A Generative Model for Inverse Design of

Metamaterials
Zhaocheng Liu1, Dayu Zhu1, Sean P. Rodrigues1,2, Kyu-Tae Lee1, and Wenshan Cai1,2*
1
School of Electrical and Computer Engineering, Georgia Institute of Technology, Atlanta, Georgia 30332
2
School of Materials Science and Engineering, Georgia Institute of Technology, Atlanta, Georgia 30332
* Correspondence should be addressed to W.C. ([email protected])

Abstract: The advent of two-dimensional metamaterials in recent years has ushered in a

revolutionary means to manipulate the behavior of light on the nanoscale. The effective
parameters of these architected materials render unprecedented control over the optical
properties of light, thereby eliciting previously unattainable applications in flat lenses,
holographic imaging, and emission control among others. The design of such structures, to
date, has relied on the expertise of an optical scientist to guide a progression of
electromagnetic simulations that iteratively solve Maxwell’s equations until a locally
optimized solution can be attained. In this work, we identify a solution to circumvent this
intuition-guided design by means of a deep learning architecture. When fed an input set of
optical spectra, the constructed generative network assimilates a candidate pattern from a
user-defined dataset of geometric structures in order to match the input spectra. The
generated metamaterial patterns demonstrate high fidelity, yielding equivalent optical
spectra at an average accuracy of about 0.9. This approach reveals an opportunity to
expedite the discovery and design of metasurfaces for tailored optical responses in a
systematic, inverse-design manner.

1. Introduction

Near the close of the last century, discoveries in light-matter interactions on the nanoscale
unlocked optical phenomena that would help to confine light to sub-wavelength scales, opening a
gateway to a new era of optical design. The metamaterial, a member of this family of new
nanophotonic devices, is capable of generating periodic dipoles in order to manipulate the behavior

1
of light in a non-classically predicted manner. As such, a well-designed metamaterial can tailor
1-5
the transmittance and phase delay of electromagnetic waves over any wavelength spectrum .
The realization of these materials has led to a vast number of applications in perfect absorption 6,
super resolution imaging 7,8, beam steering 9,10, and nonlinear optical generation 11,12.

As these nanostructured materials require labor intensive fabrication, an accurate prediction of

the optical spectrum and structure of the envisioned metamaterial must be preemptively articulated.
However, the complicated physical mechanisms that describe these light-matter interactions at the
nanoscale cannot be resolved by generalized theory and as such the prediction of a material’s
optical properties and approximate structure relies on advanced iterative calculations achieved by
Finite-Element Modeling (FEM) or Finite-Difference Time-Domain (FDTD) methods. Moreover,
this conventional metasurface design process is innately flawed by human guided error. Not only
is the initial design realized based on physical insights and intuitive reasoning, but the finalized
geometric and material parameters are ultimately achieved by a means of trial-and-error. The
design of such optical systems demands a working knowledge base of optics in order to moderate
iterative simulations that scan multi-dimensional parameter spaces. Thanks to rapid developments
in artificial intelligence (AI), some scientific problems that classically required human perception
or intricate mechanisms have recently been solved by AI 13-17. Such methods have translated into
18-20 21
the field of optics, by employing optimization methods and evolutionary algorithms to
expedite the design of photonic devices. To accelerate the design process without extensive
computation (numerical or analytical) of Maxwell’s equations, data driven methods, especially
deep neural networks, have been gradually incorporated into the design of microwave and
nanophotonics devices 22-25.

With growing interest of new phenomena and applications using metasurfaces, there is an ever-
pressing need to develop efficient methods that expedite the discovery and design of novel
metasurface structures with custom-defined functionality. As illustrated in Figure 1a, this work
aims to leverage deep neural networks to approximate the spectra of a metasurface, and more
importantly, to generate metasurface patterns that yield customer-defined spectra at the input. The
latter is the long-sought-after goal of inverse optical design, in which a working structure is to be
generated directly based on the desired optical responses of the designer. In doing so, the need for
extensive parameter scans or trial-and-error procedures is bypassed. However, due to the enormous

2
degrees of freedom in typical metasurface patterns, conventional neural network schema, like
25
backpropagation in trained simulator networks , are impotent in the inverse design of
metasurfaces. In addition, the trained simulator defines the problem as a deterministic system, such
that a single simulator will inevitably lead to a fixed outcome for a given input condition.
Meanwhile, multiple solutions may exist for the same target spectrum fed to the simulator at the
input, thereby imposing an unnecessary constraint on the diversity of the optimized structures.

To mitigate these challenges, here we adopt a generative adversarial network (GAN) in the
network model. A GAN is an unsupervised learning architecture consisting of two networks,
namely a generator and a critic, that contesting with each other to create authentic image and video
26-29
datasets . Instead of employing a single simulator to tackle the inverse design problem, we
incorporate a GAN to jointly seek the structures for the intended spectra at the input. We have also
constructed a geometric dataset that is comprised of images of random shapes. When training the
GAN-based network with the geometric data, the resultant patterns from the network will resemble
some samples in the geometric dataset. Since constructing such geometric data is largely effortless,
given a sufficiently large dataset of the geometries, we will be able to identify and refine proper
shapes of the structures in order to replicate the demanded spectra at the input of the networks.
Compared to existing practices of optical design with neural networks, which are only able to
optimize a few geometrical parameters of a fixed structure represented as vectors (i.e., thickness
distribution of layered systems), our network allows for the generation of essentially arbitrary
patterns of the unit cell structure, represented as pixelwise images. Moreover, the technique
developed here can process multiple input spectra in parallel, which facilitates efficient design and
optimization of more than one optical structures simultaneously. Such a need is commonly
encountered in metasurfaces research, where gradient meta-structures of varying unit cells are
distributed in a two-dimensional space for wavefront shaping applications such as metaholograms,
vortex generation, and planar lenses.

2. Network architecture

To realize the AI-based optical design described above, we have constructed a network
architecture as illustrated Fig. 1b. We divide the network into three parts: a simulator (S), a
generator (G) and a critic (D). The primary goal is to train an overfitted generator, which produces
metasurface patterns in response to given input spectra T such that the Euclidean distance between

3
the spectra of the generated pattern T’ and the input spectra T is minimized. All three networks are
convolutional neural networks with delicate differences in detailed structures. The simulator is a
pretrained model with fixed weights, taking the generated patterns as input and approximating
their transmittances spectra Tˆ without the use of electromagnetic simulations. It is built to control
the accuracy of the optical spectra of the generated structures when training the generator. The
generator and the critic together constitute a GAN. The critic of this GAN accepts both the user-
defined geometric data and the patterns generated from the generator, and then yields a value l
which is essential to compute the distance between the distributions of the two sets of data. By
minimizing this distance, the critic network guides the generator to produce patterns that share
commons features with the input geometric data. During the training process, we update the
weights in the generator by backpropagation from the losses defined by the simulator and the critic.
Valid patterns produced by the generator are documented throughout the training process; this
occurs whenever the losses of the simulator and critic are sufficiently small. The generated patterns
are finally smoothed to binary images as candidates of the metasurface design. Detailed network
configurations and training methods are presented in the Supplementary Information.

In essence, the critic learns the distribution of the geometric data and restricts the generator to
produce patterns in the image space where the geometric data resides in. There are several benefits
of including such a critic into the overall network architecture. First, it excludes a large number of
unacceptable patterns that are deemed unrealistic in actual nanofabrication. Second, it allows us to
control the overall shape of the generated patterns by feeding similar geometries into the critic.
While such forethoughts are not necessary for AI-guided optical design, as demonstrated in the
later part of this work, certain geometrical constraints at the input may help to narrow down the
potential candidates and thereby expedite the convergence to a solution. Finally, since the
production of various categories of geometric data is straightforward and largely effortless, we can
modify the distributions of the geometric data to avoid degeneracy. This may occur when multiple
metasurface patterns of different topologies possess identical optical spectra to the input spectra,
within an acceptable margin of deviation.

As a representative and generalizable case study, we will apply the strategy outlined above to
the design of metasurfaces with prescribed spectral behavior under linearly polarized illumination.
The general unit cell of the metasurface used in this model is shown in Figure 2a, which has a

4
single layered gold pattern in a square lattice, situated on a glass substrate. Other common
parameters include a lattice constant of w = 340 nm and a thickness of the gold layer set to d = 50
nm. To train the simulator with sufficient data, we carried out 6,500 full wave finite element
method (FEM) simulations for metasurfaces with a wide variety of shapes that replaced the metal
in the unit cells. The simulation was run over a frequency span from f = 170 THz to 600 THz (i.e.,
500 nm to 1.8 µm in the wavelength domain), which covers a major portion of the visible and the
near-infrared spectral range. These FEM simulations yield the transmittance magnitude spectra Tij

of each metasurface under x- and y-polarized illumination, where i and j indicates the polarization
directions for the incidence and the detection, respectively. Throughout the training, the unit cell
structure is represented as a binary image of 64 × 64 pixels, in which 1 stands for gold and −1 for
void (air). Each transmission spectrum Tij ( f ) is represented as a 32-entry vector with equal

frequency intervals. The simulator after the training process is able to approximate the
transmittance T with an average absolute error of less than 0.01 at each frequency point. An
example for the effectiveness of the simulator is illustrated in Fig. 2b, where the solid lines are the
results from the FEM electromagnetic simulations for the ellipse particle array shown in Fig. 2a.
Here, the circles represent the spectral approximation of the same structure obtained by the neural
network.

To demonstrate the mechanism and performance of the network, we classify the geometric data
into several classes such as circles, arcs, crosses, ellipses, rectangles, etc. A full list of the training
data and corresponding samples are made available in the Supplementary Information. Figure 2c
illustrates a series of generated patterns during a training process after certain iterations, with and
without the critic network. For this example, the critic is fed with a class of geometric data only
comprised of crosses. When the critic is on, a cross pattern emerges and gradually adjusts itself to
meet the requirement of the input spectrum. In sharp contrast, when the critic network is turned
off, the generated pattern is a cluster of random pixels and stays stabilized after a few hundred
iterations.

3. Results and analysis

Our network architecture is capable of generating metasurface patterns in response to the

arbitrarily-input, on-demand optical spectra, whose Tij components and frequency range of interest

5
are defined by the user. At the input of the generator, we specify a set of transmittance spectra Tij

from 170 to 600 THz, which corresponds to a wavelength range of 500 nm to 1.8 µm. In the
following discussion, we define a set of patterns to be used as a test set and denote this test set as
s, the spectra of each s as T. Once these spectra are passed through the network architecture, a
pattern is retrieved which we denote as s’ in correspondence to s. In verification, the generated set
s’ are FEM-simulated and defined as T’. Unless mentioned otherwise, in the following experiments
we set the number of patterns being parallelly searched at each run to be 40 and the total iterations
of the training to be 50,000. For each target spectrum, valid patterns may occur at different stages
of the training process whenever the losses of the simulator and critic are both reasonably small.
On a machine with a single GPU Quadro P5000, it takes approximately 10 minutes to carry out
10,000 iterations of training on average.

As an initial demonstration, to illustrate the overall competence of the proposed framework, we

use the spectra, T, of randomly selected test samples from each geometric class as the input, and
allow the network to seek proper patterns based on these spectra. In this situation, we ensure the
existence of solutions by using actual spectra of real patterns as the “target” or input. This
constraint will be removed when we perform the inverse design for on-demand spectra, as
presented in a later part of the paper. In each test, the critic is fed with 1000 data points (i.e.,
geometrical shapes) that are randomly generated from the same geometric class. Figure 3a shows
representative samples from such experiments for each class of the geometry. The first row depicts
the test samples s, and the second row shows the corresponding samples s’ generated by the
network. The geometric patterns in each pair of the two rows agree very well, partially because
the full spectra input Tij to the generator substantially narrows down the possible solutions. Since

the generator does not receive any direct information of the geometry for the input spectra, it may
uncover equivalent patterns s’ that are different from the test structures s while yielding the same
spectral behavior. Such examples can be found in the cross, sector and arc cases in Fig. 3a, where
the discovered patterns s’ are mirror-flipped counterparts of s with the same optical responses
under linearly polarized illuminations.

Figure 3b and 3c show the spectra of an ellipse test sample s and those of the discovered pattern
s’, with the corresponding unit cell of the metasurface placed as insets. Comparison between the
two sets allows us to conclude that the network has successfully identified the correct structure to

6
replicate the spectra with only minor deviations. We also note that the geometric data fed into the
critic network does not necessarily contain the right shape of the resultant solutions. If the type of
the right pattern s for the required spectra is contained in the geometric data at the input, the nature
of the GAN will lead to a decent chance of identifying the structure s as a proper candidate;
otherwise, the critic will guide the generator to produce patterns with geometric features similar to
those of the right geometry 26.

To illustrate the ability of our model to optimize structures of any shape, we further test our
network with handwritten digit dataset MNIST as the unit patterns of metasurfaces 30. We modify
patterns from the MNIST dataset with rotation and shifting, and then feed them into the critic
during the training. In this experiment, we intentionally exclude digit “5” in the geometric dataset
at the input of the critic, and ask the network to generate a pattern that replicates the spectral
features of a metasurface with digit “5”. Figure 3d and 3e show the spectra of the test sample s –
a rotated digit “5”, and the discovered pattern s’ – a modified digit “3”, respectively. The
topologies of the s and s’ differ considerably, but the overall transmittance behaviors of the two
samples, especially Txx and Tyy , possess similar features in terms of both the spectral location and

the amplitude. By removing the input structure from the trained data set, the network proves its
overarching utility in producing a set of spectra that can competently match that of the input
structure.

To quantify the performance of the network, we quantitatively define three types of accuracy as
follows. 1) Geometric accuracy: the portion of s’ that can be recognized as within the same class
of the input geometric data; 2) Average accuracy 1 − eave , where eave is the average absolute error

of the transmittance per frequency point; 3) Minimum accuracy 1 − emax , where emax represents the

largest absolute error of the transmittance over all frequencies. Figure 4a displays the three
accuracies in our experiments for various classes of geometry. The average and minimum
accuracies are calculated based on the correct ones in s’ in terms of the geometric accuracy. We
also note that these accuracies are data-dependent and may vary when the distribution of geometric
data changes.

Next, to further exemplify the generality and versatility of the constructed network, we feed into
the critic mixed data from all classes of geometry with over 8,000 data points, and the test samples

7
are also randomly selected from all possible geometries. As shown in the last column of Fig. 4a,
the accuracy in this situation does not degrade compared to prior studies with only a single
geometry input. An example of a discovered structure with mixed geometric input is presented in
Fig. 4b and 4c. Just as before, Fig. 4b and 4c indicate the FEM-simulated spectra of the test pattern
s and the generated pattern s’ in response to the spectral demand, respectively. Although no
specific class of geometry is indicated during the training, the generator is able to reach a pattern,
s’, that not only geometrically resembles s, but more importantly, possesses transmittance spectra,
T’, nearly identical to that of T. In general, if the input geometric data contains more than one
topology that satisfies the spectral demand, the network may generate some or all of them in a
probabilistic manner. Moreover, by changing the distribution of the geometric data, the user may
achieve diverse solutions in response to the same spectral request, thereby mitigating the potential
of degeneracy, as described earlier.

As a final example, we demonstrate the efficacy of our approach in the reverse design of a
metasurface for user drawn spectral responses, T. In practice, the desired spectra at the input are
user defined, and the existence of solutions is not guaranteed. This is particularly true when certain
constraints are applied to the metasurface design. For instance, in the present study, parameters
such as the materials used, the unit cell size, and the thickness of the patterned layer are all
predefined in the training data. Nevertheless, when the simulator is sufficiently robust, the network
is still able to unearth the best possible pattern that yields spectra T’ with minimized deviation
from the input spectra T. To demonstrate this feature, here we design a metasurface with the desired,
user-defined spectra behavior shown in the Fig. 4d: i) Txx and Tyy are two Gaussian-like

resonances with randomly chosen mean µ , variance σ , and amplitude a, and ii) Txy and Tyx are

zero. The generated pattern along with its spectra T’ is shown in the Fig. 4e. Although there exists
no exact solution to spectral demand described above, the network eventually generates patterns
whose spectra share common features with the input spectra including the resonance frequency,
the spectral bandwidth, and the transmission magnitude.

4. Conclusion and outlook

In this work, we have proposed a generative, deep, network model which efficiently discovers
and optimizes unit cell patterns of metasurfaces in response to user-defined, on-demand spectra at

8
the input. Our model addresses the pursuit of reverse design technology for photonic structures,
and helps to relieve computational and specialist resources in traditional metasurface design from
iterative simulations and parameter sweeping, up to generalizing the process for users lacking a
solid knowledge base in optics. The model developed here is mostly based on the unsupervised
learning, which guarantees efficient generation of structural patterns independent of human
experience. This feature is crucial in the investigation of new structures and novel phenomena in
optics and beyond. In addition, as our model can process multiple input spectra without the loss of
efficiency, the workload for complicated problems that require multiple metasurfaces or gradient
structural distributions will be significantly reduced. The performance of the model can be further
improved by applying certain refinements, such as the use of a more sophisticated network
configuration and the introduction of loss functions with more practical physical meanings.

The methodology we have developed is readily extended to the pursuit of desired complex
values of the reflection and transmission coefficients, which is essential to the applications like
meta-lenses and meta-holograms where both the amplitude and the phase of the light waves matter.
In the current work, we restrict the unit pattern of the metasurface as a single metallic particle with
continuous topology. By revising the geometric dataset and refining the simulator, unit cells with
multiple particles and complicated structures can be designed and optimized as well. The increased
arbitrariness of topology can significantly improve the spectral accuracy of the outcome from our
model. Moreover, if sufficient data for the training of the simulator is available, greater degrees of
freedom in the design can be made by slightly modifying the network architecture to include
modifications of the lattice constant or optimization of the pattern thickness. The developed
framework can be adapted to many other applications in optics and materials science, including
photonic crystals, 3D metamaterials, imaging systems, phase transitions, etc. We envision broad
and growing utilization of the deep learning technology in the physics realm, so that scientists and
engineers will be largely relieved from tedious processes of trial and error, and instead focus more
on truly creative thoughts yet to be broached by the machine.

9
Figures

Figure1 | Transitioning metasurface design from conventional trial-and-error approaches to

neural network mediated inverse design. (a) Both, simulation and inverse design, enact
structure-property relationships to generate an optical spectrum from a metasurface and vice-versa.
In this work, both processes will be replaced by deep neural networks. (b) Architecture of the
proposed network for AI-based optical design. Three networks, the generator (G), the simulator
(S), and the critic (D) constitute the complete architecture. The generator accepts the spectra T and
noise z and produces possible patterns. The simulator is a pretrained network that approximates
the transmittance spectrum Tˆ for a given pattern at its input, and the critic evaluates the distance
of the distributions between the geometric data and the patterns from the generator. While training
the generator, the produced patterns vary according to the feedback obtained from S and D. Valid
patterns are documented during the training process, and are smoothed to qualify as candidate
structures.

10
Figure 2 | Effectiveness of the engineered simulator and critic networks. (a) The unit cell of
the metasurface used in our case study of the inverse design problem. The pattern is allowed to
vary for a desired spectral response, with the following structural and material constraints enforced:
gold for the pattern and glass as the substrate; unit cell size w = 340 nm; thickness of gold d = 50
nm. (b) Transmission spectra of a representative structure shown in (a), obtained by FEM
electromagnetic simulation (solid lines) and by the simulator network (circles), respectively. (c)
Generated patterns during the training process after certain iterations, with and without the critic
network. Only geometric data categorized as a cross are fed into the critic in this example.

11
Figure 3 | Generating patterns with a predesigned class of geometric data. (a) Test patterns s
are depicted in the top row and the corresponding generated patterns s’ are listed in the bottom
row. Each shape provides a sample of the different classes of geometric data input to the critic
network. (b) Transmittance spectra, T’, of a test pattern s, to be fed to the network. (c) FEM
simulated transmittance of the retrieved pattern s’, from the neural network based on the input as
in (b). The unit cells of s and s’ are shown in the lower right corner of each figure. This result is
achieved when the critic only receives geometric data of the elliptical class. (d) and (e) An example
of results with a modified MNIST handwritten digital dataset as the input, geometric data. Note
that in this experiment, we intentionally excluded digit “5” in the input geometric data.

12
Figure 4 | Statistical accuracy of the model and examples of generated patterns with mixed
classes of geometric data. (a) Geometric accuracy, average accuracy, and minimum accuracy of
experiments with the different classes of geometric data. (b) and (c) Example of results with a
mixture of different classes of geometric data used at the input. Transmittance spectra of the test
structure s and the generated pattern s’ are shown in the (b) and (c), respectively, with the unit cell
depicted as the inset in each figure. (d) and (e) Example of Inverse design of metasurfaces with
human-defined spectra. (d) Desired transmittance spectra as the input to the generator, where Txx
and Tyy are two randomly generated Gaussian-like responses with parameters a, µ, and σ, while
Txy and Tyx are 0 throughout the frequency range of interest. (e) The resultant unit cell generated
by our AI model to fit the target spectra, along with the FEM simulated transmittance spectra of
this generated metasurface.

13
References

1 Yu, N. & Capasso, F. Flat optics with designer metasurfaces. Nature Materials 13, 139 (2014).
2 Kildishev, A. V., Boltasseva, A. & Shalaev, V. M. Planar photonics with metasurfaces. Science 339,
1232009 (2013).
3 Lin, D., Fan, P., Hasman, E. & Brongersma, M. L. Dielectric gradient metasurface optical elements.
Science 345, 298-302 (2014).
4 Chu, C. H. et al. Active dielectric metasurface based on phase‐change medium. Laser & Photonics
Reviews 10, 986-994 (2016).
5 Silva, A. et al. Performing mathematical operations with metamaterials. Science 343, 160-163
(2014).
6 Landy, N. I., Sajuyigbe, S., Mock, J., Smith, D. & Padilla, W. Perfect metamaterial absorber. Physical
Review Letters 100, 207402 (2008).
7 Zhang, X. & Liu, Z. Superlenses to overcome the diffraction limit. Nature Materials 7, 435 (2008).
8 Larouche, S., Tsai, Y.-J., Tyler, T., Jokerst, N. M. & Smith, D. R. Infrared metamaterial phase
holograms. Nature Materials 11, 450 (2012).
9 Pfeiffer, C. & Grbic, A. Metamaterial Huygens’ surfaces: tailoring wave fronts with reflectionless
sheets. Physical Review Letters 110, 197401 (2013).
10 Arbabi, A., Horie, Y., Bagheri, M. & Faraon, A. Dielectric metasurfaces for complete control of
phase and polarization with subwavelength spatial resolution and high transmission. Nature
Nanotechnology 10, 937-943 (2015).
11 Butet, J., Brevet, P.-F. & Martin, O. J. Optical second harmonic generation in plasmonic
nanostructures: From fundamental principles to advanced applications. Acs Nano 9, 10545-10562
(2015).
12 Li, G. et al. Continuous control of the nonlinearity phase for harmonic generations. Nature
Materials 14, 607 (2015).
13 Carrasquilla, J. & Melko, R. G. Machine learning phases of matter. Nature Physics 13, 431 (2017).
14 Chen, Y., Lin, Z., Zhao, X., Wang, G. & Gu, Y. Deep learning-based classification of hyperspectral
data. IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing 7, 2094-
2107 (2014).
15 Segler, M. H., Preuss, M. & Waller, M. P. Planning chemical syntheses with deep neural networks
and symbolic AI. Nature 555, 604 (2018).
16 Poloni, C., Giurgevich, A., Onesti, L. & Pediroda, V. Hybridization of a multi-objective genetic
algorithm, a neural network and a classical optimizer for a complex design problem in fluid
dynamics. Computer Methods in Applied Mechanics and Engineering 186, 403-420 (2000).
17 Melnikov, A. A. et al. Active learning machine learns to create new quantum experiments.
Proceedings of the National Academy of Sciences, 201714936 (2018).
18 Lu, J. & Vučković, J. Nanophotonic computational design. Optics Express 21, 13351-13367 (2013).
19 Piggott, A. Y. et al. Inverse design and demonstration of a compact and broadband on-chip
wavelength demultiplexer. Nature Photonics 9, 374-377 (2015).
20 Sitawarin, C., Jin, W., Lin, Z. & Rodriguez, A. W. Inverse designed photonic fibers and metasurfaces
for nonlinear frequency conversion. arXiv preprint arXiv:1711.07810 (2017).
21 Wiecha, P. R. et al. Evolutionary multi-objective optimization of colour pixels based on dielectric
nanoantennas. Nature Nanotechnology 12, 163 (2017).
22 Hoole, S. R. H. Artificial neural networks in the solution of inverse electromagnetic field problems.
IEEE Transactions on Magnetics 29, 1931-1934 (1993).

14
23 Kabir, H., Wang, Y., Yu, M. & Zhang, Q.-J. Neural network inverse modeling and applications to
microwave filter design. IEEE Transactions on Microwave Theory and Techniques 56, 867-879
(2008).
24 Liu, D., Tan, Y., Khoram, E. & Yu, Z. Training deep neural networks for the inverse design of
nanophotonic structures. ACS Photonics 5, 1365-1369 (2017).
25 Peurifoy, J. et al. in Physics and Simulation of Optoelectronic Devices XXVI. 1052607 (International
Society for Optics and Photonics).
26 Goodfellow, I. et al. in Advances in Neural Information Processing Systems. 2672-2680.
27 Radford, A., Metz, L. & Chintala, S. Unsupervised representation learning with deep convolutional
generative adversarial networks. arXiv preprint arXiv:1511.06434 (2015).
28 Arjovsky, M., Chintala, S. & Bottou, L. in International Conference on Machine Learning. 214-223.
29 Gulrajani, I., Ahmed, F., Arjovsky, M., Dumoulin, V. & Courville, A. C. in Advances in Neural
Information Processing Systems. 5769-5779.
30 LeCun, Y. et al. Learning algorithms for classification: A comparison on handwritten digit
recognition. Neural Networks: the Statistical Mechanics Perspective 261, 276 (1995).

15