmon polaritons (SPPs). b | Green’s function of the an ideal differentiator. f | Spatial edge detection based on the metasurface
structure (the reflection spectrum) near ky = 0, showing a resonance dip array shown in panel d. g | Scanning electron microscope image (top and
(zero). Near this zero, the Green’s function can be approximated with a side view) of an optical computing metasurface engineered such that it
linearly varying function (the straight lines), corresponding to the transfer realizes the transfer function for second-​order differentiation. h | Transfer
function of the ideal differentiator. c | Spatial edge detection based on the function of the metasurface. i | Edge detection based on the metasurface
plasmonic spatial differentiator shown in panel a. d | Spatial differentiation shown in panel g. Panels a–c adapted from ref.43, CC BY 4.0. Panels d–f
based on a metasurface array composed of split-​ring resonators. The adapted with permission from ref.46, APS. Panels g–i with permission from
refractive indices of the dielectrics inside the SPP gaps are modulated with ref. 47, https://pub.acs.org/doi/10.1021/acs.nanolett.9b02477, further
a slowly varying function, offering a degree of freedom to enhance the permission related to the material excerpted should be directed to
operation spectral bandwidth and the overall efficiency. e | Magnitude the ACS.

212 | March 2021 | volume 6 www.nature.com/natrevmats

 Reviews

spectrum follows the transfer function of a first-​order spatial modulation applied to the metasurface profile
differentiator (the straight lines). The fact that this dif- provides an additional degree of freedom to enhance
ferentiator is only composed of a single metallic layer the bandwidth of operation. More specifically, by pre-
makes the fabrication of the system simple. In addition, cisely controlling the parameters of the modulation,
it represents a significant miniaturization compared with the corresponding higher-​order (above the third order)
signal processors based on conventional bulky Fourier Taylor coefficients can be cancelled, allowing one to
elements. At the same time, its operation is limited to a enhance the spatial bandwidth of differentiation, that
narrow spatial spectrum around the dip in reflection and is, the overall resolution of the device, and tune it to the
does not avoid the aforementioned limitations in terms desired level corresponding to the numerical aperture of
of overall resolution of the images that can be processed the optical system of interest. The corresponding opti-
and of overall efficiency. mized Green’s function, obtained using numerical sim-
Several important applications have been proposed ulations, is shown in Fig. 2e, and is in good agreement
for analogue spatial differentiators. In one dimension, with the transfer function of the ideal case. The slight
they can be used to determine the sharp variations of deviation from the ideal response might be suppressed
functions. Likewise, in two dimensions, spatial dif- by introducing more sophisticated modulation patterns
ferentiation provides the possibility of detecting the over the metasurface profile. Figure 2f demonstrates
sharp variations of incident images (2D signals), that is, the performance of this second-​order differentiator for
their edges44,45. Figure 2c demonstrates the possibil- edge detection. Compared with the plasmonic spatial
ity of performing edge detection using the plasmonic differentiator discussed earlier, this design provides a
spatial differentiator shown in Fig. 2a. To this end, the higher resolution for edge detection, thanks to its larger
incident field was modulated with a spatial light modu- operational beamwidth. It is worth mentioning that it
lator and projected onto the metal film of the plasmonic is also possible to achieve wideband first-​order differ-
differentiator. The bottom panel of Fig. 2c depicts the entiation by slightly modifying the structure of Fig. 2d.
corresponding reflected image. Indeed, the edges of To this end, one should break both the horizontal and
the incident image (Fig. 2c, top) have been resolved. Note vertical mirror symmetries of the system, for example,
that the vertical edges of the image are resolved better, by adding a horizontally misplaced array structure on
because the differentiation is performed only along the one side of the SRR array. Alternatively, the reflection
horizontal direction (y axis). Yet, as long as the edge is symmetries of the structure can be broken by tilting the
not purely vertical, its signature can be traced in the direction of the incident field, leading to a first-​order
reflected image. This limitation was later overcome with zero around which the transmission follows a linearly
a design based an all-​dielectric metasurfaces22, which varying function akin to the transfer function of the
had the additional advantage of being less affected by first-​order differentiator.
absorption losses. The metasurface discussed above was designed to
As mentioned above, as a trade-​off for its simplicity, operate in the microwave range, but a similar response
the plasmonic spatial differentiator has a few drawbacks. can be obtained at optical frequencies. An optical spatial
In particular, it works only for incident waves possessing differentiator based on a resonance-​based metasurface
a narrow spatial bandwidth (this is caused by the pres- array was realized47, composed of low-​loss silicon die-
ence of higher-​order terms in the Taylor series expan- lectric resonators placed on top of an Al2O3 substrate
sion of R(ky)). For this reason, the differentiator is not (Fig. 2g). By engineering the resonators composing the
capable of resolving edges that are very close to each metasurface, the Green’s function of the structure was
other. In a recent proposal46, a more sophisticated spatial tailored such that it approached the transfer function of
analogue differentiator going beyond these limitations the second-​order derivative operator over a wide oper-
was presented. The structure of the differentiator, shown ational bandwidth (Fig. 2h). The excellent performance
in Fig. 2d, exploits a spatially modulated resonant metas- of this optical differentiator for characterizing the edges of
urface with a strong non-​local response, a property that incident images is demonstrated by the images in Fig. 2i.
is often considered undesirable for other metasurface It should be noted that, although both the structures in
applications. The metasurface array is composed of res- Fig. 2d and in Fig. 2f work only for a specific polarization
onant particles (split-​ring resonators, SRRs, resonating of the incident wave, a recent work has demonstrated the
at a frequency f0 = 1.26 Ghz), on which a TM-​polarized possibility of performing the same kinds of operation for
wave is normally incident. A slow periodic modulation both transverse-​electric and TM polarizations48.
is applied to the relative permittivities of the dielec-
trics inside the SRR gaps, leading to the appearance of Metasurface approach
a leaky-​wave resonance with controllable non-​locality. Principle. Although the Green’s function technique is
At the leaky-​wave-​resonance frequency, the reflection a straightforward approach to realize specific-​purpose
coefficient of the structure drops to zero. Furthermore, computational functionalities, it does not generally pro-
this frequency is a function of the incident angle, owing vide a platform that can be easily adapted to an arbi-
to the non-​local properties of the leaky wave. Then, if trarily complex operator without using more complex
the operation frequency is equal to the leaky-​wave fre- geometries and optimization techniques. An alternative
quency at normal incidence, the reflection spectrum strategy, known as the metasurface approach49–68, has
changes as a function of the incident angle according to enabled the realization of a wider range of operators.
a parabolic function law, matching the transfer function The core idea of this approach, schematically sketched
of the second-​order differentiator22,35. Interestingly, the in Fig.  3a, is essentially to map the conventional 4F

NaTure RevIewS | MAteRiAls volume 6 | March 2021 | 213

Reviews

correlator approach, commonly used in Fourier optics (characterized by the transfer function H(kx)) to the
and described in the previous section, onto a more input field f(x). From a systems perspective, the rela-
compact metamaterial platform. Consider a linear, shift-​ tion between the input field f(x) and the corresponding
invariant system, aimed at applying a specific operator output g(x) is expressed as

a
f(x) Lens; Fourier transformer F(kx) Metasurface; H(kx) G(kx) Lens; inverse Fourier transformer g(x)

b c f(x) g(x)
GRIN(+) GRIN(–) First-order differentiator
1

w/2

w/2
f(x) g(x)
5λ0
1 ε 1 ε

–w/2

–w/2
0
Lens Lens
Metasurface H(kx) AZO –5λ0
–1 Lens Metasurface Lens
Si

d
d e f f(y)

200 1
Au –20° 160°
Ly Lx SiO2
0.8
Au 150
Ly (nm)

0.6

|r|
g(y)
100 0.4

0.2
50

50 100 150 200

Lx (nm)

g h i
Phase diagram
550 360
g(x) 0°
Ly (nm)

f(x) 180°
325
f(x)

180° 180

0°
Lens 0
Reflection amplitude
550 1
0°
Ly (nm)

325 180° 0.5

g(x)

SiO2 180°

Silicon nanoresonator 0°
0
100 325 550 x
Lx (nm)

Fig. 3 | Metasurface approach for wave-based analogue computing. a | Block diagram of a computing system based on
the metasurface approach, consisting of three sub-​blocks: two Fourier transformers (graded index (GRIN) and lenses) and
a metasurface realizing the transfer function of the operator of choice. b | A metasurface computing system designed to
perform first-​order differentiation. c | First-​order spatial differentiation based on the metasurface computing system in
panel b. d | Analogue computing based on a reflective metasurface array, consisting of silicon nanobricks arranged on a
silica substrate and a thick metallic layer. e | The amplitude and phase of the field reflected by the metasurface array can
be manipulated independently by varying the length (Lx) and width (Ly) of the nanobricks. The plot shows the amplitude
of the reflection coefficient, r. The green lines correspond to the phase of the reflection. f | Experimental demonstration of
the performance of the metasurface computing system in panel d. A Gaussian derivative incident field as the input field
(top) results in a reflected field that is its derivative (bottom). g | A dielectric metasurface computing system, consisting of
silicon nanobricks deposited on top of a silica substrate. h | Variation of the amplitude and phase of the reflection coefficient
as a function of the length, Lx, and width, Ly, of the silicon nanobricks. i | Demonstration of the operation of the metasurface
computing system as a second-​order spatial differentiator. Panels b and c adapted with permission from ref.4, AAAS.
Panels d–f adapted with permission from ref.71, ACS. Panels g–i adapted with permission from ref.72, © The Optical Society.

214 | March 2021 | volume 6 www.nature.com/natrevmats

 Reviews

  metasurface array and R0 is a constant. The parameters

  of the metasurface can be tailored to achieve the des­
g (x ) = IFT  H (k x ) ⋅ FT [ f (x )]  (3)
⏟    
Block3  Block2 Block1  ired reflection profile. The proper operation of the
corresponding computing system was experimentally
Equation 3 governs the working of analogue computers demonstrated by exciting the structure with a Gaussian
based on the metasurface approach, which essentially derivative incident signal; the corresponding reflected
consist of three distinct sub-​blocks: a spatial Fourier field was the derivative of the incident field, confirming
transformer, FT, that takes the Fourier transform of the proper functionality of the system (Fig. 3f).
the input field f(x) (as explained previously, the Fourier Although this plasmonic metasurface can perform
transformation can be performed using an optical lens); various mathematical operations, it suffers from some
a properly designed metasurface with position-​ limitations: it has high absorption and low conversion
dependent transmission (or reflection) coefficient, cor- efficiency, which stems from the use of lossy plasmonic
responding to the transfer function of the operator of materials. In addition, it is not compatible with com-
choice; and an inverse Fourier transformer, IFT, that plementary metal–oxide–semiconductor (CMOS)
takes the inverse Fourier transform of G(kx) = F(kx)H(kx), technology, hindering its integration within silicon
yielding the desired output field g(x). photonic devices. These disadvantages were overcome
by developing an all-​dielectric metasurface computing
Examples and applications. An example of a metasur- system72. The metasurface was built from silicon nano­
face computing system, designed to calculate the first-​ resonators placed on top of a silica spacer and a thick
order derivative (∂/∂x) of input signals, is shown in layer of silver (Fig. 3g). The amplitude and phase of the
Fig. 3b. The system includes two graded-​index dielectric reflection coefficient of the metasurface as a function of
lenses69 that have a parabolic variation of permittivity the width (Ly) and length (Lx) of the silicon nanobricks
ε(x) = εc(1 − (πx/2L)2), where L is the focal length of the are shown in Fig. 3h. The associated reflection coefficient
lens. In the paraxial approximation, such an inhomo- spans the full phase range of 0 to 2π, whereas the ampli-
geneous material acts as a Fourier transformer69. The tude of the reflection can vary from 0 to 1 by varying Ly
metasurface block is composed of a layered structure and Lx. Coupled to graded index lenses, these features
of two alternating materials, aluminium-​doped zinc enable the realization of arbitrary mathematical func-
oxide (AZO) and silicon70, with different dissipation tionalities. As an example, a second-​order differentia-
losses. The geometrical parameters of the metasurface tor was designed by properly structuring the nanobricks
are tailored such that it provides a position-​independent of the metasurface. When a sinc-​shaped electric field,
transmission coefficient akin to the transfer function f(x) = sinc(x/6.8 × 10−6), was used as the input field, the
of the targeted operator, H(kx) = ikx. The simulated per- reflected field was, indeed, its second-​order derivative
formance of the system is shown in Fig. 3c, and con- (Fig. 3i).
firms that the transmitted field is, indeed, the derivative
of the incident field distribution. Compared with the Acoustic computational metamaterials
analogue differentiators based on 4F correlators, this In addition to their development in optics, computa-
system provides several advantages. In particular, tional metamaterials have been explored in another area
by properly engineering the metasurface sub-​block, of classical physics, namely, acoustics. Although acoustic
the system can be adapted to perform more complex computational metamaterials are not as fast as their opti-
mathematical operations, such as local phase control, cal counterparts, as they operate at the speed of sound,
that are not achievable with standard spatial analogue they could potentially be used to accelerate imaging and
filters. In addition, it provides higher-​resolution recon- prospection methods by allowing more processing tasks
struction, because the metasurface sub-​block can be to be performed analogically and in parallel. In addition,
deeply subwavelength. Alignment issues may also tremendous advances in surface acoustic-​wave technol-
be reduced if the entire system can be manufactured in ogy offer great potentialities for the miniaturization
one block. of acoustic analogue signal processors. For example,
Several other computing systems based on the an acoustic computing system based on the metasur-
metasurface approach were proposed and experimen- face approach73 employed a metastructure with a unit
tally demonstrated. For example, an analogue system cell composed of three tapered, labyrinthine, spiral com-
based on a plasmonic meta-​reflect array capable of ponents with varying radians, S1, S2 and S3 (Fig. 4a). By
performing a large variety of processing operations was tuning S1, S2 and S3, full control over both the amplitude
demonstrated71. The unit cell of the metasurface, shown (Fig. 4b) and the phase (Fig. 4c) of the transmitted field
in Fig. 3d, is composed of silicon nanobricks arranged on was achieved, providing the possibility of achieving a
top of a silica layer placed on an optically thick metallic broad range of analogue computational tasks. As a spe-
film. By varying the size of the nanobricks (Lx and Ly), cific example, a second-​order linear differential equa-
the amplitude and phase of the reflected field can tion solver was designed using such a metastructure. The
be independently controlled (Fig. 3e). This enables the transfer function of the targeted differential equation
realization of arbitrary transfer functions. Suppose, (Fig. 4d) was realized by precisely tuning the spiral radi-
for example, that we want to realize the derivative ans of the unit cells. The functionality of the system was
∂/∂x. The associated transfer function (H(kx) = ikx) demonstrated in simulations by exciting the metasurface
implies a position-​dependent reflection coefficient of with a Gaussian incident field (Fig. 4e), which resulted in
the form |r| = R0x/L, in which −L < x < L is the length of the a transmitted pressure field (Fig. 4f) in perfect agreement

NaTure RevIewS | MAteRiAls volume 6 | March 2021 | 215

Reviews

with the analytically predicted solution of the targeted propagation, effectively reducing the speed at which the
ordinary differential equation (ODE). acoustic wave travels75. As a result, the metamaterial acts
This computing system is based on the metasurface like a high-​index acoustic medium, confining sound and
approach. Green’s function-​based acoustic computa- guiding it via total internal reflection (this is akin to the
tion has also been implemented with metamaterials to principle of dielectric slab waveguide in optics). At a
perform specific-​purpose computational tasks using specific incident angle, the characteristic impedance of
acoustic signals. An acoustic spatial differentiator was, the metamaterial matches that of the incident medium,
indeed, demonstrated based on the Green’s function air. This leads to a dip (zero) in the reflection spectrum,
method74, and its structure is shown in Fig. 4g. It was near which the Green’s function of the structure can
built from a metamaterial composed of a square array of be approximated with a linearly varying function. At
cross-​shaped pipes. The holey structure of the metama- this condition, in real space, the structure serves as a
terial provides a reduced compressible volume for sound first-​order differentiator. The relevance of such kind

a b 1
c π

2.5 2.5

Arg (T0) (rad)

S3 (π)

S3 (π)
S2

|T0|
S1 S3 1.5 1.5

0.5 0.5
2.5 2.5 2.5
2.5
1.5 1.5 0 1.5 –π
S2 (π) 1.5
0.5 0.5 S1 (π) S2 (π) 0.5 0.5
S1 (π)

d e f
1 1 1 Simulation
1 Exact
Pressure (a.u.)
Arg (T0)

|T0|
|T0|

0.5 –1 0
–1 –0.3 0 0.3
Lens Metasurface Lens
y (m)
–0.3 0 0.3
y (m)

g h

Mask Loudspeaker
Microphone

f(x,y)
θ
θ
g(x,y)
Vertical
Horizontal

Air

Fig. 4 | Acoustic computational metamaterials. a | An acoustic metasur- incident signal shown in this panel. f | The corresponding transmitted field,
face computing system consisting of three tapered, labyrinthine following the solution of the targeted ordinary differential equation. g | An
components with varying spiral radians, S1, S2 and S3. The complex path-​ acoustic computing system based on the Green’s function approach. The
coiling of the labyrinthine structures creates strong, multiple-​scattering structure consists of a half-​wavelength high-​index metamaterial, whose
effects, leading to an extreme range of amplitudes and phases of the trans- reflection coefficient drops to zero at some specific incident angle θ. Near
mitted field. b | Amplitude of the transmission coefficient, T0, as a function this zero, the reflection spectrum can be well estimated with the transfer
of S1, S2 and S3. c | Phase of transmittance as a function of S1, S2 and S3. function of the ideal differentiator. The incident field is modulated by a mask
d | Amplitude (purple) and phase (green) of the transmission spectrum plane with a properly designed transparency pattern corresponding to the
associated with the transfer function of a second-​order ordinary differential EPFL logo. h | Spatial vertical and horizontal edge detection based on the high-​
equation solver. e | The performance of the designed metasurface index acoustic computing metamaterial in panel g. Panels a–f adapted from
computing system is evaluated by simulating its response to the Gaussian ref.73, CC BY 4.0. Panels g and h adapted from ref.74, CC BY 3.0.

216 | March 2021 | volume 6 www.nature.com/natrevmats

 Reviews

of a differentiator for acoustic-​image edge detection is By employing partial fractional decomposition, H(ω)
illustrated in Fig. 4h. can be written as a summation of the form

Emerging directions
n Ki
H (ω) = ∑ j(ω − ω0 ) + ω0/2Q i
, (8)
Wave-​based equation solvers i =1
Solving linear differential equations. Equations are
ubiquitous in many areas of science, including mathe- where Qi = −ω0/2Pi, in which Pi are the complex roots of
matics, physics and engineering. Wave-​based analogue the associated nth-​order denominator polynomial, and
computers such as those discussed in the previous sec- Ki are complex constant coefficients. Equation 8 suggests
tions represent an ideal platform for ultra-​fast equation that the transfer function of an arbitrary nth order ODE
solving76,77. As a relevant example, we start by discussing can be realized by adding the output signals of n different
first-​order ODEs, with general form Lorentzian resonators with different quality factors Qi.
The summation can be performed in a fully analogue
αf ′(t ) + βf (t ) = g (t ) (4)
fashion using, for instance, standard directional couplers78.
A wave-​based analogue system returning the solution
of Eq. 4 for given coefficients α and β is schematically Solving integral equations. Metamaterials have also
shown in Fig.  5a . Consider an arbitrary resonator been exploited to solve integral equations that cannot
with a Lorentzian spectral line shape around the be directly solved using a Fourier transform. In particu-
resonance. The Green’s function of the resonator is lar, a metamaterial platform was proposed79 to solve the
given by general integral equation
H (ω) = 1/(α j(ω − ω0 ) + β), (5) b
in which α and β are arbitrary constants that depend g (u) = Iin(u) + ∫ K (u, v )g (v )dv, (9)
a
on the quality factor of the resonator. The Green’s
function of Eq. 5 (plotted in Fig. 5b) is equivalent to in which Iin(u) is the input signal, g(v) is the unknown
the transfer function of the first-​order ODE in Eq. 4, function and K(u,v) is the kernel associated with an
implying that, in the time domain, any Lorentzian res- arbitrary operator. Note that K(u,v) is assumed to have
onator acts as a first-​order ODE solver. In acoustics, general dependence on u and v, not just on the differ-
such kind of functionality can be achieved by, for exam- ence u − v, which would correspond to a convolution,
ple, putting a defect inside a Bragg phononic crystal preventing, in general, the equation to be solved by using
(Fig. 5c). In another example of an all-​optical differen- Fourier optics. The schematic of a metamaterial-​based
tial equation solver experimentally realized based on a system solving Eq. 8 with the arbitrarily chosen kernel
resonance phenomenon74, the system was made of
a silicon microring resonator coupled to two straight 

K (u, v ) = 0.06(4 − 4i )J0(uv )
waveguides (Fig. 5d). Around the resonance frequency of 


the microring resonator, the transmission spectrum (10)
 i 2π 
of this system exhibits a Lorentzian peak (Fig. 5e), cor- + 3 exp  − (1 − 2i )
responding to a first-​order ODE with specific constant  5(u + v ) 


coefficients. Remarkably, the spectral characteristics
of the structure can be tuned via a gate voltage applied is shown in Fig. 5g. In this system, the kernel K(u,v) is
to the microring resonator, enabling a wide control over implemented using a binary metamaterial composed of
the constant coefficients. To examine the performance air and polystyrene. The inhomogeneity of the metama-
of the differentiator, the microring resonator was excited terial structure, described by an inhomogeneous rela-
with a Gaussian pulse. The corresponding transmitted tive permittivity ε(x,y), allows one to realize the desired
signal was in excellent agreement with the solution of the kernel function K(u,v) using an optimization process.
targeted ODE (Fig. 5f). N feedback waveguides provide feedback to the system,
b
Note that, in general, it is possible to solve higher-​ adding the output of the metasurface ∫a K (u, v )g (v )dv to
order ODEs by constructing a network of first-​order the input function Iin(u) at N different points across the
ODE solvers with different transfer functions. As an [a,b] interval. In this discrete form, the system imple-
example, let us suppose that we want to engineer a ments the equivalent N × N matrix equation of Eq. 9. The
transfer function of order n (n may be arbitrarily high), prototype of the metamaterial structure, in which five
corresponding to the following differential equation metallic waveguides operating at the fundamental TE10
mode were used to sample the output across the desired
f n (t ) + An −1 f n −1 (t ) + … + A1 f ′(t ) + A0 f (t ) range, is shown in Fig. 5h. The output signal measured
(6)
=Bg (t ) . at the first sampling waveguide, shown in Fig. 5i, is in
excellent agreement with the theoretical solution of the
The associated transfer function is given by targeted integral equation.

B Topological and non-​reciprocal computing

H (ω) = . Wave-​based analogue computers provide the possibil-
n n −1
(j(ω − ω0 )) + A n −1(j(ω − ω0 )) + ... (7) ity of carrying out specialized computational tasks with
+ A1(j(ω − ω0 )) + A 0 ultra-​high speed, but they suffer from an important

NaTure RevIewS | MAteRiAls volume 6 | March 2021 | 217

Reviews

limitation, which may hinder their applicability in To tackle this issue, researchers have leveraged a par-
large-​scale applications. Unlike DSPs, in which disor- ticular class of metamaterials with non-​trivial topology,
der is not an issue thanks to the binarized or discretized known as topological metamaterial insulators. Such
nature of the input, wave-​based analogue computers insulating artificial materials exhibit frequency band
are typically fragile to noise and perturbations. In par- gaps in which no bulk wave propagation is possible.
ticular, the errors caused by geometrical imperfections However, the particular topology of their band struc-
can accumulate during series operation, degrading ture guarantees the presence of modes on their bound-
the performance of the analogue computer. aries. Because the existence of these boundary modes

a b 1/β c

f(t) g(t) Input

αdf/dt+βf = g
|H(ω)|

Output

0
–10α 10α
ω

d e f

Measured amplitude Input

2
Ideal amplitude Output
Ideal output
1

Ring

Intensity (a.u.)
|T|

1
Waveguide
Waveguide

0 0
–150 –100 –50 0 50 100 150 –50 0 50

Frequency (GHz) Time (ps)

g h i
N g(u) Iin(u) Experiment (real)
1.0 Experiment (imaginary)
Theoretical (real)
2 Theoretical (imaginary)
Simulation (real)
0.5
1 Simulation (imaginary)
g(u)

0.0
g(u) ∫g(u)K(u,v)dv

–2 2
u

Fig. 5 | Wave-based analogue equation solving. a | A general equation The measured output field matches the solution of the target ODE.
solver, aimed at solving a first-​order ordinary differential equation (ODE). g | A metamaterial platform solving integral equations of general form
b
b | Green’s function of a Lorentzian resonator, matching the transfer function g(u) = Iin(u) + ∫a K (u, v ) g(v )dv , where Iin(u) is the input signal, g(v) is the
of the first-​order differential equation. c | A first-​order ODE solver based on unknown function and K(u,v) is the kernel associated with an arbitrary
resonance tunnelling through a Bragg band gap. d | An optical first-​order operator. The equation solver consists of a metamaterial realizing the kernel
ODE solver based on the resonance of a microring resonator fabricated on of the integral equation and N feedback waveguides adding the output of
a silicon wafer. e | Amplitude (yellow) of the transfer function of the the metastructure to the input signal at specific sampling points.
microring resonator around its resonance frequency, which is equivalent to h | A photograph of the metamaterial structure realizing the proposal in
that of a first-​order ODE with specific constant coefficients (black). panel g. i | Demonstration of the proper performance of the integral equation
T, transmission coefficient, f frequency. f | Experimental demonstration of solver. Panel c adapted from ref.86, CC BY 4.0. Panels d–f adapted from ref.134,
first-​order ODE solving based on the microring resonator in panel d. CC BY 4.0. Panels g–i adapted with permission from ref.79, AAAS.

218 | March 2021 | volume 6 www.nature.com/natrevmats

 Reviews

is a direct consequence of the topology of the bulk ODE by subtracting the output signals of two topological
band structure of the system, it cannot be influenced first-​order systems with tailored dissipation losses.
by continuous deformation, such as that induced by Another way to achieve robust computing is to use
the local introduction of small defects or geometrical the one-​way modes propagating in non-​reciprocal
imperfections. Instead, the topology can only change optical structures. Such non-​reciprocal effects can be
through a global, drastic modification of the system obtained by breaking time-​reversal symmetry using, for
that would entirely destroy its insulating property. example, a magneto-​optic material under an external
As a result, the edge modes of topological structures magnetic field88,89. For instance, the one-​way character of
exhibit strong immunity to small and moderate levels magnetized SPPs was used to realize a spatial differentia-
of imperfections80–85, provided that the disorder does tor immune to backscattering90. The non-​reciprocal dif-
not break a symmetry on which the topological order ferentiator consisted of an opaque medium film (ε = −1)
depends. Motivated by these advances, a new class coated on a magnetic substrate (InSb, Fig. 6d). A static
of wave-​b ased analogue signal processors leverag- magnetic field was applied to the substrate, breaking
ing these robust topological boundary modes was time-​reversal symmetry and leading to one-​way SPP
recently introduced. Such topological analogue signal excitation. The reflection coefficient of the structure has
processors86, for example, implement the Green’s func- a dip near the resonance frequency of the SPP mode,
tion method using resonance tunnelling through the which can be employed to perform spatial differenti-
boundary mode of a 1D topological wave insulator, ation, as we discussed previously. The mode profile of
thereby, featuring strong robustness against a large the structure when it is excited with a Gaussian incident
range of defect types. field at the frequency (and the momentum) of the SPP
To provide a specific example, consider the config- mode is shown in Fig. 6e. As expected, the reflected field
uration shown in Fig. 6a (left), consisting of an acoustic has a Gaussian derivative profile. Remarkably, there is
pipe, inside which a sonic crystal built from solid cylin- no backscattering, even in the presence of geometrical
drical rods is arranged. The two halves of the phononic irregularities at the interface between the two layers of
crystal include two insulating lattices with different the structure.
topological properties, inspired by a specific topolog-
ical arrangement known as the Su–Schrieffer–Heeger Metamaterials for machine learning
(SSH) scheme, initially discovered in condensed-​matter As mentioned above, one important advantage of ana-
systems 87. The difference between the topological logue approaches to perform computations is the ability
invariants of the two insulating halves of the system to process information in parallel, without restrictions
leads to an interface mode located at its centre, which on the operating speed. This advantage becomes impor-
is symmetry-​protected, as long as all the rods have the tant when a large amount of data has to be processed,
same diameter. This edge mode has a Lorentzian spectral for instance, when performing large matrix multipli-
line shape, which, as explained in the previous section, cations or decoding the information hidden in a large
can be utilized for solving first-​order ODEs. Remarkably, amount of measurements. Such decoding tasks can be
the frequency of the edge mode is pinned by topology, highly non-​trivial, especially if no analytical description
even if some relatively large disorder is imparted to the or physical modelling of the relation between the infor-
positions of the rods. Indeed, after randomly shifting mation and the data is available (for example, imaging
the rods (Fig. 6a, right), the transfer function of the equa- or recognizing objects using waves through an unknown
tion solver, and, as a result, the corresponding output multiple-​scattering environment). One way to tackle
signal, are not affected. The boundary mode is pinned such problems is to use machine learning91, a modern
to a certain frequency, as long as the position disorder is processing technique in which a computer is given the
not strong enough to turn it into an insulator, and as long opportunity to learn on some training data, for which
as the symmetry protection (identical rod diameters) the encoded information is known, before being capa-
is present86. ble of processing unknown data. Machine learning often
This system has been used to perform analogue sig- requires some particular form of non-​linearity, making it
nal processing (Fig. 6b). The measured transfer function much more complicated than the linear analogue func-
of the structure, sketched in the middle panel of the fig- tionalities described so far, because non-​linear opera-
ure, exhibits a peak at the resonance frequency of the tions cannot be tackled by Fourier optics approaches.
topological edge mode (f0). To characterize the perfor- The training process, during which the system learns the
mance of the system, the waveguide was excited with intricate links between the encoded information and
an arbitrarily shaped signal (left panel), modulated at f0. the data, and reconfigures itself to be able to process
The corresponding measured output signal, shown new data, is often based on a neural network92, an inter-
on the right side of the panel, follows the solution of the connected, reconfigurable network of artificial neurons
targeted first-​order ODE (dashed line). The function­ connecting the data and the information to be retrieved.
ality of the equation solver is immune to position dis- The neurons and their connections, called edges, are
order, as shown in the lower part of Fig. 6b. Note that, generally assembled into a layered architecture, in which
as we explained in the previous section, by constructing the output of each neuron is a weighted, non-​linear func-
a network of first-​order ODE solvers using directional tion of its inputs. The reconfigurable part of the network
couplers, one can achieve transfer functions corre- is represented by weights that increase or decrease the
sponding to higher-​order ODEs. Figure 6c, for instance, signal at a connection, a process somehow inspired by
demonstrates the possibility of solving a second-​order biological neural networks that constitute animal brains.

NaTure RevIewS | MAteRiAls volume 6 | March 2021 | 219

Reviews

Input Input
Output Same output
Disorder

Transfer
function Preserved
transfer function

b Measured Measured
~ |H(f)| Simulation Simulation
g(t)
Target Target

Amplitude
Amplitude

1 ~ 1
Ordered f(t)
0 0
–1 –1
|H(f)|
0 0.1 0.2
Time (s)

Amplitude
~ 1
Disordered f(t)
0
–1
0.05 0.1 0.15 0.2
Time (s)
c
1 f(t)
0
–1
g(t) g(t)
0 0.07 0.15
1 Time (s) 1
0 0
–1 –1

0 0.07 0.15 0 0.07 0.15

Time (s) Loss Loss Time (s)

d e
Defect

Opaque coated Incident Reflected

B0 InSb y z
x
θ0
x Defect

f(x)
g(x) = df(x)/dx
x

Fig. 6 | Robust analogue signal processing. a | A topological analogue signal processor solving first-​order differential
equations (left). The equation solver is based on the edge mode of a 1D Su–Schrieffer–Heeger (SSH)-​like array of nylon
plastic rods arranged inside an acoustic waveguide. Right: some disorder is added to the system by randomly changing
the positions of the obstacles. The topological equation solver is robust to the disorder. b | Experimental demonstration
of topological analogue signal processing using the system in panel a. The transfer function of the system (middle panel)
exhibits a peak near the resonance frequency of the edge mode, following the transfer function of a first-​order ordinary
differential equation (ODE). The corresponding output signal is, therefore, identical to the solution of the ODE (dashed
line). The system preserves its functionality even in the presence of disorder (bottom signal path). c | Realization of a second-​
order ODE solver by subtracting the output signals of two different first-​order ODE solvers (like the ones in panels a
and b) from each other. The subtraction is performed using an acoustic rat-​race coupler. d | Spatial differentiation based
on non-​reciprocal, one-​way magnetized surface plasmon polaritons formed at the interface between an opaque medium
film and the InSb substrate. e | Demonstration of the proper functioning of the differentiator and its robustness to disorder.
Panels a–c adapted from ref.86, CC BY 4.0. Panels d and e adapted with permission from ref.90, APS.

Neural networks are traditionally implemented in the of the information in the analogue domain, which
digital domain and work on digital data. However, when can yield considerably higher speed of operation and
practical data originate from a measurement operation processing efficiency, and lower power consumption.
with a large amount of analogue sensors, it may be useful In this section, we describe two categories of efforts
to perform, as much as possible, the parallel processing in which metamaterials are combined with machine

220 | March 2021 | volume 6 www.nature.com/natrevmats

 Reviews

learning for this purpose. This field is very broad and to decode these data using machine learning to extract
active, hence, we specifically focus on works in which the desired information. This second step involves a
metamaterials have a computational role, rather than series of large, linear matrix multiplications, application
on research that leverages machine learning as a design of non-​linear functions and successive adjustments of
tool for metamaterials. Let us start with an elementary weights. The green labels in Fig. 7a highlight at which
information-​theoretic description of a sensing process, steps of this process metamaterials can be used to
represented in Fig. 7a. Part of the information that we perform a computational task.
are interested in, encoded in a field that propagates in a A first category of computing metamaterials is con-
potentially complex medium, is first sampled via some cerned with the encoding part. The information carried
physical mechanism, typically involving sensors. This by a wave field is either contained in the spatial degrees
measurement process may lead to a dataset in which of freedom of the wave or in its frequency degrees of
the information of interest is present in part or in full. freedom. Metamaterials, composed of resonant scat-
Sometimes, more information than needed is contained terers with large scattering cross section, can possess
in the data. Nevertheless, the second step is generally both spatial dispersion (linked to multiple scattering)

a c
Encoding Disorder medium
Propagation Sensing

Information Measurement

Imaging
Machine
Extraction learning
Processing
Classification
b Input digits
Processing operation
Without lens

Reconstructed digits
Tunable
SRR
Neural network
Input digits
With lens

e
Reconstructed digits

Directional coupler
d Optical interference unit

Absorber

Meta-
Antenna

reflect
array Arbitrary
object

60 μm SU(4) core DMMC VNA

Fig. 7 | Machine learning with metamaterials. a | A typical sensing process, by the neural network110. c | Concept of learned sensing, where the first
where information carried by a wave field is encoded in measured data. layer of the neural network is a physically reconfigurable layer that is
These data are processed by a machine-​learning method to extract the optimized during the learning process to focus only on measuring the
desired information. b | Example of a metamaterial used to help the learning relevant information104. d | Microscope image of a silicon photonic neural
process by acting at the encoding stage. A subwavelength acoustic source network used to implement linear matrix multiplication for deep learning.
is generated by a loudspeaker array and shaped as a handwritten digit. The e | A chaotic microwave cavity with reconfigurable boundary conditions
acoustic far field is sampled at four different points using microphones. implemented using programmable metasurfaces can be used to perform
After learning, a neural network is not able to reconstruct or classify the very efficient and fast large-​matrix multiplication. DMMC, diagonal matrix
digits, because the information about subwavelength features of the source multiplication core; SRR, split-​ring resonator; VNA, vector network analyser.
(λ/30) cannot reach the far field. By placing a locally resonant metamaterial in Panel b reprinted from ref. 93, CC BY 4.0. Panel c image courtesy of
the near field (not shown in the picture), this information can be encoded Philipp del Hougne, Université Côte dʼAzur, France. Panel d adapted from
in the far field, enabling the reconstruction and classification of the digits ref.106, Springer Nature Limited. Panel e reprinted from ref.107, CC BY 4.0.

NaTure RevIewS | MAteRiAls volume 6 | March 2021 | 221

Reviews

and frequency dispersion (linked to inertial response). analogically with waves. Indeed, a metamaterial or an
Therefore, they can process both the spatial and the artificial structure can be involved at a later step, namely,
temporal information carried by the wave field, for in the heavy linear matrix multiplications involved in
instance, to make sure that a maximum amount of the the processing, in order to gain efficiency in terms of
desired information is present in the measured data. speed or power consumption. This is motivated by the
A practical example is shown in Fig. 7b. An array of ability of passive-​wave-​based systems to perform many
loudspeakers (placed in an anechoic chamber) creates operations in parallel, as an alternative to digital systems
an acoustic source with spatial features much smaller subject to the fundamental thermal limits of Moore’s law.
than the acoustic wavelength, λ, at the frequency of oper- For instance, a neural network based on silicon nano-
ation (the details are smaller than λ/30), drawing some photonic circuits has been proposed and applied to per-
handwritten digits taken out of the Modified National form deep learning106. In this system, a reconfigurable
Institute of Standards and Technology (MNIST) data- mesh of waveguides (Fig. 7d) was used to implement lin-
base. Four microphones are used to measure the acous- ear matrix multiplication, whereas the non-​linearity was
tic pressure at four points in the far field. By training a added digitally. Another interesting approach to achieve
neural network on these data, it is not possible to image reconfigurable matrix multiplication at microwave fre-
the digits, because the information about subwavelength quencies is the use of chaotic cavities, such as a standard
features is evanescent and does not reach the far field. indoor environment at a Wi-​Fi frequency. By making the
However, by using a metamaterial built from an ensem- cavity reconfigurable by placing programmable metas-
ble of Helmholtz resonators placed in the near field, the urfaces on the room walls107 (Fig. 7e), a wide variety of
information contained in the evanescent fields can be very large matrix multiplications can be performed.
encoded into information carried to the measurement Rather than engineering a material with intricate design,
points by propagating waves. In the presence of such a this approach takes a random material and tweaks it with
locally resonant metamaterial, the neural network can a programmable metamaterial to implement the desired
image and classify the digits93. functionality, thereby, removing the prohibitive sensi-
A second category of machine-​learning systems in tivity to fabrication inaccuracies. At optical frequencies,
which metamaterials or metasurfaces have a computing spatial light modulators may be used to build reconfig-
role is based on an approach called learned sensing94–96. urable metasurfaces and achieve similar multiplication
In such systems, one stops looking at the encoding tasks at higher frequencies and lower scales108,109.
and processing tasks separately, which typically leads
to the acquisition of information that is irrelevant for Conclusion and perspectives
the target task. Instead, one optimizes the entire sens- In this Review, we have discussed the recent progress in
ing cycle at once by making sure that the system only the field of computational metamaterials and outlined
acquires and processes the part of the information that several examples of its practical applications in modern
is needed, sensing and learning at the same time. This engineering, including equation solving, image pro-
idea emerged in the context of computational imaging cessing and machine learning. We close by identifying a
using metamaterials, in which, originally, the meta- longer-​term outlook on this emerging field and discuss
materials did not have any computing role related to the key challenges that should be addressed in future
machine learning but were used at the encoding level work.
to implement compressive-​s ensing strategies 97–100, An interesting idea, proposed in a few recent
exploi­ting either frequency selection or reconfigurabil- papers110–112, is multifunctional wave-​based analogue
ity. In compressive-​sensing strategies, a scene is illu- computing, in which several computational tasks are
minated with random or orthogonal patterns, which performed simultaneously on different input channels.
multiplexes all information. However, a lot of this Such kind of operation provides the unique possibility of
information may be irrelevant. A first step along the parallel processing of information, substantially enhanc-
route from compressive to learned sensing was taken ing the speed of computation. For optical signals, multi-
by incorporating metasurfaces into sensing processes functional analogue computing can be achieved by, for
leveraging artificial-​intelligence methods101–103. The example, a metasurface composed of anisotropic com-
concept of learned sensing then naturally emerged, ponents. The anisotropic behaviour of such a structure
inspired by developments in the optical-​imaging com- allows one to manipulate the reflection and transmission
munities95. The idea, represented in Fig. 7c, is to inter- of the transverse-​electric and TM polarization degrees of
pret a reconfigurable acquisition layer (for example, freedom independently. This enables performing mul-
a few reconfigurable SRRs coupled to an ensemble of tiple signal-​processing tasks at once. Extensions of par-
non-​reconfigurable resonant dipoles, represented by allel computing to an arbitrary number of channels may
the yellow dots in the figure) as an additional trainable leverage media with many spatial or temporal degrees
layer of the machine-​learning system (the rest of it may of freedom, such as disordered multiple-​s cattering
remain digital). Therefore, the jointly learned measure- systems113. Not only does the parallel operation of such
ment and processing settings will select only the rele- kinds of computing structures enable enhancing the
vant information when measuring novel data. This idea, speed of processing but it also provides the possibility
which effectively makes the computation much more of saving a lot of power, avoiding the use of large-​scale
efficient104, was recently experimentally validated105. electronic systems and the associated challenging
Finally, we stress that even some of the remaining thermal-​dissipation problems, which are currently
digital part of the sensing process can be implemented limiting the continuation of Moore’s law.

222 | March 2021 | volume 6 www.nature.com/natrevmats

 Reviews

Most of the computational metamaterials discussed Most computational metamaterials, including the
in this Review are related to linear functionalities. With ones reviewed here, are resonance based, imposing cer-
the development of learning-​based approaches, investi- tain restrictions on their performance, notably in terms
gating the possibility of performing non-​linear analogue of the operational bandwidth. Broadening the band-
processing operations with computational metama- width of computational metamaterials is important for
terials is a clear opportunity for future research. We some applications. Some steps along this direction have
highlighted the relevance of wave-​based analogue com- already been taken with wave-​based computing systems
puters for machine learning (which inherently requires designed based on wavelength-​independent phenom-
non-​linear processing), but research on reconfigurable ena such as the Brewster effect40 or relatively broadband
or non-​linear metamaterials as a way to perform part or effects like the spin Hall effect45,117,118. Yet, active wave
all of a sensing process is still in its infancy. In addi- systems may represent an unexplored way to achieve
tion, non-​linear computational metamaterials can be wideband analogue computation: because they are not
used for other purposes, such as complex non-​linear subject to restrictive sum-​rule bandwidth constraints as
equation solving, or to implement analogically many passive systems, they can push the limits on bandwidth
of the non-​linear, filtering, image-​processing methods up to what is allowed by causality and stability. The field
employed in digital technologies. Non-​linear interac- of active metamaterials119–123 may be mature enough to
tion in such kinds of computational metamaterials, offer a unique solution to maintain absorption losses
combined with reconfigurability, also represents an to acceptable levels, enhance the signal-​to-​noise ratio
opportunity for the realization of programmable ana- and the bandwidth, or provide reconfigurability123.
logue computing systems, behaving as the analogue, Machine learning systems based on computing
wave-​based counterpart of electronic field-​programmable metamaterials have only been proposed in hybrid
gate arrays. analogue-​to-​digital sensing systems, but research on
Investigating the effect of undesired disorder on fully analogue machine-​learning systems is another
computational metamaterials is another important sub- interesting route, placed at the boundary between wave
ject for future research. As we discussed in this Review, engineering and artificial intelligence124–130. For this,
topological computational structures are robust against a source of non-​linearity may be introduced in the phys-
disorder such as impurities and defects. Yet, their ical layers to enable learning, exploiting Kerr dielectrics
much-​sought robustness is ultimately restricted, when or optomechanical resonators in photonics, or con-
the imperfections are large, by Anderson localization114, trolled geometrical or electromechanical non-​linearities
which leads to a progressive filling of the band gap of in acoustics131. Analogue Ising machines132,133 may also
the structure with localized bulk states in the regime be developed to solve specific, non-​d eterministic,
of strong randomness. Although this effect seems to be polynomial-​time hard (NP-​hard) problems. On the far
harmful at first glance, the fact that introducing disor- horizon, one may dream of combining wave-​engineering
der to a system can induce a topological phase transi- techniques with machine learning to enable the reali-
tion is very encouraging: it suggests that the opposite zation of a new generation of ‘auto-​computers’ func-
transition, from a trivial structure to a topological dis- tioning without any specific manual operation and
ordered one, is also possible115. Such a transition sug- programming. Such intelligent computational systems
gests the fascinating possibility of performing specific, may enable the solution of a wide variety of physical,
well-​defined computational tasks in the regime of dom- mathematical and engineering problems that are too
inant randomness, as suggested in a recent article116. complex in reasoning or cannot be described using sim-
These findings suggest that combining topologically ple mathematical language. This suggests a very bright
robust computing metamaterials and disorder-​based future for computational metamaterials.
computations may be a promising direction for future
research. Published online 19 October 2020

1. Lim, J. S. et al. Digital Signal Processing passive phased array. Sci. Rep. 6, 35437 programmable metamaterials. Light Sci. Appl. 3,
(Massachusetts Institute of Technology, 1987). (2016). e218 (2014).
2. Solli, D. R. & Jalali, B. Analog optical computing. 11. Zhou, J. et al. Optical edge detection based on 19. Goodman, J. W. Introduction to Fourier Optics
Nat. Photon. 9, 704–706 (2015). high-​efficiency dielectric metasurface. Proc. Natl Acad. (Roberts, 2005).
3. Small, J. S. General-​purpose electronic analog Sci. USA 116, 11137–11140 (2019). 20. Athale, R. & Psaltis, D. Optical computing: past and
computing: 1945-1965. IEEE Ann. Hist. Comput. 12. Molerón, M. & Daraio, C. Acoustic metamaterial for future. Opt. Photonics News 27, 32–39 (2016).
15, 8–18 (1993). subwavelength edge detection. Nat. Commun. 6, 21. Kou, S. S. et al. On-​chip photonic Fourier transform
4. Silva, A. et al. Performing mathematical operations 8037 (2015). with surface plasmon polaritons. Light Sci. Appl. 5,
with metamaterials. Science 343, 160–163 (2014). 13. Memoli, G. et al. Metamaterial bricks and quantization e16034 (2016).
5. Engheta, N. & Ziolkowski, R. W. (eds) Metamaterials: of meta-​surfaces. Nat. Commun. 8, 14608 (2017). 22. Zhou, Y., Zheng, H., Kravchenko, I. I. & Valentine, J.
Physics and Engineering Explorations (Wiley, 2006). 14. Khorasaninejad, M. et al. Metalenses at visible Flat optics for image differentiation. Nat. Photonics
6. Pendry, J. B. Negative refraction makes a perfect lens. wavelengths: Diffraction-​limited focusing and 14, 316–323 (2020).
Phys. Rev. Lett. 85, 3966 (2000). subwavelength resolution imaging. Science 352, 23. Bykov, D. A., Doskolovich, L. L., Bezus, E. A. &
7. Gansel, J. K. et al. Gold helix photonic metamaterial 1190–1194 (2016). Soifer, V. A. Optical computation of the Laplace
as broadband circular polarizer. Science 325, 15. High, A. A. et al. Visible-​frequency hyperbolic operator using phase-​shifted Bragg grating.
1513–1515 (2009). metasurface. Nature 522, 192–196 (2015). Opt. Express 22, 25084–25092 (2014).
8. Zhang, W., Cheng, K., Wu, C., Li, H. Y. & Zhang, X. 16. Tran, M. C. et al. Broadband microwave coding 24. Liu, X. & Shu, X. Design of an all-​optical fractional-​
Implementing quantum search algorithm with metamaterial absorbers. Sci. Rep. 10, 1810 (2020). order differentiator with terahertz bandwidth based
metamaterials. Adv. Mater. 30, 1703986 (2018). 17. Fan, W., Yan, B., Wang, Z. & Wu, L. Three-​dimensional on a fiber Bragg grating in transmission. Appl. Opt.
9. Li, L. & Cui, T. J. Information metamaterials–from all-​dielectric metamaterial solid immersion lens for 56, 6714–6719 (2017).
effective media to real-​time information processing subwavelength imaging at visible frequencies. Sci. Adv. 25. Wesemann, L. et al. Selective near-​perfect
systems. Nanophotonics 8, 703–724 (2019). 2, e1600901 (2016). absorbing mirror as a spatial frequency filter for
10. Xie, Y. et al. Acoustic holographic rendering 18. Cui, T. J., Qi, M. Q., Wan, X., Zhao, J. & Cheng, Q. optical image processing. APL Photonics 4, 100801
with two-​dimensional metamaterial-​based Coding metamaterials, digital metamaterials and (2019).

NaTure RevIewS | MAteRiAls volume 6 | March 2021 | 223

Reviews

26. Karimi, P., Khavasi, A. & Khaleghi, S. S. M. 53. Minovich, A. E. et al. Functional and nonlinear optical 82. Bandres, M. A. et al. Topological insulator laser:
Fundamental limit for gain and resolution in analog metasurfaces. Laser Photonics Rev. 9, 195–213 experiments. Science 359, eaar4005 (2018).
optical edge detection. Opt. Express 28, 898–911 (2015). 83. Xia, J. P. et al. Programmable coding acoustic
(2020). 54. Gao, L. H. et al. Broadband diffusion of terahertz topological insulator. Adv. Mater. 30, 1805002
27. Zangeneh-​Nejad, F., Khavasi, A. & Rejaei, B. waves by multi-​bit coding metasurfaces. Light Sci. (2018).
Analog optical computing by half-​wavelength slabs. Appl. 4, e324 (2015). 84. Huber, S. D. Topological mechanics. Nat. Phys. 12,
Opt. Commun. 407, 338–343 (2018). 55. Huo, P. et al. Photonic spin-​multiplexing metasurface 621–623 (2016).
28. Zhang, J., Ying, Q. & Ruan, Z. Time response of for switchable spiral phase contrast imaging. Nano Lett. 85. Zangeneh-​Nejad, F. & Fleury, R. Nonlinear second-​
plasmonic spatial differentiators. Opt. Lett. 44, 4, 2791–2798 (2020). order topological insulators. Phys. Rev. Lett. 123,
4511–4514 (2019). 56. Wang, H. et al. Off-​axis holography with uniform 053902 (2019).
29. Hwang, Y., Davis, T. J., Lin, J. & Yuan, X. C. Plasmonic illumination via 3D printed diffractive optical 86. Zangeneh-​Nejad, F. & Fleury, R. Topological analog
circuit for second-​order spatial differentiation at the elements. Adv. Opt. Mater. 7, 1900068 (2019). signal processing. Nat. Commun. 10, 2058 (2019).
subwavelength scale. Opt. Express 26, 7368–7375 57. Zuo, S. Y., Wei, Q., Cheng, Y. & Liu, X. J. Mathematical 87. Xiao, Y. X., Ma, G., Zhang, Z. Q. & Chan, C. T.
(2018). operations for acoustic signals based on layered Topological subspace-​induced bound state in the
30. Dong, Z., Si, J., Yu, X. & Deng, X. Optical spatial labyrinthine metasurfaces. Appl. Phys. Lett. 110, continuum. Phys. Rev. Lett. 118, 166803 (2017).
differentiator based on subwavelength high-​contrast 011904 (2017). 88. Jalas, D. et al. What is — and what is not — an optical
gratings. Appl. Phys. Lett. 112, 181102 (2018). 58. Kamali, S. M. et al. Angle-​multiplexed metasurfaces: isolator. Nat. Photonics 7, 579–582 (2013).
31. Bezus, E. A., Doskolovich, L. L., Bykov, D. A. & encoding independent wavefronts in a single 89. Zangeneh-​Nejad, F. et al. Nonreciprocal manipulation
Soifer, V. A. Spatial integration and differentiation of metasurface under different illumination angles. of subwavelength fields in locally resonant
optical beams in a slab waveguide by a dielectric ridge Phys. Rev. X 7, 041056 (2017). metamaterial crystals. IEEE Trans. Antennas Propag.
supporting high-​Q resonances. Opt. Express 26, 59. Liu, Z., Zhu, D., Rodrigues, S. P., Lee, K. T. & Cai, W. 68, 1726–1732 (2019).
25156–25165 (2018). Generative model for the inverse design of 90. Zhang, W. & Zhang, X. Backscattering-​immune
32. Lv, Z., Ding, Y. & Pei, Y. Acoustic computational metasurfaces. Nano Lett. 18, 6570–6576 (2018). computing of spatial differentiation by nonreciprocal
metamaterials for dispersion Fourier transform in 60. Pfeiffer, C. & Grbic, A. Bianisotropic metasurfaces for plasmonics. Phys. Rev. Appl. 11, 054033 (2019).
time domain. J. Appl. Phys. 127, 123101 (2020). optimal polarization control: Analysis and synthesis. 91. Alpaydin, E. Introduction to Machine Learning
33. Eftekhari, F., Gómez, D. E. & Davis, T. J. Measuring Phys. Rev. Appl. 2, 044011 (2014). (MIT Press, 2020).
subwavelength phase differences with a plasmonic 61. Bao, L. et al. Design of digital coding metasurfaces 92. Beale, H. D., Demuth, H. B. & Hagan, M. T.
circuit — an example of nanoscale optical signal with independent controls of phase and amplitude Neural Network Design (PWS, 1996).
processing. Opt. Lett. 39, 2994–2997 (2014). responses. Appl. Phys. Lett. 113, 063502 (2018). 93. Orazbayev, B. & Fleury, R. Far-​field subwavelength
34. Guo, C., Xiao, M., Minkov, M., Shi, Y. & Fan, S. Isotropic 62. Liu, T., Chen, F., Liang, S., Gao, H. & Zhu, J. acoustic imaging by deep learning. Phys. Rev. X 10,
wavevector domain image filters by a photonic crystal Subwavelength sound focusing and imaging via 031029 (2020).
slab device. JOSA A 35, 1685–1691 (2018). gradient metasurface-​enabled spoof surface acoustic 94. Chakrabarti, A. Learning sensor multiplexing design
35. Guo, C., Xiao, M., Minkov, M., Shi, Y. & Fan, S. wave modulation. Phys. Rev. Appl. 11, 034061 through back-​propagation. Adv. Neural Inf. Process.
Photonic crystal slab Laplace operator for image (2019). Syst. 29, 3081–3089 (2016).
differentiation. Optica 5, 251–256 (2018). 63. Sun, S., He, Q., Hao, J., Xiao, S. & Zhou, L. 95. Horstmeyer, R., Chen, R. Y., Kappes, B. &
36. Davis, T. J., Eftekhari, F., Gómez, D. E. & Roberts, A. Electromagnetic metasurfaces: physics and Judkewitz, B. Convolutional neural networks that
Metasurfaces with asymmetric optical transfer applications. Adv. Opt. Photonics 11, 380–479 teach microscopes how to image. Preprint at arXiv
functions for optical signal processing. Phys. Rev. Lett. (2019). https://arxiv.org/abs/1709.07223 (2017).
123, 013901 (2019). 64. Zhang, L., Mei, S., Huang, K. & Qiu, C. W. Advances 96. Lin, X. et al. All-​optical machine learning using
37. Wu, W., Jiang, W., Yang, J., Gong, S. & Ma, Y. in full control of electromagnetic waves with diffractive deep neural networks. Science 361,
Multilayered analog optical differentiating device: metasurfaces. Adv. Opt. Mater. 4, 818–833 (2016). 1004–1008 (2018).
performance analysis on structural parameters. 65. He, Q., Sun, S., Xiao, S. & Zhou, L. High-​efficiency 97. Hunt, J. et al. Metamaterial apertures for computational
Opt. Lett. 42, 5270–5273 (2017). metasurfaces: principles, realizations, and imaging. Science 339, 310–313 (2013).
38. Fang, Y., Lou, Y. & Ruan, Y. On-​grating graphene surface applications. Adv. Opt. Mater. 6, 1800415 (2018). 98. Gollub, J. N. et al. Large metasurface aperture
plasmons enabling spatial differentiation in the 66. La Spada, L., Spooner, C., Haq, S. & Hao, Y. for millimeter wave computational imaging at the
terahertz region. Opt. Lett. 42, 3840–3843 (2017). Curvilinear metasurfaces for surface wave human-​scale. Sci. Rep. 7, 42650 (2017).
39. Idemen, M. M. Discontinuities in the Electromagnetic manipulation. Sci. Rep. 9, 3107 (2019). 99. Sleasman, T., Imani, M. F., Gollub, J. N. & Smith, D. R.
Field Vol. 40 (Wiley, 2011). 67. Zuo, S. Y., Tian, Y., Wei, Q., Cheng, Y. & Liu, X. J. Dynamic metamaterial aperture for microwave
40. Youssefi, A., Zangeneh-​Nejad, F., Abdollahramezani, S. Acoustic analog computing based on a reflective imaging. Appl. Phys. Lett. 107, 204104 (2015).
& Khavasi, A. Analog computing by Brewster effect. metasurface with decoupled modulation of phase 100. Sleasman, T. et al. Implementation and
Opt. Lett. 41, 3467–3470 (2016). and amplitude. J. Appl. Phys. 123, 091704 (2018). characterization of a two-​dimensional printed circuit
41. Chen, Z. et al. Graphene controlled Brewster angle 68. Hwang, Y. & Davis, T. J. Optical metasurfaces for dynamic metasurface aperture for computational
device for ultra broadband terahertz modulation. subwavelength difference operations. Appl. Phys. Lett. microwave imaging. Preprint at arXiv https://arxiv.org/
Nat. Commun. 9, 4909 (2018). 109, 181101 (2016). abs/1911.08952 (2019).
42. Zangeneh-​Nejad, F. & Khavasi, A. Spatial integration 69. Ozaktas, H. M. & Mendlovic, D. Fourier transforms 101. Li, L. et al. Intelligent metasurface imager and
by a dielectric slab and its planar graphene-​based of fractional order and their optical interpretation. recognizer. Light Sci. Appl. 8, 97 (2019).
counterpart. Opt. Lett. 42, 1954–1957 (2017). Opt. Commun. 101, 163–169 (1993). 102. Li, L. et al. Machine-​learning reprogrammable
43. Zhu, T. et al. Plasmonic computing of spatial 70. Monticone, F., Estakhri, N. M. & Alu, A. Full control metasurface imager. Nat. Commun. 10, 1082 (2019).
differentiation. Nat. Commun. 8, 15391 (2017). of nanoscale optical transmission with a composite 103. Liang, M., Li, Y., Meng, H., Neifeld, M. A. & Xin, H.
44. Ma, C., Kim, S. & Fang, N. X. Far-​field acoustic metascreen. Phys. Rev. Lett. 110, 203903 (2013). Reconfigurable array design to realize principal
subwavelength imaging and edge detection based 71. Pors, A., Nielsen, M. G. & Bozhevolnyi, S. I. Analog component analysis (PCA)-based microwave
on spatial filtering and wave vector conversion. computing using reflective plasmonic metasurfaces. compressive sensing imaging system. IEEE Antennas
Nat. Commun. 10, 204 (2019). Nano Lett. 15, 791–797 (2015). Wirel. Propag. Lett. 14, 1039–1042 (2015).
45. Zhu, T. et al. Generalized spatial differentiation from 72. Chizari, A., Abdollahramezani, S., Jamali, M. V. 104. del Hougne, P., Imani, M. F., Diebold, A. V.,
the spin Hall effect of light and its application in image & Salehi, J. A. Analog optical computing based Horstmeyer, R. & Smith, D. R. Learned integrated
processing of edge detection. Phys. Rev. Appl. 11, on a dielectric meta-​reflect array. Opt. Lett. 41, sensing pipeline: Reconfigurable metasurface
034043 (2019). 3451–3454 (2016). transceivers as trainable physical layer in an artificial
46. Kwon, H., Sounas, D., Cordaro, A., Polman, A. 73. Zuo, S., Wei, Q., Tian, Y., Cheng, Y. & Liu, X. Acoustic neural network. Adv. Sci. 7, 1901913 (2020).
& Alù, A. Nonlocal metasurfaces for optical signal analog computing system based on labyrinthine 105. Li, H. Y. et al. Intelligent electromagnetic sensing with
processing. Phys. Rev. Lett. 121, 173004 (2018). metasurfaces. Sci. Rep. 8, 10103 (2018). learnable data acquisition and processing. Patterns 1,
47. Cordaro, A., Kwon, H., Sounas, D., Koenderink, A. F., 74. Zangeneh-​Nejad, F. & Fleury, R. Performing 100006 (2020).
Alu, A. & Polman, A. High-​index dielectric mathematical operations using high-​index acoustic 106. Shen, Y. et al. Deep learning with coherent nanophotonic
metasurfaces performing mathematical operations. metamaterials. New J. Phys. 20, 073001 (2018). circuits. Nat. Photonics 11, 441–446 (2017).
Nano Lett. 19, 8418–8423 (2019). 75. Zangeneh-​Nejad, F. & Fleury, R. Acoustic analogues 107. del Hougne, P. & Lerosey, G. Leveraging chaos for
48. Kwon, H., Cordaro, A., Sounas, D., Polman, A. & of high-​index optical waveguide devices. Sci. Rep. 8, wave-​based analog computation: Demonstration with
Alù, A. Dual-​polarization analog 2D image processing 10401 (2018). indoor wireless communication signals. Phys. Rev. X
with nonlocal metasurfaces. ACS Photonics 7, 76. Barrios, G. A., Retamal, J. C., Solano, E. & Sanz, M. 8, 041037 (2018).
1799–1805 (2020). Analog simulator of integro-​differential equations 108. Matthès, M. W., del Hougne, P., de Rosny, J.,
49. Zhou, Y. et al. Analog optical spatial differentiators with classical memristors. Sci. Rep. 9, 12928 (2019). Lerosey, G. & Popoff, S. M. Optical complex media as
based on dielectric metasurfaces. Adv. Opt. Mater. 8, 77. Zhang, W., Qu, C. & Zhang, X. Solving constant-​ universal reconfigurable linear operators. Optica 6,
1901523 (2020). coefficient differential equations with dielectric 465–472 (2019).
50. Chen, H., An, D., Li, Z. & Zhao, X. Performing metamaterials. J. Opt. 18, 075102 (2016). 109. Pierangeli, D., Marcucci, G. & Conti, C. Large-​scale
differential operation with a silver dendritic 78. Zangeneh-​Nejad, F. & Fleury, R. Acoustic rat-​race photonic Ising machine by spatial light modulation.
metasurface at visible wavelengths. Opt. Express coupler and its applications in non-​reciprocal systems. Phys. Rev. Lett. 122, 213902 (2019).
25, 26417–26426 (2017). J. Acoust. Soc. Am. 146, 843–849 (2019). 110. Abdolali, A., Momeni, A., Rajabalipanah, H. &
51. Roberts, A., Gómez, D. E. & Davis, T. J. Optical image 79. Estakhri, N. M., Edwards, B. & Engheta, N. Inverse-​ Achouri, K. Parallel integro-​differential equation
processing with metasurface dark modes. JOSA A 35, designed metastructures that solve equations. Science solving via multi-​channel reciprocal bianisotropic
1575–1584 (2018). 363, 1333–1338 (2019). metasurface augmented by normal susceptibilities.
52. Wang, L., Li, L., Li, Y., Zhang, H. C. & Cui, T. J. 80. Zangeneh-​Nejad, F. & Fleury, R. Topological fano New J. Phys. 21, 113048 (2019).
Single-​shot and single-​sensor high super-​resolution resonances. Phys. Rev. Lett. 122, 014301 (2019). 111. Wu, Y. et al. Arbitrary multi-​way parallel mathematical
microwave imaging based on metasurface. Sci. Rep. 6, 81. Rechtsman, M. C. et al. Photonic Floquet topological operations based on planar discrete metamaterials.
26959 (2016). insulators. Nature 496, 196–200 (2013). Plasmonics 13, 599–607 (2018).

224 | March 2021 | volume 6 www.nature.com/natrevmats

 Reviews

112. Momeni, A., Rajabalipanah, H., Abdolali, A. & resonance frequency, absorption, and bandwidth. 136. de Solla Price, D. Gears from the Greeks. The
Achouri, K. Generalized optical signal processing Phys. Rev. Appl. 12, 054064 (2019). mechanism: a calendar computer from ca. 80 BC.
based on multioperator metasurfaces synthesized by 124. Zibar, D., Wymeersch, H. & Lyubomirsky, I. Machine Trans. Am. Philos. Soc. 64, 31–70 (1974).
susceptibility tensors. Phys. Rev. Appl. 11, 064042 learning under the spotlight. Nat. Photonics 11, 137. Carr, B. Astronomical clocks: probing the early
(2019). 749–751 (2017). Universe with the millisecond pulsar. Nature 315,
113. Mosk, A. P., Lagendijk, A., Lerosey, G. & Fink, M. 125. Zhou, H. et al. Self-​learning photonic signal processor 540 (1985).
Controlling waves in space and time for imaging with an optical neural network chip. Preprint at arXiv 138. I. Szalkai. General two-​variable functions on the
and focusing in complex media. Nat. Photonics 6, https://arxiv.org/abs/1902.07318v1 (2019). slide-​rule. Preprint at arXiv https://arxiv.org/abs/
283–292 (2012). 126. Hughes, T. W., Williamson, I. A., Minkov, M. & Fan, S. 1612.03955 (2016).
114. Lee, P. A. & Fisher, D. S. Anderson localization in two Wave physics as an analog recurrent neural network. 139. Jenkins, H. V. An airflow planimeter for measuring the
dimensions. Phys. Rev. Lett. 47, 882 (1981). Sci. Adv. 5, eaay6946 (2019). area of detached leaves. Plant. Physiol. 34, 532–536
115. Titum, P., Lindner, N. H., Rechtsman, M. C. & Refael, G. 127. Tahersima, M. H. et al. Deep neural network inverse (1959).
Disorder-​induced Floquet topological insulators. design of integrated photonic power splitters. Sci. Rep. 140. Chapman, R. W. A simple form of tide predictor.
Phys. Rev. Lett. 114, 056801 (2015). 9, 1368 (2019). Nature 68, 322 (1903).
116. Zangeneh-​Nejad, F. & Fleury, R. Disorder-​induced 128. Yan, T. Fourier-​space diffractive deep neural network. 141. Lawshe, C. H. Jr. A nomograph for estimating the
signal filtering with topological metamaterials. Phys. Rev. Lett. 123, 023901 (2019). validity of test items. J. Appl. Psychol. 26, 846–849
Adv. Mater. 32, 2001034 (2020). 129. Luo, Y. et al. Design of task-​specific optical systems (1942).
117. He, S. et al. Broadband optical fully differential using broadband diffractive neural networks. Light Sci.
operation based on the spin-​orbit interaction of light. Appl. 8, 112 (2019). Acknowledgements
Preprint at arXiv https://arxiv.org/abs/1910.09789 130. Zhou, Y., Chen, R., Chen, W., Chen, R. P. & Ma, Y. R.F. and F.Z.-​N. would like to acknowledge the support from
(2019). Optical analog computing devices designed by deep the Swiss National Science Foundation, under grant number
118. He, S. et al. Spatial differential operation and edge neural network. Opt. Commun. 458, 124674 (2020). 172487. A.A. and D.L.S. acknowledge the support from the
detection based on the geometric spin Hall effect 131. Guo, X., Lissek, H. & Fleury, R. Improving sound Air Force Office of Scientific Research, the Simons Foundation
of light. Opt. Lett. 45, 877–880 (2020). absorption through nonlinear active electroacoustic and the National Science Foundation. The authors acknowl-
119. Popa, B. I. & Cummer, S. A. Non-​reciprocal and highly resonators. Phys. Rev. Appl. 13, 014018 (2020). edge useful discussions with P. del Hougne about
nonlinear active acoustic metamaterials. Nat. Commun. 132. Inagaki, T. et al. A coherent Ising machine for 2000- machine-​learning applications of computing metamaterials.
5, 3398 (2014). node optimization problems. Science 354, 603–606
120. Hess, O. et al. Active nanoplasmonic metamaterials. (2016). Author contributions
Nat. Mater. 11, 573–584 (2012). 133. Marandi, A., Wang, Z., Takata, K., Byer, R. L. & All authors contributed equally to writing the article.
121. Chen, K. et al. A reconfigurable active Huygens’ Yamamoto, Y. Network of time-​multiplexed optical
metalens. Adv. Mater. 29, 1606422 (2017). parametric oscillators as a coherent Ising machine. Competing interests
122. Ghatak, A., Brandenbourger, M., van Wezel, J. Nat. Photonics 8, 937–942 (2014). The authors declare no competing interests.
& Coulais, C. Observation of non-​Hermitian 134. Yang, T. et al. All-​optical differential equation solver
topology and its bulk-​edge correspondence. with constant-​coefficient tunable based on a single Publisher’s note
Preprint at arXiv https://arxiv.org/abs/1907.11619 microring resonator. Sci. Rep. 4, 5581 (2014). Springer Nature remains neutral with regard to jurisdictional
(2019). 135. MacLennan, B. J. A review of analog computing. claims in published maps and institutional affiliations.
123. Koutserimpas, T. T., Rivet, E., Lissek, H. & Fleury, R. Univ. Tennessee https://library.eecs.utk.edu/pub/123
Active acoustic resonators with reconfigurable (2007). © Springer Nature Limited 2020

NaTure RevIewS | MAteRiAls volume 6 | March 2021 | 225