REVIEW

www.lpr-journal.org

Numerical Optimization Methods for Metasurfaces

Mahmoud M. R. Elsawy, Stéphane Lanteri,* Régis Duvigneau, Jonathan A. Fan,
and Patrice Genevet *

plasmonic or high dielectric refractive

In recent years, metasurfaces have emerged as revolutionary tools to index materials, which have thicknesses
manipulate the behavior of light at the nanoscale. These devices consist of within the range of the operating wave-
nanostructures defined within a single layer of metal or dielectric materials, length. The design of metasurfaces is
and they offer unprecedented control over the optical properties of light, generally achieved using the following
two approaches. In the first approach,
leading to previously unattainable applications in flat lenses, holographic which we will refer to as the direct
imaging, polarimetry, and emission control, amongst others. The operation design approach, a rigorous full wave
principles of metaoptics include complex light–matter interactions, often electromagnetic solver is used to study
involving insidious near-field coupling effects that are far from being different classes of meta-atoms. Large
described by classical ray optics calculations, making advanced numerical sets of key meta-atom parameters are
parametrically swept to create meta-atom
modeling a requirement in the design process. In this contribution, recent
libraries, and ensembles of meta-atoms
optimization techniques used in the inverse design of high performance from these libraries are assembled to-
metasurfaces are reviewed. These methods rely on the iterative optimization gether to create metasurfaces with given
of a Figure of Merit to produce a final device, leading to freeform layouts desired optical responses. This method
featuring complex and non-intuitive properties. The concepts in numerical works well for certain applications. How-
inverse designs discussed herein will push this exciting field toward realistic ever, it does not incorporate near-field
electromagnetic coupling effects be-
and practical applications, ranging from laser wavefront engineering to tween neighboring meta-atoms and does
innovative facial recognition and motion detection devices, including not generalize to large area, freeform
augmented reality retro-reflectors and related complex light field engineering. devices.
The second approach is inverse de-
sign. With this approach, the desired op-
tical response is defined as an objective
cost function (for example, the deflection
1. Introduction or focusing efficiencies) and the inverse problem solves for the
shape and dimensions of the metasurface in a manner that
During the last decade, metasurfaces have received lots of at-
maximizes the cost function value. In principle, the inverse
tention due to their ability to precisely control the phase, ampli-
design approach can account for near-field interactions by opti-
tude, and wavefront of light. These light–matter interactions are
mizing relatively large metasurface regions at a time. However,
mediated by ensembles of subwavelength meta-atoms, made of
it requires rigorous and computationally efficient optimization
techniques that can deal with large parameter spaces and multi-
objective cost functions. There are two main classes of optimiza-
tion methods that have currently been used in inverse design
Dr. M. M. R. Elsawy, Prof. S. Lanteri, Dr. R. Duvigneau
Université Côte d’Azur of metasurfaces: gradient-based algorithms and gradient-free
Inria, CNRS, LJAD approaches.
06902 Sophia Antipolis Cedex, France Gradient-based methods depend strongly on the initial guess
E-mail: [email protected] of the solution and are efficient in finding local optima. Gradient-
Prof. J. A. Fan based method require the knowledge of the derivatives of the
Department of Electrical Engineering
Stanford University cost function with respect to design parameters, which can be
Stanford, CA 94305, USA evaluated analytically sometimes, or approximated numerically
Dr. P. Genevet in most cases.[1] These methods have a lengthy history in
CNRS, CRHEA metasurface design and were used early in the field[2,3] to create
Université Côte d’Azur devices that maximized light diffraction at visible wavelengths.
06560 Sophia Antipolis Cedex, France
Today, several descent methods, ranging from steepest-descent to
E-mail: [email protected]
quasi-Newton methods for both constrained and unconstrained
The ORCID identification number(s) for the author(s) of this article
problems, have been applied to metasurfaces. In this short
can be found under https://doi.org/10.1002/lpor.201900445 review, we focus on the most common methods used recently
in the literature, including the objective-first[1,4] and topology
DOI: 10.1002/lpor.201900445

Laser Photonics Rev. 2020, 1900445 1900445 (1 of 17) © 2020 Wiley-VCH GmbH
www.advancedsciencenews.com www.lpr-journal.org

optimization algorithms, which have proven to be efficient and 2. Gradient-Based Optimization Techniques
rigorous design methodologies.[4–8]
Gradient-free approaches are capable of capturing global 2.1. Objective-First Algorithm
optima,[9] albeit for a limited number of parameters, thereby over-
coming the local minimum trapping issue of gradient-based al- The objective-first algorithm is a widely used optimization tech-
gorithms. In other words, with gradient-free global optimization nique in nanophotonics.[1,9] As described in Chapter 6 in Ref. [1]
techniques, the final optimized results are not influenced by the and in refs. [6, 26], the algorithm begins by defining quantities to
initialization of the optimizer. In addition, some of these algo- optimize. After specifying an objective or target function describ-
rithms can deal with discrete optimization parameters and non- ing these quantities, the algorithm searches for the spatial dis-
differentiable objective functions, which are conditions that are tribution of dielectric material that maximizes this target while
generally not handled by gradient-based algorithms. Without gra- satisfying Maxwell’s equations as accurately as possible.[6] For
dients to help guide the optimization, convergence with global most of the applications discussed in this manuscript, we con-
optimization algorithms is often considerably slower than con- sider nonmagnetic materials, such that the algorithm searches
vergence with gradient-based algorithms. To date, several global for permittivity distributions within a design window.
optimization techniques have been proposed in the context of Following the notations in Chapter 6 of Ref. [1] and assuming a
metasurface design. To properly deal with the extensive (often time-harmonic dependency of the electromagnetic field, we write
discrete) parameter space and the existence of several local op- the general optimization problem as
tima, the majority of inverse design methods of interest for meta-
surface design are stochastic and include genetic algorithms and minimize f (x)
x,p
evolutionary strategies.[10–18] (1)
In addition to the methods discussed above, emergent ap- subject to A(p(x))x − b(p(x)) = 0,
proaches, including artificial neural networks and Bayesian opti-
mization, have the potential to uncover surprising new metasur- A change of variables is applied to match this problem with
face designs. We will highlight in this review the key ideas behind electromagnetics: H → x, 𝜖 −1 → p such that A(p) = ∇ × 𝜖 −1 ∇ ×
these techniques and illustrate their versatility and advantages for −𝜇0 𝜔2 , and b(p) = ∇ × 𝜖 −1 J. J is the current density vector, 𝜇0 is
the optimization of practical metasurfaces.[19,20] the vacuum permeability, and 𝜔 is the angular frequency. Equa-
We note there have been a number of reviews about inverse de- tion (1) indicates that the minimization of the target function,
sign for nanophotonics that have been published recently.[21–23] achieved by varying H and 𝜖 −1 simultaneously, is performed
In Ref. [21], the authors focus on inverse design in the frame- while satisfying the wave equation.[1] Note that Equation (1) is a
work of deep learning. In Ref. [22], the authors provide gen- priori a non-convex problem, since this requires solving for p(x)
eral guidance for different optimization techniques, indicating and x simultaneously, which is in general a difficult problem.
their performances with respect to a range of nanophotonics The objective-first algorithm splits the optimization problem
problems. In Ref. [23], the authors present a collection inverse into two convex subproblems. One of these subproblems deals
design techniques in nanophotonics and their specific applica- with the fields: given the permittivity, it solves for the usual wave
tion to nonlinear optics, integrated optics, and topological pho- equation and determines the fields that minimize the residual.
tonics. What differentiates our review from the previous one The second subproblem solves for the permittivity distribution
is that we we focus here our on numerical methods used for given constant electromagnetic fields. After this minimization,
designing metasurfaces, providing a simple and clear illustra- both subproblems are then merged together[1] :
tion of these techniques and demonstrating their advantages
and drawbacks. Indeed, we present an overview (with a dis-
cussion on the numerical implementation) of the main and Loop:
most common optimization methods for metasurface designs. ‖ A(p)x − b(p)‖2
minimize ‖ ‖
We do not intend to present an exhaustive literature review on x
nanophotonic structures optimized using inverse design tech-
niques. This topic has already been covered in recently published subject to f (x) = fideal , (2)
works.[21–25] ‖ B(x)p − d(x)‖2
The article is organized as follows. In Section 2, we discuss minimize ‖ ‖
p
the main gradient-based optimization techniques used in the
metasurface literature: the objective-first and topology optimiza- subject to pmin ≤ p ≤ pmax ,
tion approaches. In Section 3, we introduce the most popu-
lar gradient-free techniques, that is, genetic algorithms, particle fideal is the ideal performance for function f (x). We see that
swarm optimization, and covariance matrix adaptation evolution the first subproblem tries to converge to an ideal solution that
strategy. In Section 4, we give a simple and practical introduc- satisfies the wave equation up to some residual. The second sub-
tion to artificial neural networks and how they can be applied to problem, with B(x) = ∇ × (∇ × H) − ∇ × J and d(x) = 𝜇0 𝜔2 H,
the inverse design of metasurface devices. In Section 5, we intro- seeks to solve for the permittivity with the condition that it takes
duce the concept of Bayesian optimization and discuss its appli- continuous values. In practice, the optimization domain that cor-
cation to inverse design for nanophotonics. Finally, in Section 6, responds to the physical space containing the nanostructures is
we discuss methods to incorporate robustness into the optimiza- decomposed into equally spaced pixels, each specified by a given
tion process. local permittivity value. The optimization process is iteratively

Laser Photonics Rev. 2020, 1900445 1900445 (2 of 17) © 2020 Wiley-VCH GmbH
www.advancedsciencenews.com www.lpr-journal.org

Figure 1. Optimized devices using the objective-first algorithm. a) Schematic representation of a wavelength demultiplexing grating coupler device.
This device is illuminated with a normally incident light from above, and splits the light (according to the wavelength). b) The optimization process
for directional modal coupling. The optimization process supposes, during the first evaluation round, that the permittivity varies in a continuous way
(stage 1). In the second evaluation round, the permittivity is converted to a binary-set representation between Si and air (stage 2). The right column
figure is taken from Ref. [5], in which several metadevices have been optimized using the objective-first algorithm. Here, the desired optical functionality
is defined by a set of input and output conditions at the boundaries of the design space (black dots). The top figure represents a polarization splitter,
the middle one is a bending device, and the bottom one, a device to convert an incident plane wave into lens to focus light at a single focal point. (a,b)
Reproduced with permission.[7] Copyright 2015, Springer Nature. (c) Reproduced with permission.[5] Copyright 2018, Springer Nature.

employed until convergence, and until it reaches a final structure the solutions satisfying Maxwell equation all the way along the
that reasonably satisfies the objective.[1,6,7,27,28] optimization procedure.
To summarize, and as it is mentioned in Ref. [28], the ob- We note that since the algorithm relies on a gradient-based
jective first algorithm ingeniously splits Equation (1) into two technique, it does not directly apply to structures comprising dis-
convex sub-problems as shown in Equation (2), and uses local crete representations of the permittivity. As such, a subsequent
optimization approaches based on convex optimization[29] to stage of discrete optimization, based on a binary representation
effectively explore the huge parameter space. With respect to of the structure, is required and achieved using a separate opti-
classical gradient-based optimization methods, which stick to mization method (see refs. [5, 6, 26] for more details). The latter
physically realistic solution able to satisfy the wave equation, the step is of critical importance when considering practical experi-
objective first method treats each sub-problem sequentially, em- mental device realizations.
ploying the alternating directions method, solving for p and x[30] In Figure 1, we present several examples of nanophotonic
iteratively. The resolution of the first sub-problem optimizes the devices optimized using this objective-first algorithm. In the first
performance but does not check if the solution satisfies the wave column, a demultiplexing grating waveguide that splits an inci-
equation. This “violation,” as discussed in refs. [27, 28], is regular- dent free-space Gaussian beam into left-going O-band (1300 nm)
ized through the minimization of the second sub-problem. The and right-going C-band (1550 nm) waveguided modes has been
first experimental demonstration of the objective first algorithm optimized.[7] The fabricated device has a measured splitting ratio
has been reported in Ref. [28], where, the objective first algorithm of 17 dB at 1310 nm and 12 dB at 1540 nm, whereas the designed
is followed by an adjoint-based gradient method to finely tune values at these wavelengths were 19.6 and 22.2 dB. In the second
the structure[28] by implementing classical steepest-descent al- column of Figure 1, the objective-first algorithm was used to opti-
gorithm. It has been shown that the final device performs better mize all-dielectric devices with different functionalities[5] includ-
than those optimized using only adjoint optimization method. ing polarization splitting (top figure), light deflection (middle fig-
For more details about the mathematical implementation of ure), and light focusing. Several other nanophotonic devices have
this method, we refer to Chapter 7 and Appendix. C in Ref. [30] been optimized using this technique, see for example Ref. [1] for
together with Refs. [27, 28]. optical cloaks, Ref. [6] for the optimization of a broadband optical
It is important to distinguish this objective first optimization diodes and Ref. [26] for the optimization of a 1D grating coupler.
method from other topology optimization (TO) and genetic It is worth mentioning that in most of the reported cases, the
algorithm (GA) methods discussed in the next paragraph. agreement in efficiency between optimized devices and the fab-
Objective first solutions are constrained to satisfy Maxwell ricated ones is quite low. Discrepancies are generally related to
equations using convergence process of initially nonphysical the difficulties in properly addressing the change from continu-
solutions while other methods, including TO and GA, optimize ous to the binary representation. To correctly binarize designs in

Laser Photonics Rev. 2020, 1900445 1900445 (3 of 17) © 2020 Wiley-VCH GmbH
www.advancedsciencenews.com www.lpr-journal.org

a precise way, subsequent efficient optimization techniques that Metagrating devices that maximize deflection efficiency at
can mimic this clear-cut simplification must be considered.[26] high deflection angles and near-infrared wavelengths have
been optimized using TO and demonstrated experimentally in
Ref. [37] The optimization algorithm starts with a random and
2.2. Topology Optimization continuous distribution of permittivity values between the values
of Si and air, and these permittivity are iteratively optimized in a
In this section, we will review a second class of optimization manner that optimizes the cost function value. As can be seen
methods, namely topology optimization (TO), used in the inverse from Figure 2a, the gray scale values of the dielectric constant
design of nanophotonic devices.[31] TO historically has been ap- are pushed toward air or Si as the optimization proceeds, which
plied to a broad range of physical systems such as mechanical is driven by the use of penalty terms in the cost function. As a
structures, MEMS, and materials design Ref. [32], and it was in- result, final devices have a binary layout of Si in air. These de-
troduced to nanophotonics in the early 2000s in the context of vices are able to achieve high efficiencies due to non-trivial mul-
photonic crystal-based technologies.[33,34] More recently, it has tiple scattering effects mediated by the presence of high order
been applied to the design of linear[35] and nonlinear[36] metasur- optical modes in high contrast dielectric structures.[38,39] TO has
faces. TO is based on a gradient-based algorithm that is able to been readily extended to other variants of periodic diffractive op-
produce freeform geometric configurations. This method works tical structures, including those exhibiting ultra-high anomalous
as a local optimizer, starting from an initial guess for the config- refraction[40] and the diffraction of different wavelengths to dif-
uration of the device and then undergoing an iterative process to ferent diffraction angles.[41]
achieve a locally optimal configuration.[35] TO-based metasurfaces have also been extended to the design
With TO, the device is subdivided into an ensemble of pixels, of wavelength-scale scatterers with defined scattering directions
for which each pixel is associated to a design variable such as the and phases, as shown in Figure 2b.[42] The final devices comprise
dielectric permittivity. The dimensions of these pixels can be as single crystal silicon,[43] scattering light at visible wavelengths
small as a few square nanometers for visible light devices, such with strong directionality in the desired direction, in a process
that the final devices can comprise smooth, curvilinear shapes. that is mediated by strong near-field interactions between nanos-
Often, the goal is to produce a final device that consists of di- tructures. These wavelength-scale building blocks can be stitched
electric materials with discrete permittivity values, such as an together to produce high efficiency aperiodic metasurfaces, such
isotropic semiconducting or insulating material with air voids. as metalenses. Aperiodic metasurfaces have also been considered
However, as TO is a gradient-based algorithm, the optimization in refs. [4, 36], in which efficient metalens devices and photonic
is required to be performed on a continuum of permittivity val- fibers for nonlinear frequency conversion are optimized using
ues. Consider as an example a final device that comprises two TO. However, near field coupling is crucial and highly affects the
dielectric materials with permittivities, 𝜖1 and 𝜖2 . These permit- performance of the final optimized design. In Ref. [44], the au-
tivities can be normalized to take values of “0” and “1,” respec- thors introduced an overlap technique to their topology optimiza-
tively, and the actual parameter undergoing optimization will be tion approach to take into account the near field coupling from
a continuous parameter 0 ≤ s ≤ 1.[31,32] This continuous repre- neighbouring unit cells. They applied their technique to optimize
sentation leads to optimized regions with intermediate dielectric a large-scale metalens, demonstrating higher efficiency with re-
values between the two desired discrete permittivities. To ensure spect to the solution obtained with the classical local approxima-
that TO produces fully discrete devices, terms can be added to the tion technique.[45]
objective function that penalize the presence of grayscale dielec- The influence of the initial guess (initial geometry) on the over-
tric values.[31,32] all performance of the optimized design has been discussed in
We present below a general and quick overview of the opti- Ref. [35]. It is shown that conventional metasurface devices serv-
mization problem in the framework of TO (see also refs. [31, 36]). ing as starting points for optimization do not produce highly ef-
The optimization problem reads[4] : ficient topology optimized devices. Instead, random initial guess
geometries have the potential to yield final devices with ultrahigh
efficiencies (see right column in Figure 2).
minimize F(E, 𝜖)
̃
𝜖̃
(3)
subject to G(E, 𝜖)
̃ ≤ 0, 0 ≤ 𝜖̃ ≤ 1.
3. Gradient-Free Optimization Techniques
Here, 𝜖̃ is the normalized dielectric permittivity associated to 3.1. Genetic Algorithm
each pixel in a specific volume. The values of 𝜖̃ are related to the
position-dependent dielectric profile via some linear interpola- A Genetic Algorithm (GA) is a metaheuristic inspired by the pro-
tion function.[36] Note that both the objective function (F) and the cess of natural selection that belongs to the larger class of evo-
constraints (G) are a function of the permittivity and the electric lutionary algorithms (EA). GAs are commonly used to generate
field E, which is a solution to Maxwell’s equations. Equation (3) high quality solutions to optimization and search problems by re-
can be solved using mathematical techniques such as the Method lying on bio-inspired operators such as mutation, crossover, and
of Moving Asymptotes.[31] However, in the framework of TO, selection. In a GA, a population of candidate solutions to an opti-
the derivative of the objective function and the constraints with mization problem, called individuals, creatures, or phenotypes, is
respect to 𝜖̃ have to be computed at each pixel. This can be treated evolved toward better solutions. Each candidate solution has a set
using methods such as adjoint variable method.[31,35,36] of properties (i.e., its chromosomes or genotype) that is iteratively

Laser Photonics Rev. 2020, 1900445 1900445 (4 of 17) © 2020 Wiley-VCH GmbH
www.advancedsciencenews.com www.lpr-journal.org

Figure 2. Metasurface design using topology optimization. a) Metagrating efficiency as a function of iteration number. The insets show the dielectric
distribution in the course of the optimization process. Reproduced with permission.[37] Copyright 2017, American Chemical Society. b) Design of a
wavelength-scale scatterer with a scattering direction of 20 degrees and phase response of 45 degrees. Reproduced with permission.[42] Copyright 2019,
Springer Nature. c) Topology-optimized metagratings designed using different initial dielectric distributions. Reproduced with permission.[35] Copyright
2017, Optical Society of America.

mutated and altered. Traditionally, solutions are represented in in which a random chromosome location is first chosen. Then,
binary as strings of 0s and 1s, but other encoding methods such the chromosome of child 1 consists of an initial chromosome
as mixed-integer heuristics, have been proposed.[46,47] GAs can segment from parent 1 spanning the start of the chromosome
effectively deal with multiobjective optimization problems and to the random location, followed by the chromosome segment
multimodal functions.[48] The capabilities of GAs for optimiza- from parent 2 spanning the random location to the end of the
tion problems involving electromagnetic waves have been clearly chromosome. The chromosome of child 2 has a similar struc-
demonstrated in several works.[10,48–51] More recently, GAs have ture, except that it starts with a chromosome segment from
been applied to metasurface design.[11,12,14] parent 2 followed by one from parent 1. The the probability of
Here, we summarize the mains steps of GA and review some crossover should lie between 0.6 and 0.8).[48] The next step is
recent works on the development and application of GAs for de- to apply a mutation operator, in which each gene in the chro-
signing efficient and practical metasurface devices. A GA algo- mosome of an offspring is randomly changed. In case of a bi-
rithm can be summarized as follows: nary gene representation, each “1” becomes “0” and each “0”
becomes “1” upon mutation. This mutation operation should
• Initial generation. A GA starts by creating a number of random occur at low probability, between 0.01 and 0.08.[48] Figure 3,
candidate solutions (e.g., different metasurface designs). Each represents an example for the crossover mechanism using a
design (solution) is characterized by a chromosome that com- binary representation of the parameters in the chromosomes.
prises the optimization parameters, such as the device width, • Now the children replace the parents and one has the same
height, period of the gratings, etc. Each parameter in the chro- number of chromosomes (devices) as in the previous genera-
mosome is coded by genes. Very often, this coding is binary tion. Then, one evaluates the fitness function for each chromo-
value-based (0 or 1), but alternative approaches are possible some in the new generation using an electromagnetic solver.
as well. Each chromosome in the initial generation is associ- The GA can then continue by selecting the survivors in the
ated with a value of a so-called fitness function (for example, new generation. The termination of the algorithm can depend
the transmission or reflection coefficient). Then, the chromo- either on an appropriate convergence threshold or a number
somes are ranked according to their fitness function value. of iterations (generations).
• Selection. After ranking each of the chromosomes in the cur-
rent generation according to their fitness function values, one General considerations to implementing the algorithm in-
needs to select the most promising chromosomes to be used clude the following:
for producing the next generation (i.e., survival of the fittest).
For instance, one can decide that only 50% of the chromo- • The representation of each parameter in the chromosome. To
somes that are closer to the target fitness value are kept while operate effectively, GAs require the use of properly defined
the rest of the chromosomes are discarded. The selected chro- coding schemes that map metasurface parameters to genes.[48]
mosomes are considered as parents and are used to obtain the Note that with respect to binarization of the dielectric values
next generation of devices. The next step consists in mating the obtained after TO, as discussed previously, the purpose of us-
parents to generate new children. Different mating techniques ing binary encoding in GA algorithm is to assign the physical
can be applied, for instance, a mating between devices ordered information to a sequence of bits in the chromosome, and do
within the ranking list or random mating between devices. not readily correspond to the value of the material parameters.
• Crossover and mutation. The objective of these two operators The most common method is to use a binary representation in
is to generate two children from two parents. The question is: which each of the parameter values is represented by a binary
how do we generate the children, which correspond to new de- analog, such that a string of bits represents each parameter.
signs? One method is the basic one-point crossover operator, The evaluation of the cost function, which is determined by

Laser Photonics Rev. 2020, 1900445 1900445 (5 of 17) © 2020 Wiley-VCH GmbH
www.advancedsciencenews.com www.lpr-journal.org

Figure 3. Illustration of the crossover mechanism in a GA for two chromosomes with a binary representation of the parameters (genes).

performing a fullwave electromagnetic simulation, requires a enhance the transmission and increase the deflection angle for
protocol in which the binary representation of the parameters all-dielectric metasurfaces made of Si nanodisks by optimizing
is converted into real physical values. According to Ref. [48], two parameters: the radius and thickness of the nanodisk[12]
one can consider the following decoding method for any pa- (see Figure 4a). The authors show theoretically that transmis-
rameter p in the optimization problem: sion efficiencies up to 87.2% are obtained in visible spectrum
(580 nm) and up to 82% at the telecommunication wavelength
( )
pmax − pmin ∑ n
N−1 (1550 nm).
p= 2 bn + pmin , (4) Another application of GAs is discussed in Ref. [14] for the
2N − 1 n=0
design of a highly efficient beam deflector in the visible regime
using an extended unit cell approach. The building block cell
pmax and pmin represent the upper and lower limits of p, re- is made of elliptical nanoantennas, and the optimization is per-
spectively, and represent boundaries to parameters (e.g., fea- formed for five parameters: the minimum and maximum radii
ture size, refractive index, etc.) describing the physical device. of the ellipses, the x and y position for the center of the el-
N is the number of bits in the binary representation of the pa- lipses, and the orientation angle of the major axes (see Figure 4b).
rameter p and bn is the n − 1th bit. The role of Eq. (4) is to scale The authors have compared their GA variant with another bio-
the binary representation of the parameter p to a real value in inspired algorithm called the Artificial Bee Colony (ABC), which
a manner that takes into account its lower and upper limits.[48] is a global optimization technique based on the concept of swarm
• Number of chromosomes. It is also important to have a suf- intelligence.[14] They conclude that the two methods provide
ficient number of chromosomes at each generation to ensure nearly the same efficiency; however, the ABC method converges
sufficient diversity in the gene pool. It has been shown recently faster to the global optimal solution.
that more than 100 bits can be used, as discussed in refs. [15, Recently, another class of metasurfaces, referred to as binary
17] metasurfaces, has been introduced. Binary metasurfaces are
• Generating the random list of bits. During the initial genera- described in Ref. [59] and are based on a binary coding of the con-
tion of parent chromosomes, the bits in the chromosomes are stituent meta-atoms (see Figure 4c,d). As systems that explicitly
randomly generated. In many cases, a Poisson distribution is utilize a binary coding scheme, they are naturally amenable to
used for this task.[10] GA design methods. These concepts have been applied generally
• Convergence and probability of mutation. Mutation is impor- to the tailoring of scattering patterns and specifically to the reduc-
tant to explore the parameter space and avoid local maxima. tion of device radar cross sections. With this class of devices, the
It has been shown that mutating 1% of the chromosome bits optical response is specified by a meta-atom coding sequence.[59]
at each GA iteration is a reasonable choice of mutation rate For example, a 1-bit coding describes a sequence of binary coding
that enables sufficient exploration of the design space.[48] In particles that utilize elements represented by “0” and “1.” The
order to terminate the algorithm, one needs to either specify “0” and “1” represent meta-atoms with “0” and “𝜋” phase shift
the number of total generations or specify a threshold value for responses, respectively. A 2-bit binary representation can also
the fitness function. For example, for a transmissive metasur- be considered: “00,” “01,” “10,” and “11” represent meta-atoms
face, one can set a target threshold value for the transmission with “0,” “𝜋∕4,” “𝜋,” and “3𝜋∕4” phase shifts, respectively.[58,59]
coefficient at a given frequency. As the behaviour of a GA algo- The material choice and dimensions of each meta-atom are indi-
rithm is stochastic, it is necessary to run the algorithm several vidually optimized prior to GA optimization of the array coding
times to confirm its convergence to a global maximum. For sequence,[58,59] as shown in Figure 4c,d. In this case, this ap-
cases where different solutions are obtained for different opti- proach does not properly account for near-field coupling between
mization runs, decreasing the number of mutations or adding meta-atoms, which limits the overall device performance.
some physical constraints based on the problem at hand can GAs have also been used to enhance light-matter interactions
improve the convergence.[10] such as magnetic effects in the visible regime. In Ref. [60],
the authors identified the optimal configuration of a binary
More resources on the implementation of GAs to electromag- configuration, made of Si and air, which maximizes the mag-
netics problems are discussed in refs. [10, 48, 49, 52]. netic field intensity. The geometry and its binary representation
Figure 4 highlights important works that have used GAs are shown in Figure 5a, in which each “1” represents Si
for the design of metasurfaces. The GA has been applied to rectangular meta-atoms and each “0” represents an air void. As

Laser Photonics Rev. 2020, 1900445 1900445 (6 of 17) © 2020 Wiley-VCH GmbH
www.advancedsciencenews.com www.lpr-journal.org

Figure 4. Examples of optimized metasurface devices using a GA or its adaptive form. The readers can refer to refs. [21–23] for a broad overview
of optimized configurations based on GA or, more generally, on advanced evolutionary strategies as in refs. [53–57]. a) An all-dielectric metasurface
configuration made of Si nanodisks. The parameters to be optimized were the radii and the thicknesses of the nanodisks. The authors demonstrated an
optimized beam deflector metasurface with nearly 82% at the telecommunication wavelength. b) A 3D sketch of a beam deflector metasurface made of
TiO2 elliptical nanoantennas. (a,b) Reproduced with permission.[35] Copyright 2017, Optical Society of America. The parameters to be optimized were
the minimum radius, maximum radius, x and y coordinates of the ellipse in the unit-cell, and the orientation angle of the major axes. The authors showed
that a 60% efficiency can be obtained for beam steering with an angle as large as 50 degree at 𝜆 = 520 nm). c,d) example of an optimized metasurface
geometry with its corresponding binary representation. (c,d) Reproduced with permission.[58] Copyright 2018, Optical Society of America. e) A plasmonic
metasurface comprising a backside Au mirror (yellow bottom region) and Si spacer (blue region) and a binary gold pattern (top yellow regions). This
geometry has been used to optimize a beam reflector metasurface with nearly 92% of performance efficiency. Reproduced with permission.[11] Copyright
2019, Springer Nature. f) A reconfigurable metasurface pattern designed using a GA in order to switch from highly transmissive mode with efficiency
(80%) to highly absorptive modes with efficiency as high as (76%). Reproduced with permission.[13] Copyright 2017, Optical Society of America.

an initial step, a random population consisting of 20 geometries It has been revealed that the optimized geometry exceeds the
is considered, each simulated using an electromagnetic solver state of the art reference plasmonic geometry by more than
to compute the magnetic intensity enhancement at the center a factor two (see Figure 5d–f). For more details about the im-
of the geometry (see red point in Figure 5a). The five best plementation of the inverse design, the readers can refer to
geometries, providing the highest magnetic enhancement, are Ref. [15].
kept and used to generate the new population. This process Various devices have been recently optimized using GA
is repeated until an optimized geometry is obtained. For a or using some advanced evolutionary strategies as shown
GA run of nearly 350 generations, with each generation cod- in refs. [53–56, 61]. To highlight the basic concepts of the
ing 20 different geometries, 7000 independent simulations numerical optimization methods and mention briefly some
are performed using an electromagnetic solver. The authors possible applications, the readers can refer to recent re-
conclude that the magnetic power density obtained using the view works about metasurfaces and nanophotonics (refs. [21–
optimized geometry (see Figure 5) can be increased by a factor 23]), discussing in details different applications relying on
of five, compared to state-of-the-art dielectric nanoantennas. In inverse design.
similar circumstances, the authors in Ref. [15], employed GA to So far, our discussion has focused on the optimization of only
improve the near field intensity in a plasmonic configuration. a single target in the objective function. A more challenging

Laser Photonics Rev. 2020, 1900445 1900445 (7 of 17) © 2020 Wiley-VCH GmbH
www.advancedsciencenews.com www.lpr-journal.org

Figure 5. a) Binary matrix representation of a metasurface geometry that enhances the magnetic field intensity at the center of the structure (red dot in
right and bottom figure). Each “1” represents a rectangular nanostructure made of Si while each “0” corresponds to an air void. b) Magnetic intensity
enhancement for the distribution of parent devices as a function of generation number. Each generation contains 20 individual devices. c) The magnetic
field distribution of the optimized geometry after 248 generations. (a–c) Reproduced with permission.[60] Copyright 2019, Wiley–VCH. d–f): another
example to illustrate the usefulness of utilizing evolutionary approach based technique in enhancing near field intensity. d) refers to the reference slot
gold antenna configuration (top), and the bottom shows the optimized geometry. The small point refers to where the field enhancement is calculated.
e): represents the comparison between the reference geometry (black curve) and the optimized one (red curve) as indicated by the field maps in (f).
(d–f) Reproduced with permission.[15] Copyright 2012, American Physical Society.

question is: how can we deal with multiple targets? such as mul- For the GA approaches presented here, all require fullwave cal-
tiple optical functionalities or optimization at different operat- culations of Maxwell’s equations for all devices in each genera-
ing wavelengths? In the following, we present a quick survey of tion, which adds up to at least a few thousand of simulations.
the attempts to include multiobjective optimization in the frame- Therefore, this technique is computationally expensive and has
work of GAs for metasurface designs. We refer to Ref. [62] for to be used with efficient fullwave solvers, especially when consid-
more information about multiobjective optimization for evolu- ering the optimization of 3D structures. We note that the previ-
tionary strategies. ously mentioned ABC evolutionary algorithm has been used to
The most straightforward way is to combine the targets into optimize metasurface designs, and when compared with a clas-
a single objective function and weigh each target with a normal- sical GA,[14] the ABC method is 35% faster while producing com-
ization term based on the prioritization of the target. However, parable results to GA.
according to Ref. [11], this is sub-optimal. A more effective
method to perform this multi-objective optimization is to con-
sider an adaptive form of GAs, which has been applied to design 3.2. Particle Swarm Optimization for Metasurfaces
binary metasurfaces with two targets in the objective function
(see Figure 4e). The main idea behind the adaptive approach is The Particle Swarm Optimization (PSO) algorithm is an iterative
to initially perform the GA that considers only one target in the global optimization technique in which the population (swarm)
objective function. After this first round of optimization is done, consists of a predefined number of small particles, each of which
the objective function is modified to account for the second target are coordinates in the search space.[1] These particles try to im-
and the GA is applied again to the whole problem until it reaches prove their location in the search space by remembering their
a satisfactory solution for both targets simultaneously. This adap- best location and sharing this information with the other mem-
tive GA is used to theoretically optimize different metasurface bers of the population. The PSO has been used to optimize differ-
devices, including binary pattern reflect-arrays and dual beam ent photonic devices including diffraction grating structures,[67]
aperiodic leaky wave antennas.[11] photonic crystal waveguides,[68] and to optimize metal nanopar-
There have been other attempts to apply GAs to the multi- ticles to obtain broadband plasmonic field enhancement over the
objective optimization of metasurface devices, including a re- entire visible regime.[69]
configurable metasurface device that has been demonstrated In Ref. [70], the PSO algorithm was coupled to an FDTD solver
experimentally[13] (see Figure 4f). The main goal was to achieve a to realize metasurfaces consisting of etched features within ex-
tunable metasurface configuration, using a specific composition tended slab waveguides. This metasurface architecture, which
of the Chalcogeneide glass, which changes its response from be- is also discussed theoretically in Ref. [14], is complementary
ing highly transparent to being highly absorptive at 𝜆 = 1.55 μm to more traditional metasurface layouts based on structurally-
as a function of temperature. The optimization of the device is isolated nanostructures. With PSO, the radii of ten nanoholes and
done using an adaptive GA together with a full-wave electromag- their relative positions are initially optimized to maximize light
netic solver based on the periodic finite element boundary inte- deflection at the wavelength 𝜆 = 4.2 𝜇m, using a predefined num-
gral method.[63] Another study aiming at optimizing the design ber of iterations. At the end of this step, the best device within the
of colour pixels based on Si nanostructures used an evolutionary population is identified and further locally optimized using a gra-
algorithm coupled to a frequency-domain Maxwell solver to treat dient descent-based technique (see Figure 6). This work showed
a multi-objective function.[64] Recently, some advanced evolution- for the first time the connection of the high forward scattering ef-
ary strategies have also been extended to design multifunctional ficiency of a cell with the well-known Kerker conditions that exist
metasurfaces.[53,65,66] for isolated scatterers (see also Ref. [71]).

Laser Photonics Rev. 2020, 1900445 1900445 (8 of 17) © 2020 Wiley-VCH GmbH
www.advancedsciencenews.com www.lpr-journal.org

Figure 6. a) Beam deflector Si metasurface made of spherical nanoholes. b) Top view of the structure, in which the supercell is delineated by the black
rectangle. the parameter d in the figure represents center-to-center distance. c,d) Normalized far field radiation pattern before and after optimization.
(a-d) Reproduced with permission.[70] Copyright 2017, Optical Society of America. e) A polarization beam splitter based on metasurface-assisted silicon
nitride Y-junction. The input port is I0 , TM0 can be converted to TE0 through the path I0 − O1 similarly, TE0 can be converted to TM0 through the path
I0 − O2 . The metasurface (red part) needs to be optimized to enhance the efficiency of the device. f,g): represent the simulated TM0 and TE0 injected
modes and their output responses, respectively. (e–g) Reproduced with permission.[72] Copyright 2019, Elsevier.

PSO has also been applied to optimize a polarization beam strategies operate with fixed parameters during the optimization
splitter based on metasurface-assisted silicon nitride Y-junction process.[10–12,52]
for mid-infrared wavelengths[72] (see Figure 6e. The main objec- The Covariance Matrix Adaptation Evolution Strategy
tive of this device is to convert the fundamental TM0 mode to the (CMA-ES) is an alternative evolutionary strategy that is ca-
TE0 and vice versa with high efficiency. In order to maximize the pable of adapting its internal optimization parameters during
mode conversion efficiency, the Y-junction is patterned with sil- the optimization process. CMA-ES is population-based and
icon metasurfaces (see red parts in Figure 6e). The optimization operates by iteratively evolving a Gaussian distribution of design
results are validated using numerical simulations, as indicated in candidates within the search space in order to find the global
Figure 6f,g. maxima.[1,73,74] This Gaussian distribution is fully defined by its
mean and covariance matrix, the latter describing the shape of
the distribution. This advanced evolutionary strategy uses several
automatically adjustable parameters that allow the covariance
3.3. Covariance Matrix Adaptation Evolution Strategy (CMA-ES) matrix to adapt to the local characteristics of the function to be
optimized. Starting from an initial random guess, the algorithm
With the evolutionary optimization strategies presented thus searches for the global maxima by reshaping and resizing its
far, we have seen that the internal optimization parameters, Gaussian sampling automatically every few iterations.
such as the number of generations and mutation protocol, Recently, CMA-ES has been applied to the design of phase
must be carefully chosen. The convergence of these methods gradient metasurfaces operating at the visible light regime.[75]
can be accelerated by tuning these internal parameters, but The CMA-ES algorithm has also been used to optimize several
this task is usually tedious in practice and is computationally metasurface devices such as infrared broadband quarter-wave
costly. While it is possible to automatically adjust these internal plate metasurfaces,[76] metasurface absorbers,[77] and apochro-
parameters during optimization, most classical evolutionary matic singlets metasurface-augmented GRIN lenses.[78]

Laser Photonics Rev. 2020, 1900445 1900445 (9 of 17) © 2020 Wiley-VCH GmbH
www.advancedsciencenews.com www.lpr-journal.org

Figure 7. Neural network architectures for the inverse design of nanophotonics systems. a) deep, fully connected neural networks as a surrogate electro-
magnetic simulator. Reproduced with permission.[83] Copyright 2018, American Association for the Advancement of Science. b) optimizer that combines
neural network simulators with particle swarm algorithms. Reproduced with permission.[87] Copyright 2019, Wiley-VCH. c) tandem neural network com-
bining an inverse and forward deep network. Reproduced with permission.[84] Copyright 2018, American Chemical Society. d) freeform nanostructure
inverse design using generative adversarial networks. Reproduced with permission.[88] Copyright 2019, American Chemical Society. e) global topology
optimization networks (GLOnets) are global, population-based optimizers that train using simulations from the adjoint variables method. Reproduced
with permission.[89] Copyright 2019, American Chemical Society.

Algorithm 1 Simple illustration for the CMA-ES steps For electromagnetic design problems described by a small
number of geometric parameters, discriminative networks can
1. Start with an initial guess (mean, and shape of the distribution) accurately map the explicit relationship between a geometry and
– m: mean of the distribution its electromagnetic response. Discriminative networks are super-
– C (matrix): shape of distribution (C = I is the identity matrix for the initial vised learning algorithms that learn from a training set, and the
guess) learning process can be mathematically described as mini batch
2. Generate population: xi = m + Ni (𝜎 2 , C) gradient descent. In basic form, these algorithms are deep neural
3. Evaluate the objective function using the electromagnetic solver for all the networks[83,84] that comprise multiple layers of interconnected
individuals of the population nodes, called neurons, which perform non-linear mathematical
4. Sort the generation (choose the most fitted individuals) operations on a set of weighted inputs to produce an output
5. Update mean, variance, and the covariance matrix value (Figure 7a). More sophisticated algorithms utilize convolu-
6. Repeat steps (2 − 5) until convergence is achieved
tional neural networks,[85] in which convolution operations are
performed by neurons. The weights and convolutional kernels
are learnable parameters that are determined from the network
training process. Given the non-linear responses of individual
neurons and their ensembles, the highly non-linear relationship
4. Inverse Designs Using Artificial Neural between geometry and response can be properly captured.
Networks To date, discriminative networks have been demonstrated to
accurately model a wide range of nanoscale electromagnetic sys-
The machine learning revolution has transformed the way large tems, including the scattering and chiral properties of plasmonic
datasets are handled and processed in all fields of technology structures,[83,85] silicon photonic devices,[86] and metasurfaces.[19]
and science. The electromagnetic device geometries and their Trained discriminative networks can be used to optimize
responses can be treated as large datasets, making the use of electromagnetic systems in a variety of ways. One way is to
machine learning highly applicable and relevant. The last two treat the discriminative network as a high speed electromagnetic
years have witnessed rapid growth in applying deep learning in solver and embed it into classical iterative optimization schemes,
the field of nanophotonics.[24,25,79–82] However, in this section, we such as genetic[90] and particle swarm[87] algorithms (Figure 7b).
will examine how machine learning can aid in the inverse design Compared to conventional electromagnetic solvers, a trained
and optimization of metasurface structures. We will focus on two discriminative network can model the electromagnetic response
types of deep learning architectures, discriminative and genera- of a system with order-of-magnitude faster times. Another way
tive networks. is to directly optimize the electromagnetic device using iterative

Laser Photonics Rev. 2020, 1900445 1900445 (10 of 17) © 2020 Wiley-VCH GmbH
www.advancedsciencenews.com www.lpr-journal.org

backpropagation, which performs optimization using gradient accurately differentiating between the generated devices and
descent.[83] The idea is to first start with a random geometric those from the training set. Upon the completion of GAN
input, calculate its optical response with the neural network, and training, the generator will be able to generate devices that
iteratively backpropagate the difference between the actual and match the distribution of training set devices. GANs have been
desired optical response. During backpropagation, the network used to realize freeform geometries with tailored reflection
weights are fixed and the input geometry is modified to reduce and transmission spectra.[88,92] They can also learn from im-
the optical response error. In a third approach, discriminative ages of topology-optimized dielectric metasurfaces to generate
networks can be configured to directly solve the inverse problem topologically complex devices with high performance.[93]
without requiring an iterative process by specifying the inputs of Generative networks can also be trained directly, without a
the network to be the desired optical response and the outputs to training set, using calculations based on the adjoint variables
be the device layout. To help ensuring the stability and accuracy method. These global topology optimization networks, termed
of this inverse network, tandem architectures that combine GLOnets, use the training of a generative neural network to per-
an inverse network with a forward solver network (Figure 7c) form global topology optimization.[89,94] The concept is outlined
are effective at producing properly trained networks.[84] In this in Figure 7e for metagratings as a model system. In one itera-
algorithm, the forward solver network is first trained, using tion of the optimization process, the generator produces a batch
supervised learning with a training set, to create a high speed of devices, from which the efficiencies are calculated using an
surrogate solver. An inverse network is then attached to the electromagnetic solver and the efficiency gradients are evaluated
trained forward solver network, which has fixed weights, and using the adjoint variables method. These efficiencies and effi-
the inverse network is trained in the framework of the tandem ciency gradients are then used to update the network weights
network. Tandem networks have been used to design wavelength through backpropagation. Upon training completion, the genera-
filters[84] and topological photonic devices.[91] tive network maps the desired optical parameters and latent noise
The main drawback of discriminative networks is the extreme vectors to an ensemble of high performance devices. Bench-
amount of required training data. Empirically, for electromag- marking of GLOnets with iterative-only topology optimizers for
netic problem settings that are described by ten geometric pa- metagratings indicates that GLOnets can generate ultra-high effi-
rameters, approximately 10 000 system configurations need to ciency devices, with efficiencies higher than those produced from
be considered as inputs for the network to accurately converge. many instances of iterative-only topology optimization. While it
As a typical example, a metagrating described by sixteen geo- is not possible to determine whether these devices are globally
metric parameters requires 90 000 devices in the training set for optimal, due to the non-convexity of our optimization problem,
proper network training.[19] As such, even though a trained neu- they clearly have exceptional performance metrics. We antici-
ral network can serve as a fast electromagnetic solver, it requires a pate that hybrid machine learning concepts, which properly in-
very computationally costly off-line training phase. Furthermore, corporate physical knowledge into neural networks through the
strategies based on such networks cannot practically scale to com- use of physics-based calculations and that can even directly solve
plex electromagnetic geometries due to the curse of dimension- physics-based differential equations,[95] will play a large role in
ality. This concept, which is a long-standing problem in machine the future of electromagnetics inverse design.
learning, states that the design space and training set scale ex-
ponentially with the complexity of the system being modeled. As
such, discriminative network approaches cannot practically ap- 5. Bayesian Optimization
plied to structures described by freeform geometric layouts.
Generative neural networks are an alternative type of network A central benchmarking standard for any inverse design tech-
architecture that can be used in the design of complex electro- nique is the computational cost. For most photonics problems,
magnetic devices described by tens to hundreds of geometric these techniques require rigorous and expensive electromagnetic
parameters. Typically, these networks have desired optical pa- solvers to accurately compute the objective function at each op-
rameters as inputs and high resolution images of the devices as timization step. The computational cost of the electromagnetic
outputs. A key feature of generative networks is that, in addition solver can be mitigated by using an Artificial Neural Network
to the desired optical parameters, a high dimensional latent ran- (ANNs) as a surrogate solver, but a significant number of elec-
dom variable is also used as an input to the network. As such, for tromagnetic simulations still must be performed to train the
a given desired optical parameter, a wide range of devices can be network prior to optimization.[19,84] In this section, we discuss
generated, each of them mapping onto a unique latent random an alternative optimization strategy based on surrogate mod-
variable value. This mapping of an optical parameter to a distri- eling, named Efficient Global Optimization (EGO),[96,97] which
bution of devices in the design space is fundamentally different has been introduced recently in the context of the design of
from the mechanics of discriminative networks, and allows for nanophotonic devices.[98] The EGO algorithm is a global opti-
very high dimensional structures to be modeled and generated. mization algorithm that substitutes the complex and costly it-
There are various methods to train generative networks. One erative electromagnetic evaluation process with a simpler and
way is to use a generative adversarial network (GAN), in which cheaper model.[96,97] Its aim is to maximize a specific statistical
a generative neural network is trained against a discriminative criterion related to the optimization target, which is referred to
neural network using a training set (Figure 7d). During training, as the merit function. EGO involves two main phases. The first
the generative network generates devices and feeds them into one is called the Design Of Experiment (DOE), in which an ini-
the discriminator network. Its objective is to fool the discrimi- tial database is generated using a sampling of photonic devices
nator network. The discriminator is a classifier with the goal of within the design space. These devices are simulated using an

Laser Photonics Rev. 2020, 1900445 1900445 (11 of 17) © 2020 Wiley-VCH GmbH
www.advancedsciencenews.com www.lpr-journal.org

Figure 8. Illustration of the EGO algorithm using a simple 1D example. The black curve is the exact analytical function to be minimized. a) The initial
step. The blue points are the DOE database elements, and the orange curve is the first surrogate model that fits the blue points. b) Similar to (a), except
that the merit function is computed for each point on the orange curve (each parameter value). It is represented by the green curve and the right label.
As can be seen, at x ≈ 0.62, the merit function is maximized (see red dashed line). It means that this value of x corresponds to the highest probability
to improve the results, i.e obtain a better minimum. c) The objective function is computed at x ≈ 0.62 using the solver and the new value is added to
the database (red point). A new model is constructed (see orange curve in (c)). d) One computes the merit function based on this new model and all
the steps are repeated until the value of the merit function approaches zero, shown here. In this example, the surrogate model (orange curve) coincides
with the exact analytical function (black curve). For this simple example, only 20 iterations are required for the system to converge.

electromagnetic solver to compute their corresponding objective times the number of parameters, while the increased number of
function values, which are stored in the database. In the second iterations requires costly simulations for the metamodel adapta-
phase, a Gaussian Process (GP) model is constructed by inter- tion after each iteration. Moreover, the model training becomes
polating the database points. Internal model parameters are cal- computationally expensive when too many points are fitted.
ibrated according to a maximum likelihood principle.[99] Once Recently in Ref. [75], we applied the EGO algorithm to the
this GP model is defined, one can obtain an estimation of the optimization of metasurfaces. The objective was to maximize
objective function at any point of the design space, which rep- the light deflection efficiency at 𝜆 = 600 nm using metasurface
resents the model mean, and an estimate of the prediction un- designs based on rectangular and spherical nanopillars. Opti-
certainty, which is the model variance. These quantities are then mizing up to eight parameters describing arrays of cylindrical
used to define a statistical merit function, called the Expected Im- nanopillars, one can obtain more than 85% efficiency for the
provement (EI), whose maximum corresponds to the next design deflection of both TM and TE polarizations. In addition, in
parameters to be evaluated using an electromagnetic solver. After using rectangular-shaped antennas and optimizing twelve pa-
simulating this new point, this new data is added to the database. rameters, one obtains more than 88% efficiency for incident TM
This process in the second phase is repeated until convergence. polarized waves, as indicated in Figure 9. Moreover, in Ref. [75],
In Figure 8, we present a simple 1D example that illustrates the we have shown that several optima may exist for this problem,
basic mechanisms of the EGO algorithm. and that the use of EGO allows for the identification of all the
In Ref. [98], EGO was applied to optimize 3D nanoparticle physically relevant global optima related to the geometry under
shapes to design the morphology of metal nanoparticles. The consideration.
main target was to maximize the average electric fields on their This optimization approach, based on the iterative construc-
surfaces. The optimization is performed by changing the shapes tion of a database and an associated model, can be considered as
of the particles and the excitation wavelength. Several plasmonic a statistical learning strategy. The main feature of EGO is the use
materials were considered in the optimization, including gold of internal uncertainty estimation (i.e., variance) to drive both the
and silver. We would like to mention that a comparison between search for the optimum and the improvement of the model accu-
five benchmarking global optimization methods have been per- racy simultaneously. This concept is different from ANNs, which
formed in Ref. [100], in which it is shown that the Bayesian-based aim to construct an accurate model within the whole design space
optimization techniques require less simulation time compared prior to optimization, which is very expensive when the param-
to other techniques. The main limitation of the Bayesian opti- eter space is large. On the contrary, EGO focuses on the most
mization is the cost to construct the model when a large number promising areas of the design space. It is therefore far less ex-
of observations is included. It is usually related to the difficulty pensive in terms of electromagnetic solver calls, and in practice,
in handling a large parameter space. Indeed, this requires a only a few hundred electromagnetic simulations are typically re-
large database in the design of experiments phase, nearly ten quired for EGO. For instance, it was demonstrated in Ref. [75]

Laser Photonics Rev. 2020, 1900445 1900445 (12 of 17) © 2020 Wiley-VCH GmbH
www.advancedsciencenews.com www.lpr-journal.org

Figure 9. Optimization of rectangular-shaped nanoantennas using the EGO algorithm. a) The geometry under consideration consists of rectangular
nanoridges made of GaN (dark red) on top of a semi-infinite substrate made of Al2 O3 (green region). The twelve red circles represent the optimization
parameters. Under normal illumination, we aim at maximizing the diffraction efficiency for the first order mode 𝜂 TM (0, −1) at 𝜆 = 600 nm. b) Optimization
process using EGO as a function of the number of fullwave solver calls. The blue points represent the DOE (shaded region), the black points represent
the value of the objective function at each iteration, and the green solid line indicates the best performance obtained so far. c) Diffraction efficiency as a
function of the wavelength for the optimized design. We notice that it is maximal at the desired wavelength indicated by the black line. d) Field map at
𝜆 = 600 nm for the optimized design. (a–c) Reproduced with permission.[75] Copyright 2019, Springer Nature.

that one needs in total 150 solver calls for both the iterative en- ness criteria show that robust devices are relatively insensitive
richment and DOE (80 points) in order to optimize a structure to differing levels of over- and underexposure from the lithog-
with twelve parameters. raphy fabrication process. We note that there exists a trade off
between robustness and maximum device performance, which
does place practical limits on metasurface performance. It is
6. Robustness worth mentioning that, if to the best of our knowledge very few
papers are discussing this issue for metasurface designs, several
This burgeoning field of research in computational nanophoton- other works have been reported, applying sensitivity analysis in
ics applied to the design of metasurfaces is currently booming, optimizing various electromagnetic devices.[28,102,103]
offering new design perspectives. It is nevertheless important to An alternative methodology for the optimization of robust
point out that the uncertainties related to fabrication errors are metasurfaces leverages a concept termed Uncertainty Quantifi-
not considered for most of the optimization methods discussed cation (UQ).[104] In this reference, the authors optimize a 2D
so far, and that they play an important role to addressing errors (i.e., periodic only in one direction), high contrast, subwavelength
and uncertainties during the fabrication process. grating comprising gallium nitride (GaN) in order to maximize
In gradient-based topology optimization, geometric robust- the light deflection at a given fixed angle at visible regime. The
ness can be incorporated by considering the performance of results are summarized in Figure 10. With UQ, the authors op-
the eroded and dilated versions of the device throughout the timize the influence of the manufacturing process in order to
optimization process.[101] By incorporating the performance of obtain robust structures that are insensitive to small manufac-
these geometric variants into the objective cost function, the final turing imperfections.
devices become relatively insensitive to geometric perturbations. As a first step, the authors optimize several designs by comput-
Experimental characterization of devices designed with robust- ing the electromagnetic response of arrays of 2D subwavelength

Laser Photonics Rev. 2020, 1900445 1900445 (13 of 17) © 2020 Wiley-VCH GmbH
www.advancedsciencenews.com www.lpr-journal.org

Figure 10. a) Illustration of the 2D phase gradient metasurface made of GaN nano-resonators (orange ridges) placed over a substrate made of Al2 O3 .
The height is fixed to 1000 nm. The main goal is to optimize the thicknesses (𝛿Xi ) and the positions of the nanoridges Xi . b) UQ results for uniform
input distributions. (a,b) Reproduced with permission.[105] Copyright 2019, Optical Society of America.

ridges using the rigorous coupled wave analysis method, together techniques, CMA-ES tunes its internal parameters during the
with a gradient-free pattern search algorithm,[104] implemented optimization process, making it a suitable global optimization
in the Matlab optimization toolbox. In a second step, they per- strategy for complex problems with large parameter spaces.
form a UQ analysis and explore the impact of geometrical varia- Third, we discussed the utilization of artificial neural networks
tions in technologically relevant parameters such as ridge widths (ANNs) and deep learning architectures, namely discriminative
𝛿Xi and relative ridge positions Xi . A Monte Carlo ensemble of and generative networks, for the inverse design of metasurface
a million random device realizations are numerically simulated devices. For systems described with a small number of parame-
to evaluate the sensitivity of the optimal design with respect to ters, discriminative networks can accurately determine the rela-
manufacturing imperfections through statistical indicators, in- tionship between the optimization parameters and their electro-
cluding the mean deflection efficiency, standard deviation to the magnetic responses, thereby serving as computationally efficient
mean value, and confidence intervals. The results show that the surrogate electromagnetics solvers. However, this type of ANN
number of elements per phase gradient period plays a consider- requires a high computational overhead in order to effectively
able role in the reliability of the structures with respect to small train the network and it cannot be practically applied to complex
±5 nm uncertainties in the widths and positions.[105] problems. With generative neural networks, a high dimensional
latent random variable is used as an input to the network, mean-
ing that a wide range of devices can be generated and outputted.
7. Conclusion This mapping of a latent space to a distribution of devices allows
the amount of training data to be reduced and high dimensional
As a general conclusion of this work, we gave an up-to-date structures to be modeled and generated. This concept has been
overview of optimization techniques used in the field of meta- used recently in the framework of global topology optimization.
surface designs. We focused on the most general and common Finally, we discussed the concept of Bayesian Optimization as
inverse design techniques used in the literature. We hope that an alternative approach for the inverse design of metasurfaces.
the discussion provided herein will be helpful and useful even for More precisely, we focused on a widely used approach, which is
readers that are non-experienced in the field of inverse design. called Efficient Global Optimization (EGO). The EGO algorithm
First, we introduced gradient-based optimization techniques, is a global optimization algorithm based on a surrogate model,
including the objective-first algorithm and topology optimiza- and it replaces the costly evaluation process by a simpler and
tion. These algorithms can produce freeform geometrical shapes, computationally cheaper model. EGO uses internal statistical cri-
and can deal with a very large number of design variables in an teria in order to choose correctly the new evaluations that will en-
efficient manner. They can lead to sophisticated, complex, and rich the model to miminize/maximize of the objective function
non-intuitive designs, yet with high efficiencies. Nevertheless, and improve its accuracy as well. EGO uses trained data in its
there are still some challenges in fabricating the produced non- initial phase, however the number of trained data is far smaller
intuitive designs. In addition, these techniques depend strongly than the ones used in the ANNs framework. We also touch on the
on the initial design. importance of optimizing robust devices the implementation of
Second, we presented genetic algorithms (GA) and the co- Uncertainty Quantification to capture the sensitivity of manufac-
variance matrix adaptation evolution strategy (CMA-ES), which turing imperfections in the device design process. Among all of
are widely used global optimization techniques. These methods these optimization methods, a great deal of attention will soon
are iterative and based on a population of designs that represent be devoted to the simultaneous optimisation of multifunctional
the optimization parameters. In general, GAs and CMA-ESs metasurfaces, i.e. metasurface designs involving more than one
can deal with large parameter space at both the continuous and objective function, in particular to find trade-off designs able to
the discrete levels. However, they require expensive electromag- resolve the poor efficiency of broadband metasurfaces. Methods
netic solver calls, especially when dealing with large parameter such as multiobjective programming, multicriteria optimization,
spaces. CMA-ES has the particular advantage in that it is a self- multiattribute optimization, vector optimization or Pareto-front
adapted global optimization method. Unlike GA and other global optimization, which have already been applied to many fields of

Laser Photonics Rev. 2020, 1900445 1900445 (14 of 17) © 2020 Wiley-VCH GmbH
www.advancedsciencenews.com www.lpr-journal.org

science and engineering, would lead realistic metadevice designs [20] Z. Liu, D. Zhu, S. Rodrigues, K. T. Lee, W. Cai, Nano Lett. 2018.
with increased performance and capabilities. [21] K. Yao, R. Unni, Y. Zheng, Nanophotonics 2019, 8, 339.
[22] S. D. Campbell, D. Sell, R. P. Jenkins, E. B. Whiting, J. A. Fan, D. H.
Werner, Opt. Mater. Express 2019, 9, 1842.
Acknowledgements [23] S. Molesky, Z. Lin, A. Y. Piggott, W. Jin, J. Vuckovic, A. W. Rodriguez,
arXiv preprint arXiv:1801.06715 2018.
S.L., R.D., and P.G. acknowledge support from French defence pro- [24] K. Yao, R. Unni, Y. Zheng, Nanophotonics 2019, 8, 339.
curement agency under the ANR ASTRID Maturation program, grant [25] P. R. Wiecha, O. L. Muskens, Nano Lett. 2019.
agreement number ANR-18-ASMA-0006-01. P.G. acknowledges funding [26] L. Su, R. Trivedi, N. V. Sapra, A. Y. Piggott, D. Vercruysse, J. Vučkovíc,
from the European Research Council (ERC) under the European Union’s Opt. Express 2018, 26, 4023.
Horizon 2020 research and innovation programme (Grant agreement
[27] J. Lu, J. Vučković, Opt. Express 2013, 21, 13351.
no. 639109, 874986). J.F. is supported by the U.S. Air Force under
[28] A. Y. Piggott, J. Lu, K. G. Lagoudakis, J. Petykiewicz, T. M. Babinec,
Award Number FA9550-18-1-0070 and the David and Lucile Packard
J. Vučković, Nat. Photonics 2015, 9, 374.
Foundation.
[29] S. Boyd, S. P. Boyd, L. Vandenberghe, Convex optimization, Cam-
bridge university press, 2004.
[30] J. Lu, Nanophotonic Computational Design, PhD thesis, University of
Conflict of Interest Stanford, 2013.
[31] J. S. Jensen, O. Sigmund, Laser Photonics Rev. 2011, 5, 308.
The authors declare no conflict of interest.
[32] M. P. Bendsøe, O. Sigmund, in Topology Optimization, Springer,
2004, pp. 1–69.
[33] J. S. Jensen, O. Sigmund, Appl. Phys. Lett. 2004, 84, 2022.
Keywords [34] M. Burger, S. Osher, E. Yablonovitch, IEICE TRANSACTIONS ON
ELECTRONICS 2004, E87C, 258.
artificial intelligence, bayesian optimization, inverse design, metasurface, [35] J. Yang, J. A. Fan, Opt. Lett. 2017, 42, 3161.
optimization methods [36] C. Sitawarin, W. Jin, Z. Lin, A. W. Rodriguez, Photonics Res. 2018, 6,
B82.
Received: December 18, 2019 [37] D. Sell, J. Yang, S. Doshay, R. Yang, J. A. Fan, Nano Lett. 2017, 17,
Revised: June 30, 2020 3752.
Published online: [38] J. Yang, J. A. Fan, Opt. Express 2017, 25, 23899.
[39] J. Yang, D. Sell, J. A. Fan, Annalen der Physik 2018, 530, 1700302.
[40] D. Sell, J. Yang, E. W. Wang, T. Phan, S. Doshay, J. A. Fan, ACS Pho-
tonics 2018, 5, 2402.
[1] Numerical Methods for Metamaterial Design, Topics in Applied Physics [41] D. Sell, J. Yang, S. Doshay, J. A. Fan, Adv. Opt. Mater. 2017, 5,
(Ed: K. Diest), Vol. 127, Springer Netherlands, Dordrecht 2013. 1700645.
[2] P. Lalanne, S. Astilean, P. Chavel, E. Cambril, H. Launois, Opt. Lett. [42] T. Phan, D. Sell, E. W. Wang, S. Doshay, K. Edee, J. Yang, J. A. Fan,
1998, 23, 1081. Light: Science & Applications 2019, 8, 48.
[3] P. Lalanne, S. Astilean, P. Chavel, E. Cambril, H. Launois, JOSA A [43] D. Sell, J. Yang, S. Doshay, K. Zhang, J. A. Fan, ACS Photonics 2016,
1999, 16, 1143. 3, 1919.
[4] Z. Lin, B. Groever, F. Capasso, A. W. Rodriguez, M. Lončar, Phys. [44] Z. Lin, S. G. Johnson, Opt. Express 2019, 27, 32445.
Rev. Appl. 2018, 9, 044030. [45] R. Pestourie, C. Pérez-Arancibia, Z. Lin, W. Shin, F. Capasso, S. G.
[5] F. Callewaert, V. Velev, P. Kumar, A. V. Sahakian, K. Aydin, Scientific Johnson, Opt. Express 2018, 26, 33732.
Reports 2018, 8. [46] M. Schlüter, J. A. Egea, J. R. Banga, Computers & Operations Research
[6] F. Callewaert, S. Butun, Z. Li, K. Aydin, Scientific Reports 2016, 6. 2009, 36, 2217.
[7] A. Y. Piggott, J. Lu, T. M. Babinec, K. G. Lagoudakis, J. Petykiewicz, [47] R. Li, M. T. Emmerich, J. Eggermont, T. Bäck, M. Schütz, J. Dijkstra,
J. Vučkovíc, Scientific Reports 2015, 4. J. H. Reiber, Evolutionary computation 2013, 21, 29.
[8] D. Sell, J. Yang, S. Doshay, J. A. Fan, Adv. Opt. Mater. 2017, 5, [48] J. Johnson, V. Rahmat-Samii, IEEE Antennas and Propagation Maga-
1700645. zine 1997, 39, 7.
[9] J. Lu, S. Boyd, J. Vučkovíc, Opt. Express 2011, 19, 10563. [49] R. Haupt, IEEE Transactions on Antennas and Propagation 1994, 42,
[10] R. Haupt, IEEE Antennas and Propagation Magazine 1995, 37, 7. 993.
[11] S. Jafar-Zanjani, S. Inampudi, H. Mosallaei, Scientific Reports 2018, [50] C. Forestiere, A. J. Pasquale, A. Capretti, G. Miano, A. Tamburrino,
8, 11040. S. Y. Lee, B. M. Reinhard, L. Dal Negro, Nano Lett. 2012, 12, 2037.
[12] V. Egorov, M. Eitan, J. Scheuer, Opt. Express 2017, 25, 2583. [51] A. Mirzaei, A. E. Miroshnichenko, I. V. Shadrivov, Y. S. Kivshar, Appl.
[13] A. V. Pogrebnyakov, J. A. Bossard, J. P. Turpin, J. D. Musgraves, H. J. Phys. Lett. 2014, 105, 011109.
Shin, C. Rivero-Baleine, N. Podraza, K. A. Richardson, D. H. Werner, [52] R. L. Haupt, D. H. Werner, Genetic Algorithms in Electromagnetics,
T. S. Mayer, Opt. Mater. Express 2018, 8, 2264. John Wiley & Sons, 2007.
[14] K. D. Donda, R. S. Hegde, Prog. Electromagn. Res. M 2017, 60, 1. [53] D. Z. Zhu, E. B. Whiting, S. D. Campbell, D. B. Burckel, D. H. Werner,
[15] T. Feichtner, O. Selig, M. Kiunke, B. Hecht, Phys. Rev. Lett. 2012, 109, ACS Photonics 2019, 6, 2741.
127701. [54] C. Liu, S. A. Maier, G. Li, ACS Photonics 2020, 7, 1716.
[16] M. D. Huntington, L. J. Lauhon, T. W. Odom, Nano Lett. 2014, 14, [55] M. J. Wallace, S. T. Naimi, G. Jain, R. McKenna, F. Bello, J. F. Done-
7195. gan, Opt. Express 2020, 28, 8169.
[17] T. Feichtner, O. Selig, B. Hecht, Opt. Express 2017, 25, 10828. [56] Z. Li, L. Stan, D. A. Czaplewski, X. Yang, J. Gao, Opt. Lett. 2019, 44,
[18] P. R. Wiecha, C. Majorel, C. Girard, A. Cuche, V. Paillard, O. L. 114.
Muskens, A. Arbouet, Opt. Express 2019, 27, 29069. [57] Z. Jin, S. Mei, S. Chen, Y. Li, C. Zhang, Y. He, X. Yu, C. Yu, J. K. Yang,
[19] S. Inampudi, H. Mosallaei, Appl. Phys. Lett. 2018, 112, 241102. B. Luk’yanchuk, S. Xiao, C.-W. Qiu, ACS nano 2019, 13, 821.

Laser Photonics Rev. 2020, 1900445 1900445 (15 of 17) © 2020 Wiley-VCH GmbH
www.advancedsciencenews.com www.lpr-journal.org

[58] S. Sui, H. Ma, Y. Lv, J. Wang, Z. Li, J. Zhang, Z. Xu, S. Qu, Opt. Express [81] Z. A. Kudyshev, A. V. Kildishev, V. M. Shalaev, A. Boltasseva, arXiv
2018, 26, 1443. preprint arXiv:1910.12741 2019.
[59] T. J. Cui, S. Liu, L. Zhang, J Mater Chem C 2017, 5, 3644. [82] Y. Kiarashinejad, S. Abdollahramezani, A. Adibi, npj Comput. Mater.
[60] N. Bonod, S. Bidault, G. W. Burr, M. Mivelle, Adv. Opt. Mater. 2019, 2020, 6, 1.
7, 1900121. [83] J. Peurifoy, Y. Shen, L. Jing, Y. Yang, F. Cano-Renteria, B. G. DeLacy,
[61] H. Li, G. Wang, L. Zhu, X. Gao, H. Hou, Optics Communications p. J. D. Joannopoulos, M. Tegmark, M. Soljačić, Sci. Adv. 2018, 4,
124601 (2020). eaar4206.
[62] K. Deb, Multi-objective optimization using evolutionary algorithms, [84] D. Liu, Y. Tan, E. Khoram, Z. Yu, ACS Photonics 2018, 5, 1365.
John Wiley & Sons, 2001. [85] W. Ma, F. Cheng, Y. Liu, ACS Nano 2018, 12, 6326.
[63] T. F. Eibert, J. L. Volakis, D. R. Wilton, D. R. Jackson, IEEE Transactions [86] D. Gostimirovic, W. N. Ye, IEEE J. Sel. Top. Quantum Electron. 2019,
on Antennas and Propagation 1999, 47, 843. 25, 1.
[64] P. R. Wiecha, A. Arbouet, C. Girard, A. Lecestre, G. Larrieu, V. Pail- [87] Q. Zhang, C. Liu, X. Wan, L. Zhang, S. Liu, Y. Yang, T. J. Cui, Adv.
lard, Nature Nanotechnology 2017, 12, 163. Theory Simul. 2019, 2, 1800132.
[65] D. Tang, L. Chen, J. Liu, Opt. Express 2019, 27, 12308. [88] Z. Liu, D. Zhu, S. P. Rodrigues, K. T. Lee, W. Cai, Nano Lett. 2018,
[66] E. B. Whiting, S. D. Campbell, D. H. Werner, P. L. Werner, in 2019 18, 6570.
IEEE International Symposium on Antennas and Propagation and [89] J. Jiang, J. A. Fan, Nano Lett. 2019, 19, 5366.
USNC-URSI Radio Science Meeting, 2019, pp. 1815–1816. [90] Y. Liu, T. Lu, K. Wu, J. M. Jin, in 2018 IEEE 27th Conference on Electri-
[67] M. Shokooh-Saremi, R. Magnusson, Opt. Lett. 2007, 32, 894. cal Performance of Electronic Packaging and Systems (EPEPS), IEEE,
[68] S. M. Mirjalili, K. Abedi, S. Mirjalili, Optik 2013, 124, 5989. 2018, pp. 261–263.
[69] C. Forestiere, M. Donelli, G. F. Walsh, E. Zeni, G. Miano, L. Dal Ne- [91] Y. Long, J. Ren, Y. Li, H. Chen, Appl. Phys. Lett. 2019, 114, 181105.
gro, Opt. Lett. 2010, 35, 133. [92] S. So, J. Rho, Nanophotonics 2019, 8, 1255.
[70] J. R. Ong, H. S. Chu, V. H. Chen, A. Y. Zhu, P. Genevet, Opt. Lett. [93] J. Jiang, D. Sell, S. Hoyer, J. Hickey, J. Yang, J. A. Fan, ACS Nano 2019,
2017, 42, 2639. 8, 8872.
[71] Q. Yang, S. Kruk, Y. Xu, Q. Wang, Y. K. Srivastava, K. Koshelev, I. [94] J. Jiang, J. A. Fan, Nanophotonics 2019, Ahead of Print.
Kravchenko, R. Singh, J. Han, Y. Kivshar, I. Shadrivov, Adv. Funct. [95] M. Raissi, P. Perdikaris, G. E. Karniadakis, J. Comput. Phys. 2019,
Mater. 2020, 30, 1906851. 378, 686.
[72] B. Zhang, W. Chen, P. Wang, S. Dai, H. Li, H. Lu, J. Ding, J. Li, Y. Li, [96] D. Jones, J Global Optimization 1998, 13.
Q. Fu, T. Dai, Y. Wang, J. Yang, Opt. Commun. 2019, 451, 186. [97] D. Jones, J Global Optimization 2001, 21, 345.
[73] N. Hansen, A. Ostermeier, Evolutionary Computation 2001, 9. [98] C. Forestiere, Y. He, R. Wang, R. M. Kirby, L. Dal Negro, ACS Pho-
[74] N. Hansen, S. Muller, P. Koumoutsakos, Evolutionary Computation tonics 2015, 3, 68.
2003, 11, 1. [99] D. J. MacKay, Neural Computation 1991, 4.
[75] M. M. R. Elsawy, S. Lanteri, R. Duvigneau, G. Brière, M. S. Mo- [100] P. I. Schneider, X. Garcia Santiago, V. Soltwisch, M. Hammer-
hamed, P. Genevet, Scientific Reports 2019, 9. schmidt, S. Burger, C. Rockstuhl, ACS Photonics 2019, 6, 2726.
[76] P. E. Sieber, D. H. Werner, Opt. Express 2014, 22, 32371. [101] E. W. Wang, D. Sell, T. Phan, J. A. Fan, Opt. Mater. Express 2019, 9,
[77] I. Martinez, A. H. Panaretos, D. H. Werner, G. Oliveri, A. Massa, in 469.
2013 7th European Conference on Antennas and Propagation (EuCAP), [102] Z. Ren, M. T. Pham, C. S. Koh, IEEE Trans. Magn. 2012, 49, 851.
IEEE, 2013, pp. 1843–1847. [103] J. Jung, IEEE Photonics Technol. Lett. 2015, 28, 756.
[78] J. Nagar, S. Campbell, D. Werner, Optica 2018, 5, 99. [104] J. P. Hugonin, P. Lalanne, “Reticolo software for grating analysis,”
[79] S. So, T. Badloe, J. Noh, J. Rho, J. Bravo-Abad, Nanophotonics 2020. www.lp2n.institutoptique.fr. accessed: January 2014.
[80] Y. Kiarashinejad, M. Zandehshahvar, S. Abdollahramezani, O. Hem- [105] N. Schmitt, N. Georg, G. Brière, D. Loukrezis, S. Héron, S. Lanteri,
matyar, R. Pourabolghasem, A. Adibi, Advanced Intelligent Systems C. Klitis, M. Sorel, U. Römer, H. D. Gersem, S. Vézian, P. Genevet,
2020, 2, 1900132. Opt. Mater. Express 2019, 9, 892.

Mahmoud M. R. Elsawy received his Ph. D. degree at University of Aix-Marseille, France in 2017 with
specialization in optics, photonics, and image processing. The topic of the Ph. D. was related to mod-
eling and improvements of complex nonlinear plasmonic waveguides. He spent one year as a post-
doctoral researcher at Institut Fresnel, Marseille, France. His research activities were dedicated to the
field of nonlinear plasmonics and modeling of hollow-core negative curvature optical fibers. Since Oc-
tober 2018, he is working as a postdoctoral visitor researcher at Inria Sophia Antipolis. His research
focuses on numerical optimization of metasurfaces and complex nanophotonic devices.

Laser Photonics Rev. 2020, 1900445 1900445 (16 of 17) © 2020 Wiley-VCH GmbH
www.advancedsciencenews.com www.lpr-journal.org

Stéphane Lanteri received his Ph.D. degree in engineering sciences at University of Nice Sophia An-
tipolis, France, in 1991. After a postdoc in the Aerospace Engineering Department of the University
of Colorado at Boulder, he was appointed as a research scientist at Inria Sophia Antipolis, in 1993.He
is currently a senior research scientist and heading research activities on innovative mathematical
methods for the numerical modeling of nanoscale light-matter interactions, including high order fi-
nite element methods (Discontinuous Galerkin methods) and high performance multiscale solvers of
full-wave differential models. He also coordinates the development of the DIOGENeS software suite
(https://diogenes.inria.fr/) dedicated to computational nanophotonics.

Régis Duvigneau received his Ph. D. degree at Ecole Centrale de Nantes (France) in 2002, whose topic
concerned shape optimization algorithms in fluid mechanics. After a post-doc period at Ecole Poly-
technique in Montreal (Canada), he became permanent researcher at Inria Sophia-Antipolis (France)
in 2005. His research activity was then focused on numerical methods for design optimization for
systems governed by partial differential equations, with main applications in aerodynamics.

Jonathan A. Fan received his Ph. D. degree at Harvard University in 2010. After a postdoc at the Univer-
sity of Illinois at Urbana-Champaign, he became an assistant professor in the Department of Electrical
Engineering at Stanford University in 2014. His research activities have focused around the growth,
assembly, and computational design of new photonic materials.

Patrice Genevet received his Ph.D. degree at the université Côte d’Azur, France in 2009 on localized
spatial solitons in semiconductor lasers. He did five years as a research fellow (2009–2014) in the
Capasso group (Harvard University) in collaboration with Prof. Scully (Texas A&M University). In
2014, he worked as senior research scientist in A∗ STAR, Singapore, to join CNRS (France) in 2015. His
research activities concern the development of metamaterials, metasurfaces and their applications.

Laser Photonics Rev. 2020, 1900445 1900445 (17 of 17) © 2020 Wiley-VCH GmbH