Research Article Vol. 30, No.

13 / 20 Jun 2022 / Optics Express 22901

Data-driven design of thin-film optical systems

using deep active learning
YOUNGJOON H ONG 1 AND DAVID P. N ICHOLLS 2,*
1 Departmentof Mathematics, Sungkyunkwan University, Suwon 440-746, Republic of Korea
2 Departmentof Mathematics, Statistics, and Computer Science, University of Illinois at Chicago, Chicago,
IL 60607, USA
* davidn@uic.edu

Abstract: A deep learning aided optimization algorithm for the design of flat thin-film multilayer
optical systems is developed. The authors introduce a deep generative neural network, based
on a variational autoencoder, to perform the optimization of photonic devices. This algorithm
allows one to find a near-optimal solution to the inverse design problem of creating an anti-
reflective grating, a fundamental problem in material science. As a proof of concept, the authors
demonstrate the method’s capabilities for designing an anti-reflective flat thin-film stack consisting
of multiple material types. We designed and constructed a dielectric stack on silicon that exhibits
an average reflection of 1.52 %, which is lower than other recently published experiments in the
engineering and physics literature. In addition to its superior performance, the computational
cost of our algorithm based on the deep generative model is much lower than traditional nonlinear
optimization algorithms. These results demonstrate that advanced concepts in deep learning can
drive the capabilities of inverse design algorithms for photonics. In addition, the authors develop
an accurate regression model using deep active learning to predict the total reflectivity for a given
optical system. The surrogate model of the governing partial differential equations can then be
broadly used in the design of optical systems and to rapidly evaluate their behavior.

© 2022 Optica Publishing Group under the terms of the Optica Open Access Publishing Agreement

1. Introduction
Multilayered flat thin-film diffraction gratings are essential optical elements in nanoplasmonic
and photonic devices as they modulate light intensity and spectral composition in such systems
[1]. In practice it is often important to determine the best multilayer design among a wide choice
of dielectrics and metals of varying thicknesses to achieve a desired reflection and transmission
spectrum. This problem has a long and distinguished history, beginning at least with Baumeister’s
characterization of optimal coating design as an optimization problem in 1958 [2]. Numerous
inverse algorithms have been constructed to find such optimal designs for use in efficient photonic
devices [3]. For instance, a multitude of conventional optimization methods perform well at this
particular task including genetic algorithms [4,5], topology optimization [6,7], and the needle
optimization technique [8,9] which, e.g., have delivered the remarkable results on design of
anti–reflective coatings found in [10,11]. However, it is well known that this problem is highly
nonlinear and non-convex, featuring numerous suboptimal local minima which make it difficult
to find the global optimum [3].
In the past decade the artificial intelligence technique of machine learning, in particular deep
learning, has revolutionized many fields of computational science, and nano-phototonics is
no exception. There are already a multitude of survey articles on the topic, e.g., [12], and a
staggering number of techniques have been brought to bear on this task. Of particular note,
we point out the work of Peurifoy et al. [13] who trained an artificial neural network (ANN)
and then optimized this using conventional techniques, and Liu et al. [14] who considered a
tandem approach which begins the same, but then trains a second ANN to produce desired
reflectivity values. We also mention the work of Unni et al. [15] on convolutional mixture

#459295 https://doi.org/10.1364/OE.459295
Journal © 2022 Received 23 Mar 2022; revised 26 May 2022; accepted 26 May 2022; published 8 Jun 2022
 Research Article Vol. 30, No. 13 / 20 Jun 2022 / Optics Express 22902

density networks and the paper of Barry et al. [16] demonstrating how nature itself identifies
optimal structures for its own purposes. Recently, deep generative models including (conditional)
variational autoencoders have been introduced to optimize the shape of two-dimensional unit
cells giving a two-dimensional binary image [17–20]. As deep neural networks deliver such
superior performance in computer vision, one expects that shape optimization algorithms can
also be improved. We close by mentioning the deep learning approach that has been considered
to find optimal designs using a ResNet generative model [21].
In this paper, we develop an alternative approach based on deep learning strategies which is
both fast and efficient, and which can readily be extended to more general thin-film structures
featuring, e.g., corrugated layer interfaces. In this way we view our new algorithm as particularly
promising in the design of new metamaterials [22]. More specifically, in this paper we develop a
novel and effective inverse design algorithm with the aid of deep learning to identify m-layer
flat thin-film stacks composed of materials with varying refractive indices and thicknesses. In
this scheme we begin by constructing a structure/response database using a rapid and accurate
classical Fresnel solver [23]. Then, we make use of a generative deep neural network (DNN),
a conditional variational autoencoder (CVAE), to obtain a nearly optimal design of the optical
system. The goal of this CVAE is to minimize the average reflection of the structure over a range
of incident illumination angles (0 ≤ θ ≤ π/3) and wavelengths (400 nm ≤ λ ≤ 1600 nm), which
we denote ∫ π/3 ∫ 1600
3 1
O(p) = R(λ, θ, p) dλ dθ, (1)
π 1200 0 400
where p is the design vector and R is specified in (4). We also propose a deep active learning
algorithm to effectively search for the optimal solution. While our method effectively and
efficiently delivers a configuration with quite a small reflectivity, we cannot guarantee that it is
the “best” design with the smallest possible reflectivity.

2. Methods
In this section we present the governing equations for a thin-film multilayer optical system and
an accurate method for numerically approximating its solution based upon the classical Fresnel
equations [23]. We then present a novel approach to coupling this to not only a Deep Learning
(DL) algorithm (we consider a CVAE), but also an active learning methodology.

2.1. Governing equations

We consider a thin-film multilayer optical system consisting of flat layers of varying materials
with rather arbitrary thicknesses. The variety of dielectrics and thicknesses leads to a considerable
practical design challenge of finding a “best” grating structure. To be more precise, dielectrics
occupy each of the m-many domains

S1 := {y>g1 }; Sm := {y<gm−1 }; and Sj := {gj <y<gj−1 }, 2 ≤ j ≤ m − 1;

where the (flat) interface locations are given by {y = gj }j=1

m−1 . This structure is illuminated by

incident radiation of frequency ω and angle θ in the uppermost layer, S1 , of the form

vinc = e−iωt+iαx−iβy , α = k0 sin(θ), β = k0 cos(θ), k0 = ω/c0 ,

and c0 is the speed of light in the vacuum. Factoring out time-dependence of the form exp(−iωt)
we define the (reduced) scattered fields

vj = vj (x, y) in Sj for 1 ≤ j ≤ m,
 Research Article Vol. 30, No. 13 / 20 Jun 2022 / Optics Express 22903

and seek α-quasiperiodic, outgoing (upward/downward propagating) solutions of the following

system of Helmholtz equations [23]

∆vj + kj2 vj = 0, in Sj , 1 ≤ j ≤ m, (2a)

v1 − v2 = −eiβg1 eiαx , at y = g1 , (2b)

∂y v1 − τ12 ∂y v2 = (iβ)e iβg1 iαx
e , at y = g1 , (2c)
vj − vj+1 = 0, at y = gj , 2 ≤ j ≤ m − 1, (2d)
∂y vj − τj2 ∂y vj+1 = 0, at y = gj , 2 ≤ j ≤ m − 1, (2e)
where kj = nj k0 = nj (ω/c0 ) is the wavenumber in layer j (with refractive index nj ) and
{︄
1, for Transverse Electric (TE) polarization,
τj = 2
(kj /kj+1 ) , for Transverse Magnetic (TM) polarization,

for 1 ≤ j ≤ m − 1.

2.2. Numerical method: the Fresnel solver

It is well known [23] that the most general solutions of (2a) are
(︂ j j
)︂
vj (x, y) = Uj eiβ y + Dj e−iβ y eiαx , D1 = Um = 0,

where βj := (kj )2 − α2 , Im(βj ) ≥ 0, and the upward/downward propagating wave conditions

√︁

enforce D1 = Um = 0, respectively. The remaining constants {U1 , D2 , U2 , . . . , Dm−1 , Um−1 , Dm }

are determined from the boundary conditions (2b)–(2e). This results in a linear system of
equations to be solved

A⃗v = ⃗r, ⃗v := (U1 , D2 , U2 , . . . , Dm−1 , Um−1 , Dm )T , ⃗r := eiβg1 (−1, (iβ), 0, . . . , 0)T , (3)

and A is pentadiagonal with readily derived entries [23]. We denote the direct solution of the
Fresnel equations, A⃗v = ⃗r, as the Fresnel Solver which, we point out, can be accomplished in
linear (in m) time via the Thomas Algorithm [24]. Of the many quantities that one can compute
from this solution, the one of paramount importance to our current study is the reflectivity

R = |U1 | 2 . (4)

We point out that the solution of these equations is the order-zero approximation produced by
our recently developed High-Order Perturbation of Surfaces (HOPS) algorithm [25] implemented
with a Transformed Field Expansions approach. As this methodology is designed for structures
with corrugated interfaces the full power of this algorithm is not necessary in the current context,
however, in a forthcoming publication we will describe the extension of our algorithm to the case
of corrugated interfaces which will require the full HOPS methodology.

2.3. Variational autoencoder

Generative models in combination with neural networks, such as variational autoencoders (VAEs),
are used to learn complex distributions underlying datasets, see e.g. [26,27]. Described simply,
VAEs consist of an encoder, a latent space, and a decoder (see Fig. 1). VAEs presume that a given
training sample is generated from a latent representation, which is then sampled by a decoder
(with a prior Gaussian distribution).
 Research Article Vol. 30, No. 13 / 20 Jun 2022 / Optics Express 22904



  
 
   

 


  
 
 


 

 




 


 


 

     

 

Fig. 1. Depiction of a varitational autoencoder.

As a generative neural network, VAEs have been successfully utilized in various domains
from image generation and natural language processing to anomaly detection and clustering
tasks (see, e.g., [26] and references therein). VAEs are regularized autoencoders which also
feature an encoder and decoder. The encoder maps high-dimensional data to low-dimensional
latent vectors that capture principal features, then the decoder maps the latent vector back
to the high-dimensional space. While there are many applications of autoencoders (such as
dimensionality reduction, anomaly detection, and noise removal) they are not generally adequate
as generative models [28]. Indeed, once the autoencoder is trained there is no opportunity to
produce any new content with both encoder and decoder. By contrast, a VAE regularizes the
encoding distribution to ensure that its latent space has good properties to generate a new dataset.
More precisely, the encoder in the VAE maps input data points not to the latent space but to the
distribution of the latent space. Then, the encoder produces the mean and covariance matrix
values that are a function of the input data. The decoder exploits the latent space distribution as
an input to generate distributions of data. In the VAE, the loss function consists of reconstruction
loss and regularization loss. The reconstruction loss is identical to that used by autoencoders,
while the regularization loss is the Kullback-Leibler (KL) divergence between the Gaussian
distribution from the encoder and a standard Gaussian distribution [29].
For the optical system we consider here the input vector x consists of a collection of refractive
indices and layer thicknesses. The VAE architecture aims to learn a stochastic mapping between
the observed data space x and a latent space z which can be interpreted as a directed model with
a joint distribution pθ (x, z) such that pθ (x, z) = pθ (x|z)pθ (z), where θ is a learnable parameter and
pθ (z) is the prior distribution of the latent variable. The conditioned distribution pθ (x|z) can be
parameterized by a decoder but the distribution is generally intractable. To resolve this issue a
VAE introduces another deep neural network (encoder) to map x back to the latent vector z by
approximating the posterior distribution (see Fig. 1). With the encoder and decoder networks,
the likelihood function for the training has a tractable representation and can be derived by the
evidence lower bound (ELBO):

Loss = Eq(z|x) [log(p(x|z))] − DKL [q(z|x)||p(z)],

where DKL stands for the KL divergence. In this paper, we employ a conditional variational
autoencoder (CVAE), which is an extension of the VAE suitable for incorporating a control on a
specified condition [27]. The CVAE is believed to insert label information in the latent space to
force a deterministic, constrained, representation of the learned data. In contrast to a VAE, a
 Research Article Vol. 30, No. 13 / 20 Jun 2022 / Optics Express 22905

CVAE has control on the data generation process so, by changing the conditional variable (which
refers to the reflectivity in our model), inputs of an optical system for a specified reflectivity can
be generated.

2.4. Active learning

Active learning is a machine learning strategy which interactively queries a user to generate new
data points with desirable properties. These queries are usually in the form of unlabeled data and,
to improve the underlying machine learning model, it is crucial for the system to propose records
for interactive labelling effectively. Active learning is often called optimal experimental design
(OED) in the engineering literature (see, e.g., [30,31]) as it finds the most uncertain points and
adds them into the training set in an iterative way [32]. In this paper we develop surrogate models
using deep learning techniques (implemented by a multilayer perceptron (MLP)) to identify
efficient device designs. However, in order to improve these designs, additional training of the
surrogate model becomes increasingly expensive.
We aim to obtain a better model by adding minimal additional training data, and, for this
purpose, we adopt an active learning strategy that selects only training points which enhance the
accuracy of the MLP surrogate model. To produce such training samples, the CVAE is used to
generate an optimal system design which hopefully gives small reflectivity (see Fig. 2). To be
more precise, let L be the number of training set data points initially generated by the Fresnel
solver, and, with these L points, we train our CVAE model to generate K many optical systems
(which hopefully possess small reflectivity) by imposing the smallness of the reflectivity as the
condition on the CVAE. Few (say P many) of these K optical systems generated by the CVAE
produce the reflectivity expected by the condition enforced since the randomly generated training
set does not contain sufficient data close to our target. However, they are sufficiently close that
they merit further investigation with our (relatively expensive) Fresnel solver and we add them to
our training set; this is then repeated M times. This algorithm is depicted in Fig. 2. The initial
reflectivity histogram of randomly generated data (the “initial training set”) is plotted in Fig. 3(a)
while the augmented dataset generated by our CVAE and active learning process is shown in
Fig. 3(b). The second figure is, naturally, nearly identical to the first figure as all of the original
data is retained while only “desirable” (i.e., small reflectivity) configurations are added.

 
   

 
  

 

  
 


  
 
 

  

   
 
   

  
  



   

Fig. 2. Depiction of an active learning implementation.

 Research Article Vol. 30, No. 13 / 20 Jun 2022 / Optics Express 22906

5000 5000

4000 4000
frequency

frequency
3000 3000

2000 2000

1000 1000

0 0
0.0 0.1 0.2 0.3 0.4 0.5 0.6 0.0 0.1 0.2 0.3 0.4 0.5 0.6

reflectivity reflectivity

(a) Histogram of the original dataset. (b) Histogram of the CVAE augmented dataset.

Fig. 3. Histogram of the (a.) original and (b.) CVAE augmented data sets.

3. Numerical experiments
As we described above, the CVAE architecture consists of an encoder and decoder network. The
encoder network can be described as

INPUT → FC → sigmoid → FC → BN → sigmoid → FC → sigmoid → FC → OUTPUT,

while the decoder architecture is given by

INPUT → FC → RELU → FC → BN → RELU → FC → RELU → FC → OUTPUT,

where “FC” stands for a fully connected layer and “BN” denotes batch normalization. In more
detail:
• The input of the encoder consists of refractive indices and thickness for each layer with
reflectivity as a condition.
• The dimension of the output of the encoder is 4 × 2, which is the same as the dimension of
the latent space.
• In the encoder the output sizes of the FCs are 384, 384, 128, and 4, respectively.
• The input of the decoder consists of samples of the four latent variables (for which we have
four means and four variances).
• In the decoder the output sizes of the FCs are 128, 384, 384, and 6, respectively.
• The BN has momentum = 0.1 and stability constant ε = 10−5 .
We applied our algorithm to the design of an anti-reflection (AR) coating for a silicon solar
cell consisting of three layers of dielectrics [3]. This thin-film stack was designed to minimize the
average reflection at an air-silicon interface over the incident illumination angle range [0, π/3] and
wavelength range [400, 1100] nm in TM polarization as in (1). As a benchmark, we compared
our results with those from [3] which provides a guaranteed global optimum solution using a
parallel branch-and-bound method. Their algorithm required extensive searching through the full
design space and utilized more than two weeks of CPU time to solve for the global optimum.
To be consistent with [3] we generated an L = 50, 000 member training set from our Fresnel
solver whose refractive indices and layer thicknesses were randomly selected from the intervals
[1.09, 2.60] and [5, 200] nm, respectively. Here we supplemented with active learning using
K = 5000 at each of M = 3 iterations resulting in P = 2000 additional datapoints. A histogram
 Research Article Vol. 30, No. 13 / 20 Jun 2022 / Optics Express 22907

of the resulting reflectivities are given in Fig. 4 which shows that the optimized devices generated
from our CVAE and deep active learning algorithms have average reflectivities from approximately
1.5% to 3%, which is quite a small range of values compared to a randomly generated set. A
fraction of the suggested devices were near the global optimum, and the best device had an
efficiency of 1.52 %. This best device had layer thicknesses

151.722nm, 91.258nm, 61.637nm,

and refractive indices

1.187, 1.771, 2.717,
from top layer to bottom. The total computing time that our CVAE and deep active learning
algorithm required for training was less than 30 minutes with a single GPU (NVIDA RTX-3090).
All of the devices sampled from our algorithm were near the global optimum, showing the
ability of the generative network to produce a narrow distribution of devices. Another feature
of such biased data generation is the construction an accurate regression model using the MLP.
Figures 5(a)–5(d) show that the MLP model trained by the augmented data set (generated by the
active learning algorithm) gives a better result than the model trained by the original (normally
distributed) dataset.

40

30 1.52%
Frequency

20

10

0
0.014 0.016 0.018 0.020 0.022 0.024 0.026 0.028 0.030
Reflectivity

Fig. 4. Histogram of reflectivities generated by our CVAE and deep active learning
algorithm.

The reflectivity map of the best device (with efficiency 1.52 %) is depicted in Fig. 6 with the
full set of incidence angles, and in Fig. 6(b) for three choices of the angle. We remark that the
data-driven design under consideration can be generalized to an arbitrary number of layers, and
the machine learning procedure becomes more effective since the computational cost is very
expensive in this case.
 Research Article Vol. 30, No. 13 / 20 Jun 2022 / Optics Express 22908

0.10
0.5 Rel L 2 error = 0.0033 Rel L 2 error = 0.0557
0.08
0.4

0.06
Prediction

Prediction
0.3

0.04
0.2
0.02
0.1
0.00
0.1 0.2 0.3 0.4 0.5 0.00 0.02 0.04 0.06 0.08 0.10
Numerical simulation Numerical simulation

(a) Regression using deep learning with the original (b) Regression using deep learning with the original
dataset. This plot shows reflectivities which range dataset. This plot shows only reflectivities which range
from 2 % to 50 %. from 2 % to 7 %.

0.5 0.10
Rel L 2 error = 0.0007 Rel L 2 error = 0.0263
0.08
0.4

0.06
Prediction

Prediction
0.3

0.04
0.2
0.02
0.1
0.00
0.1 0.2 0.3 0.4 0.5 0.00 0.02 0.04 0.06 0.08 0.10
Numerical simulation Numerical simulation

(c) Regression using deep active learning with the (d) Regression using deep active learning with the
augmented dataset. This plot shows reflectivities which augmented dataset. This plot shows only reflectivities
range from 2 % to 50 %. which range from 2 % to 7 %.

5.5.Regression
Fig.
Fig. Regressionmodel
modelcomparison
comparison with the original
with the originaland
andaugmented
augmenteddatasets.
datasets.Clearly,
Clearly,
thethe
relative L2 -errors
relative 𝐿 2 -errors
of the
of regression model
the regression with with
model the augmented datasets
the augmented are lower
datasets than the
are lower
original regression
than the original model.
regression model.

Reflectivity map
1100 0.08

0.08 0 degree
30 degree
60 degree
0.07
1000
0.07

900
0.06
0.06
Reflectivity map
Wavelength (nm)

1100 0.08
0.05
0.05
0 degree
Reflectivity

800
0.08
0.07 0.04 30 degree
1000 0.04

0.07 60 degree
700

0.06
0.03 0.03
900 0.06
600
0.02
0.05
Wavelength (nm)

0.02
0.05
800
Reflectivity

0.01
0.04 0.04
500
0.01

700 0.00
400
0.03 0.03
0 10 20 30 40 50 60 400 500 600 700 800 900 1000 1100
600 Wavelength (nm)
Incident angle (deg)

0.02 0.02

500
(a) Optimal grating reflectivity versus incident 0.01
0.01 angle (b) Optimal grating reflectivity versus three choices of
(degrees) and wavelength (nanometers). 0.00 angle (degrees) and wavelength (nanometers).
incident
400
0 10 20 30 40 50 60 400 500 600 700 800 900 1000 1100
Incident angle (deg) Wavelength (nm)
Fig. 6. Optimal grating reflectivity versus incident angle and wavelength.
(a) Optimal grating reflectivity versus incident angle (b) Optimal grating reflectivity versus three choices of
(degrees) and wavelength (nanometers). incident angle (degrees) and wavelength (nanometers).
203 Disclosures. The authors declare no conflicts of interest.
Fig. 6. Optimal grating reflectivity versus incident angle and wavelength.
204 Data Availability Statement. Data underlying the results presented in this paper are not publicly available
205 at this time but may be obtained from the authors upon reasonable request.
 Research Article Vol. 30, No. 13 / 20 Jun 2022 / Optics Express 22909

4. Conclusions
In this paper we introduced a deep learning aided optimization algorithm for thin-film multilayer
optical systems. We constructed a deep generative neural network, based on a variational
autoencoder, to perform optimization of photonic devices. The incorporation of the variational
autoencoder helps to improve our search for the optimal grating design. Benchmark calculations
of our algorithm for the problem of designing anti-reflection coatings show that the generative
model is effective in searching for global optima, is computationally efficient, and outperforms a
number of alternative design algorithms.
Funding. Basic Science Research Program through the NRF funded by Ministry of Education (NRF-2021R1A2C1093579);
Korean Government (MSIT) (2022R1A4A3033571); SungKyunKwan University and the BK21 FOUR (Graduate School
Innovation) funded by the Ministry of Education (MOE) and National Research Foundation of Korea; National Science
Foundation (DMS-2111283).
Disclosures. The authors declare no conflicts of interest.
Data Availability. Data underlying the results presented in this paper are not publicly available at this time but may
be obtained from the authors upon reasonable request.

References
1. S. A. Maier, Plasmonics: Fundamentals and Applications (Springer, New York, 2007).
2. P. Baumeister, “Design of multilayer filters by successive approximations,” J. Opt. Soc. Am. 48(12), 955–958 (1958).
3. P. Azunre, J. Jean, C. Rotschild, V. Bulovic, S. G. Johnson, and M. A. Baldo, “Guaranteed global optimization of
thin-film optical systems,” New J. Phys. 21(7), 073050 (2019).
4. J. Dobrowolski and R. Kemp, “Refinement of optical multilayer systems with different optimization procedures,”
Appl. Opt. 29(19), 2876–2893 (1990).
5. S. Martin, J. Rivory, and M. Schoenauer, “Synthesis of optical multilayer systems using genetic algorithms,” Appl.
Opt. 34(13), 2247–2254 (1995).
6. J. Jensen and O. Sigmund, “Topology optimization for nano-photonics,” Laser Photonics Rev. 5(2), 308–321 (2011).
7. A. M. Hammond, A. Oskooi, M. Chen, Z. Lin, S. G. Johnson, and S. E. Ralph, “High-performance hybrid
time/frequency-domain topology optimization for large-scale photonics inverse design,” Opt. Express 30(3),
4467–4491 (2022).
8. A. V. Tikhonravov, M. K. Trubetskov, and G. W. DeBell, “Application of the needle optimization technique to the
design of optical coatings,” Appl. Opt. 35(28), 5493–5508 (1996).
9. A. V. Tikhonravov and M. K. Trubetskov, “Modern design tools and a new paradigm in optical coating design,” Appl.
Opt. 51(30), 7319–7332 (2012).
10. J. Dobrowolski and B. Sullivan, “Universal antireflection coatings for substrates for the visible spectral region,” Appl.
Opt. 35(25), 4993–4997 (1996).
11. F. Lemarquis, T. Begou, A. Moreau, and J. Lumeau, “Broadband antireflection coatings for visible and infrared
ranges,” CEAS Space J. 11(4), 567–578 (2019).
12. P. R. Wiecha, A. Arbouet, C. Girard, and O. L. Muskens, “Deep learning in nano-photonics: inverse design and
beyond,” Photonics Res. 9(5), B182–B200 (2021).
13. J. Peurifoy, Y. Shen, L. Jing, Y. Yang, F. Cano-Renteria, B. G. DeLacy, J. D. Joannopoulos, M. Tegmark, and M.
Soljacic, “Nanophotonic particle simulation and inverse design using artificial neural networks,” Sci. Adv. 4(6),
eaar4206 (2018).
14. D. Liu, Y. Tan, E. Khoram, and Z. Yu, “Training deep neural networks for the inverse design of nanophotonic
structures,” ACS Photonics 5(4), 1365–1369 (2018).
15. R. Unni, K. Yao, and Y. Zheng, “Deep convolutional mixture density network for inverse design of layered photonic
structures,” ACS Photonics 7(10), 2703–2712 (2020).
16. M. A. Barry, V. Berthier, B. D. Wilts, M.-C. Cambourieux, P. Bennet, R. Pollés, O. Teytaud, E. Centeno, N. Biais,
and A. Moreau, “Evolutionary algorithms converge towards evolved biological photonic structures,” Sci. Rep. 10(1),
12024 (2020).
17. W. Ma, F. Cheng, Y. Xu, Q. Wen, and Y. Liu, “Probabilistic representation and inverse design of metamaterials based
on a deep generative model with semi-supervised learning strategy,” Adv. Mater. 31(35), 1901111 (2019).
18. Z. Liu, L. Raju, D. Zhu, and W. Cai, “A hybrid strategy for the discovery and design of photonic structures,” IEEE J.
on Emerg. Sel. Top. Circuits Syst. 10(1), 126–135 (2020).
19. W. Ma and Y. Liu, “A data-efficient self-supervised deep learning model for design and characterization of
nanophotonic structures,” Sci. China Physics, Mech. Astron. 63(8), 284212 (2020).
20. Z. A. Kudyshev, A. V. Kildishev, V. M. Shalaev, and A. Boltasseva, “Machine learning–assisted global optimization
of photonic devices,” Nanophotonics 10(1), 371–383 (2021).
21. J. Jiang and J. A. Fan, “Multiobjective and categorical global optimization of photonic structures based on ResNet
generative neural networks,” Nanophotonics 10(1), 361–369 (2020).
 Research Article Vol. 30, No. 13 / 20 Jun 2022 / Optics Express 22910

22. L. Wang, Y.-C. Chan, F. Ahmed, Z. Liu, P. Zhu, and W. Chen, “Deep generative modeling for mechanistic-based
learning and design of metamaterial systems,” Comput. Methods Appl. Mech. Eng. 372, 113377 (2020).
23. P. Yeh, Optical waves in layered media, vol. 61 (Wiley-Interscience, 2005).
24. R. J. LeVeque, Finite difference methods for ordinary and partial differential equations (Society for Industrial and
Applied Mathematics (SIAM), Philadelphia, PA, 2007). Steady-state and time-dependent problems.
25. Y. Hong and D. P. Nicholls, “A high–order perturbation of surfaces method for scattering of linear waves by periodic
multiply layered gratings in two and three dimensions,” J. Comput. Phys. 345, 162–188 (2017).
26. D. P. Kingma and M. Welling, “An introduction to variational autoencoders,” Foundations Trends Mach. Learn.
12(4), 307–392 (2019).
27. K. Sohn, H. Lee, and X. Yan, “Learning structured output representation using deep conditional generative models,”
in Advances in Neural Information Processing Systems, vol. 28 C. Cortes, N. Lawrence, D. Lee, M. Sugiyama, and R.
Garnett, eds. (Curran Associates, Inc., 2015), pp. 1–9.
28. I. Goodfellow, Y. Bengio, and A. Courville, Deep Learning (MIT Press, 2016). http://www.deeplearningbook.org.
29. D. M. Blei, A. Kucukelbir, and J. D. McAuliffe, “Variational inference: A review for statisticians,” J. Am. Stat. Assoc.
112(518), 859–877 (2017).
30. H. Ranganathan, H. Venkateswara, S. Chakraborty, and S. Panchanathan, “Deep active learning for image classifica-
tion,” in 2017 IEEE International Conference on Image Processing (ICIP), (2017), pp. 3934–3938.
31. B. Settles, “Active learning,” Synth. Lect. on Artif. Intell. Mach. Learn. 6(1), 1–114 (2012).
32. R. Pestourie, Y. Mroueh, T. V. Nguyen, P. Das, and S. G. Johnson, “Active learning of deep surrogates for pdes:
application to metasurface design,” npj Comput. Mater. 6(1), 164 (2020).