Made available for a limited time for personal research and study only License.

pubs.acs.org/acsaom Article

Machine-Learning-Assisted Design of Multiband Terahertz

Metamaterial Absorber
Mukund Soni and Shikhar Misra*

Cite This: ACS Appl. Opt. Mater. 2023, 1, 1679−1687 Read Online

ACCESS Metrics & More Article Recommendations

See https://pubs.acs.org/sharingguidelines for options on how to legitimately share published articles.
Downloaded via INDIAN INST OF TECH ROORKEE on February 3, 2024 at 12:06:27 (UTC).

ABSTRACT: The design of metamaterials involves engineering

the electromagnetic properties on a subwavelength scale, making
their behavior challenging to predict with models that are time-
consuming and computationally expensive. In recent years, there
has been growing interest in developing various machine learning
(ML) models for designing metamaterials. In this research, we
demonstrate an ML-based framework for the optimal design of
metamaterial absorbers in the terahertz range. Herein, the optical
properties of a metamaterial structure with an open polygon-
shaped meta-atom are studied using numerical simulations. In
addition, machine learning models are developed, specifically, a
forward and an inverse model to predict absorption spectra from
the structure and vice versa using different machine learning
algorithms, including neural networks, decision trees, KNN, and linear regression. The mean square error was as low as 0.003 for the
forward model and 0.02 for the inverse model. This work can be expanded for designing other complex fabricated electromagnetic
structures for use in metamaterials and metasurfaces.
KEYWORDS: metamaterials, terahertz, absorber, machine learning, photonics, light−matter interaction

■ INTRODUCTION
Engineering light−matter interaction at the nanoscale allows
ing, and inefficient. These approaches rely on analytical and
numerical methods such as FDTD simulations and genetic
unprecedented control over its propagation, absorption, algorithms for optimization. While these methods have
reflection, refraction, and emission, enabling exciting applica- successfully designed metamaterials with specific properties,
tions in the field of photonics, quantum technologies, defense, they have limitations. The complex behavior of metamaterials
optoelectronics, renewable integration, and sensing.1−4 Meta- can make it difficult to achieve optimal designs by using these
materials, a type of artificially engineered materials possessing methods. Additionally, these approaches can be computation-
unique and extraordinary properties that are not found in ally expensive and require extensive simulations and manual
naturally occurring materials, are composed of subwavelength- parameter adjustments. To overcome these issues, various
sized structures that can effectively tailor light−matter machine learning techniques, including deep learning, have
interaction resulting in a wide range of unusual properties emerged as a powerful tool in the design and optimization of
such as negative refraction,56 absorber,7 and super-resolution metamaterials.2021 Machine learning (ML) techniques offer
imaging in the microwave to infrared frequency regime.348−11 various applications in the field of metamaterials, such as in the
discovery and design of new metamaterial compositions and
Among them, metamaterial absorbers (MMAs) operational in
structures by predicting properties and optimizing material
the THz range (ranging from 0.1 to 10 THz) have attracted
configurations.2223 It can also characterize the performance of
significant research interest due to their different applications
metamaterials by analyzing experimental or simulation
in high-speed communication, sensing, defense, energy
data.24,25 Further, ML algorithms can solve inverse design
harvesting, and imaging.3410−14
Designing metamaterials poses a significant challenge, as
they consist of subwavelength structures that interact with Received: July 21, 2023
electromagnetic waves, resulting in unique and counterintuitive Revised: August 26, 2023
properties. Traditional approaches to designing and realizing Accepted: September 18, 2023
metamaterials, based on a combination of computational, Published: October 2, 2023
patterning, and self-assembly techniques,15−19 often involve a
trial-and-error process that can be challenging, time-consum-

© 2023 American Chemical Society https://doi.org/10.1021/acsaom.3c00246

1679 ACS Appl. Opt. Mater. 2023, 1, 1679−1687
ACS Applied Optical Materials pubs.acs.org/acsaom Article

Figure 1. (a) Three-dimensional (3-D) view of the proposed metamaterial unit cell. The top and bottom layers are gold and the middle layer is
silicon dioxide. (b) Two-dimensional (2-D) view of the proposed metamaterial. Each unit cell is the meta-atom shown in (a). The arrows indicate
that the structure is expanding in all four directions. (c, d) Dimensional parameters of the proposed metamaterial structure. Out of all of the
parameters, h1, h2, h3, h4, s, and tr are variables.

problems, enabling the engineering of metamaterials with mimic infinite or periodic systems by enforcing periodicity at
desired electromagnetic or acoustic responses.26−29 ML the boundaries, were used on all of the sidewalls. Scattering
techniques have also been applied to optimize manufacturing boundary conditions, which describe how waves or particles
processes to improve robustness and generalization and interact with boundaries, accounting for reflection, trans-
enhance the overall understanding and advancement of mission, or absorption, were used at the top wall. A perfect
metamaterials.30−32 By leveraging ML, researchers can electric conductor, which is an idealized boundary condition
accelerate the development of unique materials with representing a material with infinite conductivity and zero
extraordinary properties for applications in telecommunica- electric field penetration, was applied to the bottom wall. It
tions, sensing, energy harvesting, and more. reflects electromagnetic waves completely without allowing for
In this study, a novel metamaterial perfect absorber (MPA) penetration into the conductor. The simulation was performed
structure, operational in the THz range, with an open polygon- in the frequency domain and uses Maxwell’s equations, which
shaped meta-atom, has been proposed and studied using are a foundational set of equations that describe electro-
numerical simulations as shown in Figure 1a. Practically, such magnetic phenomena. They encompass four fundamental
designs can be realized through a combination of lithography equations that relate electric and magnetic fields to their
and etching techniques. The parameters of this MPA structure sources, conveying essential insights into the behavior of
are tunable, enabling us to model various meta-atom designs, electromagnetic waves and interactions.
and the absorption spectra of these meta-atoms have been
Data Preprocessing
studied. The variations in absorption spectra range from a
single band to multiband spectra, depending on the meta- Absorbance plots corresponding to the metamaterial structure
atom. A forward machine learning model has been proposed, are obtained for approximately 6000 different structural
which predicts the electromagnetic spectrum for a specific parameters as shown in Figure 2e. This was followed by the
design of metamaterial in addition to an inverse machine data preprocessing step before giving the input to the ML
learning model that predicts a novel structure of a models. First, the graphical data are converted into one-
metamaterial absorber operating in a 0.1−6 THz frequency dimensional (1-D) arrays of input and output data. In the data
regime. set, the input point is the structural parameters of the

■ METHODOLOGY
The proposed MPA is simulated using the wave optics module
metamaterial, and the output is the projection of the
absorption spectrum, which is a 1-D array. These absorption
spectra are received as a text file from COMSOL and further
in COMSOL Multiphysics 6.0. The proposed MIM structure processing is performed on this text file. By using the
was parametrized using 10 different structural parameters, out projection, only the points on the y-axis of the absorption
of which six were variables, allowing the simulation of different spectrum are considered. For example, in Figure 2f, the point
structures as shown in Figure 1. The six variable parameters (x, y) denotes that x is the frequency and y is the absorbance
consist of four height parameters (h1, h2, h3, and h4), thickness corresponding to that frequency (there are 100 such points in
of the split-ring (t), and the split (s) as shown in Figure 1. The the plot). Therefore, only the y points are considered in the
simulation domain was defined by using different boundary data set, thus making it a 1-D array. We can obtain the plot
conditions. Periodic boundary conditions, which are used to from the array using 100 equidistant points as x-values. Figure
1680 https://doi.org/10.1021/acsaom.3c00246
ACS Appl. Opt. Mater. 2023, 1, 1679−1687
ACS Applied Optical Materials pubs.acs.org/acsaom Article

Figure 2. (a−d) Different designs of unit cells obtained after varying the structural parameters h1, h2, h3, and h4 of the proposed metamaterial
structure. (e) The flow of data generation in COMSOL. Inputs are the different metamaterial structures and outputs are the corresponding
absorption spectrum. (f) Method to convert a 2-D array into a 1-D array. (g) Conversion of absorption spectrum obtained from COMSOL, into a
1-D array, to use it in ML models. The 1-D array corresponding to the absorption spectrum and the 1-D array corresponding to structural
parameters are used as inputs and outputs for ML models, suitably.

2g shows the final deductions from the plots. The structural function that uses the forward model. In both the NNs, there is
array is a 1-D array of structural parameters, having six values an Input layer, an output layer, five hidden layers, and two
in the following order, h1, h2, h3, h4, s, and t, whereas the dropout layers in between hidden layers. The input layer has a
absorption spectrum is a 1-D array containing 100 equidistant hundred nodes, the output layer has six nodes, and the five
values. The total data set is divided into three categories: hidden layers have 250, 250, dropout (p = 0.2), 250, 128,
training data, test data, and validation data in the ratio of 8:1:1. dropout (p = 0.2), and 50 nodes, respectively. For NN with
Forward Model custom loss, the loss function has two components. To
calculate the first component, the output of the inverse model
Several machine learning techniques such as linear regression, was put into the pretrained forward model and then the mean
KNN, decision trees, and neural networks were used and square error of this output and the input of the inverse model
compared to make the forward model. In doing so, input is the was taken. The second component is the usual mean squared
structural parameter, and output is the absorption spectrum. error that in addition to the first component gives the overall
Scikit-learn was used for KNN and decision trees, while custom loss function, on which the model was trained.
TensorFlow was used for neural networks. The neural network Schematics of all of the inverse models are shown in Figure 3.
input layer has six nodes (the structural parameters), the
output layer has a hundred nodes, and the five hidden layers
have 64, 128, 512, 1024, and 512 nodes, respectively. MSE was
used to compare the performances of the models. Schematics
■ RESULTS AND DISCUSSION
MIM structures are a type of metamaterial structure consisting
of all of the forward models are shown in Figure 3. of two thin metal layers separated by a thin insulating layer.
The thickness of the metal layers and the insulating layer can
Inverse Model be tuned to create a resonant structure that strongly absorbs
Similarly, linear regression, KNN, decision trees, and neural light over a specific frequency range. MIM structures can be
networks were also implemented for the inverse models. The used to design perfect absorbers with a high absorption
inverse neural network model was of two types: one with a efficiency over a broad frequency range. The strong absorption
default loss function and the other with a customized loss arises from the coupling between the plasmonic resonances in
1681 https://doi.org/10.1021/acsaom.3c00246
ACS Appl. Opt. Mater. 2023, 1, 1679−1687
ACS Applied Optical Materials pubs.acs.org/acsaom Article

Figure 3. Architecture of all of the models. (Forward) - Forward model that predicts absorption spectrum from structures. (Inverse) - Inverse
model that predicts structures from absorption spectrum. (Forward + Inverse) - An inverse neural network model with a customized loss function.

the metal layers and the Fabry−Perot resonances in the As shown in Figures 1 and 2, there are a total of 10 structural
insulating layer. They have several advantages over other parameters that define the geometry of the unit cell: p, t1, t2, t3,
metamaterial structures, as they can be fabricated using s, t, h1, h2, h3, and h4. The first four parameters (p, t1, t2, and t3)
standard thin-film deposition techniques, making them easy are fixed while the remaining six parameters (s, t, h1, h2, h3, and
to manufacture. They can also be engineered to have thin and h4) are variables to generate different types of patterns from
lightweight structures, which is advantageous in applications this structure for creating different simulation geometries and
where weight and size constraints are critical, such as aerospace store this data for the training of the ML model. Herein, p is
or portable electronics. the periodicity of the unit cell, which is fixed at 60 μm. t1, t2,
The overall goal is to develop a methodology to design and t3 are the thicknesses of the top, middle, and bottom
metamaterials with desired electromagnetic response and vice layers, and their values are fixed at 4, 10, and 4 μm,
versa. The electromagnetic property that is of interest is respectively. s is the length of the split in the ring made by the
absorbance. Since the MIM structures are suitable for top layer, and its value lies between 8 and 12 μm. t is the
designing perfect absorber metamaterials, the MIM structure thickness of the ring, and its value lies between 3 and 8 μm. h1,
was chosen as the proposed metamaterial structure. The unit h2, h3, and h4 are the parameters that define the shape of the
cell of our proposed structure of the metamaterial is shown in ring and they vary in the direction of the y-axis. The values of
Figure 1. It is a three-layered structure, having metal as the top h1 lie between 0 and 25 μm, h2 between 5 and 25 μm, h3
layer, insulator as the middle layer, and metal as the bottom between 5 and 25 μm, and h4 between 0 and 25 μm. There are
layer, thus making it an MIM structure. In the MIM structure, also some constraints while producing different geometries
the metallic layers are made up of gold (Au) and the with the range of values, that is, if a set of values for h1, h2, h3,
nonmetallic layer is silicon dioxide (SiO2) as shown in Figure and h4 is assigned to a structure (X, Y, Z, W μm) and the same
1a. The top metallic layer is patterned as a split-ring structure. values are assigned in the reverse order (W, Z, Y, X μm) in
This split-ring structure creates a resonant circuit that behaves another structure, then both structures will be symmetrically
in a specific way based on the size and shape of the ring, as well related and give the same absorption spectrum while
as the frequency of the incoming wave, which gives unique simulating in COMSOL, owing to the rotational symmetry,
properties to metamaterials. They are widely used in the field and are considered identical. To ensure that each of these
of metamaterials to engineer materials with unique and useful structures is unique, different values of the parameters were
properties. The materials of each layer are shown in Figure 1. chosen such that they do not become identical while reversing
1682 https://doi.org/10.1021/acsaom.3c00246
ACS Appl. Opt. Mater. 2023, 1, 1679−1687
ACS Applied Optical Materials pubs.acs.org/acsaom Article

the order. All of the structural parameters are summarized in can be seen that the linear regression model shows the worst
Table 1. Representative patterns are shown in Figure 2a−d as a performance with an MSE of 0.035 while the neural network
model shows the best performance with an MSE of 0.0028.
Table 1. Dimensions of All of the Structural Parameters Similarly, all of these models in addition to the NN with a
custom loss function were implemented and tested for the
parameter dimension (μm)
inverse model. The inverse model aims to predict the structural
t1 4 parameters of a metamaterial structure with a desired property.
t2 10 The inverse model takes a specific absorption spectrum as
t3 4
input and outputs the resulting structural parameters. Figure 4c
p 60
presents the MSE of all of the inverse models. It can be
h1 (0−25)
concluded that to predict the structural parameters of a
h2 (5−25)
metamaterial for a custom absorption spectrum, the neural
h3 (5−25)
network model with default loss function will be used, which
h4 (0−25)
has the lowest MSE of 0.023. Interestingly, the inverse model
s (8−12)
with the custom loss function has an MSE of 0.036, higher
t (3−8)
than that of the model with the default loss function, which
may be due to the error that is present in the pretrained
result of modifying only h1, h2, h3, and h4, other parameters
forward model that may have translated further while training.
being constant. By adjusting the parameters, as demonstrated
The forward model’s output serves as an input to the inverse
by these examples, we can produce a wide range of patterns. In
model, influencing the custom loss computation and,
terms of symmetry, there are both symmetrical and nonsym-
consequently, the backpropagation of gradients through both
metrical structures among the four images, including square-
models. When the pretrained forward model’s predictions are
like, triangle-like, hexagon-like, and skewed structures.
Since the unit cell’s periodicity is 60 μm, in order for the used as inputs to the inverse model, errors in these predictions
length of the structure to be compatible with the unit cell, it can have an amplified effect during backpropagation. This may
has to be smaller than 60 μm. Therefore, the maximum length be due to the nature of how gradients flow backward through
was taken to be 50 μm with h1 ranging from 0 to 25 μm. The the network layers during training. If the forward model makes
same process was followed for h4. Further, to realize the split- some errors in its predictions, these errors get propagated
ring structure, the range of h2 and h3 was taken from 5 to 25 through the inverse model’s layers as gradients, and they can
μm. The thickness, kept under 10 μm, ensures compatibility become magnified.
within the unit cell dimensions. In order to confirm that the model is not biased toward any
After generating the training data, the preprocessed data was particular variable, a correlation heatmap between input
given as input to the forward model. The main objective of the (structural parameters) and output parameters (averaged
forward model is to predict the absorption spectrum from the absorption values for every 10 points) was calculated and is
structural parameters using a machine learning architecture. shown in Figure 5. The correlation coefficient measures the
For making the forward model, several machine learning strength and direction of a linear relationship between two
techniques such as linear regression, KNN, decision trees, and numerical variables. Several libraries are used to compute
neural networks were used and compared. In doing so, the these. In our case, Seaborn’s heatmap in combination with
input is an array of structural parameters, and the output is an Pandas was used.33 The most commonly used correlation
array of absorption spectrum. Epoch vs loss plot (for the NN coefficient is the Pearson correlation coefficient, which ranges
model), shown in Figure 4a, confirms the consistent from −1 (perfect negative correlation, which means that if one
convergence of the loss of test data and the convergence of variable increases, then the other will decrease, and the extent
loss on validation data shows that there is no overfitting in the of change will depend upon the correlation coefficient) to 1
model. Figure 4b summarizes the mean squared error (MSE) (perfect positive correlation, which means that if one variable
of all of the forward models. By comparing the MSE values, it increases, then the other will also increase), with 0 indicating

Figure 4. (a) Epoch vs loss plot for forward neural network model, inverse neural network model with default loss function, and inverse neural
network model with custom loss function. (b) Bar plot summarizing mean squared errors of all of the forward models. (c) Bar plot summarizing
mean squared errors of all of the inverse models.

1683 https://doi.org/10.1021/acsaom.3c00246
ACS Appl. Opt. Mater. 2023, 1, 1679−1687
ACS Applied Optical Materials pubs.acs.org/acsaom Article

no linear correlation. In general, values greater than 0.70

indicate a very strong correlation between the two values,
indicating that the parameters are highly dependent on each
other, which may lead to inconsistency that can impact the
performance of our model. However, in our case, the highest
correlation value is approximately 0.45, which is a weak
correlation indicating that none of the output features depend
significantly on the input features, indicating that our model is
not skewed toward a particular input variable.
Some of the sample predictions of all of the forward models
are depicted in Figure 6. There are four example plots for each
model where the blue plot is original and the red is predicted.
It can be clearly seen that the prediction is not good in the case
of linear regression, evident by the poor overlapping between
the red and blue plots, but the overlap increases subsequently
for KNN and DT, and best for NN. The predictions of the
inverse models are depicted in Figure 7. Similarly, the overlap
for linear regression is the least and highest for NN (default
Figure 5. Heatmap of the correlation between input and output loss), showing that the NN with the default loss function
parameters. The vertical axis represents the input structural
performs the best among all of the inverse models.
parameters and the horizontal axis represents 10 output values or
predicted values, each of them is the average of 10 output points (out Metamaterial design problems often involve highly nonlinear
of 100) as in the forward model. relationships between the design parameters and the desired
material properties. Linear regression assumes that the input
variables and output variables have a linear relationship. It
seeks to identify the optimal line that minimizes residual errors.

Figure 6. Comparison between the predictions of all of the forward models (linear regression, decision tree regression, KNN regression, and neural
network). Comparison is done using four sample examples for each model.

1684 https://doi.org/10.1021/acsaom.3c00246
ACS Appl. Opt. Mater. 2023, 1, 1679−1687
ACS Applied Optical Materials pubs.acs.org/acsaom Article

Figure 7. Comparison between the predictions of all of the inverse models (linear regression, decision tree regression, KNN regression, neural
network with default loss function, and neural network with custom loss function). Comparison is done using two sample examples for each model.
For each sample, there is also a plot that validates the prediction. It is done by comparing the input absorption spectrum with the spectrum that is
obtained after COMSOL simulation of the predicted structural parameters.

While linear regression can accommodate basic linear relation- the previous layer and passing them through nonlinear
ships, it has difficulty capturing complex nonlinear mappings. It activation functions. The hidden layers in neural networks
is restricted to representing relationships as straight lines or allow for the composition of multiple nonlinear functions,
hyperplanes. KNN is a nonparametric that makes predictions
enabling them to model highly complex and nonlinear
based on the similarity. It can capture nonlinear relationships
by considering the neighbors with similar input features. While relationships. With appropriate architecture design, neural
KNN can capture nonlinear relationships, it relies heavily on networks can approximate any continuous function to arbitrary
the local structure of the data and may not generalize well in precision, making them powerful tools for capturing nonlinear
high-dimensional spaces or when dealing with noisy data. By mappings. Metamaterial design also involves a large number of
recursively partitioning the feature space, decision trees can design parameters, leading to large-dimensional input spaces.
identify nonlinear relationships between input and output
Neural networks are well suited to handle high-dimensional
variables. The branches of a decision tree may be determined
by nonlinear decision boundaries. Nevertheless, the nonlinear data and can efficiently learn representations that capture the
modeling capability of decision trees remains somewhat underlying structure. Hence, NNs are better in the design
limited in comparison to more complex models. Decision problem of metamaterials in comparison to other mentioned
trees approximate nonlinear relationships through a series of models, and it can also be seen from our results. Also, the
binary divisions, which may not capture intricate or complex experimental simulations conducted using COMSOL Multi-
nonlinear patterns, as well as other methods. In comparison to
physics took a considerable amount of time for each run.
other ML models, neural networks excel in capturing complex
nonlinear relationships due to their layered structure and However, by harnessing a trained machine learning model for
activation functions. They can learn intricate patterns and the same simulations, a remarkable reduction of over 99% in
nonlinear mappings by combining the weighted inputs from the computation time was achieved.
1685 https://doi.org/10.1021/acsaom.3c00246
ACS Appl. Opt. Mater. 2023, 1, 1679−1687
ACS Applied Optical Materials pubs.acs.org/acsaom Article

■ CONCLUSIONS
In conclusion, a metal−insulator−metal (MIM) design for a
(7) Han, C.; Zhang, B.; Wang, H.; Ding, J. Metamaterial perfect
absorber with morphology-engineered meta-atoms using deep
learning. Opt. Express 2021, 29 (13), 19955−19963.
Metamaterial absorber (MMA) in the terahertz regime is (8) Fang, N.; Lee, H.; Sun, C.; Zhang, X. Sub-diffraction-limited
proposed, and numerical simulations are performed on it to optical imaging with a silver superlens. Science 2005, 308 (5721),
determine its optical properties and create a database for 534−537.
training the ML models. Among ML models, KNN, linear (9) Pendry, J. B.; Smith, D. R. THE QUEST for the superlens. Sci.
regression, decision trees, and NN were implemented for both Am. 2006, 295 (1), 60−67.
forward and reverse models, and the forward and reverse (10) Withayachumnankul, W.; Abbott, D. Metamaterials in the
models with the best performance were proposed for designing terahertz regime. IEEE Photonics J. 2009, 1, 99−118.
multiband terahertz metamaterial absorber. For both the (11) Roh, Y.; Lee, S. H.; Kwak, J.; Song, H. S.; Shin, S.; Kim, Y. K.;
forward and reverse models, the neural network model with Kang, B.; Ju, B. K.; Wu, J. W.; Seo, M. Terahertz imaging with
the default loss function performed the best with an MSE of metamaterials for biological applications. Sens. Actuators, B 2022, 352,
0.0028 and 0.023, respectively. The application of machine 130993.
(12) Liu, X.; Chen, W.; Ma, Y.; Xie, Y.; Zhou, J.; Zhu, L.; Xu, Y.;
learning strategies for the purpose of generating forward and Zhu, J. Enhancing THz fingerprint detection on the planar surface of
inverse models of metamaterials for predicting the absorption an inverted dielectric metagrating. Photonics Res. 2022, 10 (12),
and metamaterial’s structural parameters can drastically reduce 2836−2845.
the computational cost and enable an accelerated process for (13) Xie, Y.; Ma, Y.; Liu, X.; Khan, S. A.; Chen, W.; Zhu, L.; Zhu, J.;
the design and optimization of a variety of devices, including Liu, Q. H. Dual-Degree-of-Freedom Multiplexed Metasensor Based
antennas, sensors, and optical systems. The application of ML on Quasi-BICs for Boosting Broadband Trace Isomer Detection by
in designing metamaterials can be extended to other domains THz Molecular Fingerprint. IEEE J. Sel. Top. Quantum Electron. 2023,
of materials discovery in applications such as drug delivery, 29 (5), 1−10.
energy storage, and catalysis. (14) Zhu, J.; Jiang, S.; Xie, Y.; Li, F.; Du, L.; Meng, K.; Zhu, L.;
Zhou, J. Enhancing terahertz molecular fingerprint detection by a
■ AUTHOR INFORMATION
Corresponding Author
dielectric metagrating. Opt. Lett. 2020, 45 (8), 2335−2338.
(15) Misra, S.; Wang, H. Review on the growth, properties and
applications of self-assembled oxide-metal vertically aligned nano-
Shikhar Misra − Indian Institute of Technology Kanpur, composite thin films-current and future perspectives. Mater. Horiz.
Kanpur 208016 Uttar Pradesh, India; orcid.org/0000- 2021, 8 (3), 869−884.
0001-8257-2878; Email: [email protected] (16) Zhang, D.; Misra, S.; Li, L.; Wang, X.; Jian, J.; Lu, P.; Gao, X.;
Sun, X.; Qi, Z.; Kalaswad, M.; Zhang, X.; Wang, H. Tunable Optical
Author Properties in Self-Assembled Oxide-Metal Hybrid Thin Films via Au-
Phase Geometry Control: From Nanopillars to Nanodisks. Adv. Opt.
Mukund Soni − Indian Institute of Technology Kanpur, Mater. 2020, 8 (4), No. 1901359.
Kanpur 208016 Uttar Pradesh, India (17) Huang, J.; Jin, T.; Misra, S.; Wang, H.; Qi, Z.; Dai, Y.; Sun, X.;
Complete contact information is available at: Li, L.; Okkema, J.; Chen, H. T.; Lin, P. T.; Zhang, X.; Wang, H.
https://pubs.acs.org/10.1021/acsaom.3c00246 Tailorable Optical Response of Au−LiNbO3 Hybrid Metamaterial
Thin Films for Optical Waveguide Applications. Adv. Opt. Mater.
Notes 2018, 6 (19), No. 1800510.
(18) Misra, S.; Li, L.; Zhang, D.; Jian, J.; Qi, Z.; Fan, M.; Chen, H.
The authors declare no competing financial interest. T.; Zhang, X.; Wang, H. Self-Assembled Ordered Three-Phase Au−
BaTiO 3 − ZnO Vertically Aligned Nanocomposites Achieved by a
■ ACKNOWLEDGMENTS
This work was partially supported by the IITK Start-up Fund
Templating Method. Adv. Mater. 2019, 31 (7), No. 1806529.
(19) Misra, S.; Li, L.; Jian, J.; Huang, J.; Wang, X.; Zemlyanov, D.;
Jang, J. W.; Ribeiro, F. H.; Wang, H. Tailorable Au Nanoparticles
and SERB SRG/2022/000580. The authors also thank Prof. Embedded in Epitaxial TiO2 Thin Films for Tunable Optical
Rituraj from IITK for the insightful discussions related to Properties. ACS Appl. Mater. Interfaces 2018, 10 (38), 32895−32902.
numerical simulations. (20) Chen, W.; Gao, Y.; Li, Y.; Yan, Y.; Ou, J. Y.; Ma, W.; Zhu, J.

■ REFERENCES
(1) Dhillon, A. S.; Mittal, D.; Bargota, R. Triple band ultrathin
Broadband Solar Metamaterial Absorbers Empowered by Trans-
former-Based Deep Learning. Adv. Sci. 2023, 10 (13), No. 2206718.
(21) Xiong, J.; Shen, J.; Gao, Y.; Chen, Y.; Ou, J. Y.; Liu, Q. H.; Zhu,
polarization insensitive metamaterial absorber for defense, explosive J. Real-Time On-Demand Design of Circuit-Analog Plasmonic Stack
detection and airborne radar applications. Microwave Opt. Technol. Metamaterials by Divide-and-Conquer Deep Learning. Laser Photonics
Lett. 2019, 61, 89−95. Rev. 2023, 17 (3), No. 2100738.
(2) Kang, L.; Lan, S.; Cui, Y.; Rodrigues, S. P.; Liu, Y.; Werner, D. (22) Challapalli, A.; Konlan, J.; Li, G. Inverse machine learning
H.; Cai, W. An Active Metamaterial Platform for Chiral Responsive discovered metamaterials with record high recovery stress. Int. J. Mech.
Optoelectronics. Adv. Mater. 2015, 27, 4377−4383. Sci. 2023, 244, No. 108029.
(3) Debus, C.; Bolivar, P. H. Frequency selective surfaces for high (23) Wu, L.; Liu, L.; Wang, Y.; Zhai, Z.; Zhuang, H.; Krishnaraju,
sensitivity terahertz sensing. Appl. Phys. Lett. 2007, 91 (18), D.; Wang, Q.; Jiang, H. A machine learning-based method to design
No. 184102. modular metamaterials. Extreme Mech Lett. 2020, 36, No. 100657.
(4) Xu, W.; Xie, L.; Ying, Y. Mechanisms and applications of (24) Wang, Y.; Zhu, Y.; Cui, Z.; Jiang, H.; Zhang, K.; Wang, X.
terahertz metamaterial sensing: A review. Nanoscale 2017, 9 (37), Ensemble learning: a bidirectional framework for designing data-
13864−13878. driven THz composite metamaterials. J. Opt. Soc. Am. B 2022, 39,
(5) Pendry, J. B.; Smith, D. R. Reversing light with negative 835−842.
refraction. Phys. Today 2004, 57 (6), 37−43. (25) Donda, K.; Zhu, Y.; Merkel, A.; Fan, S. W.; Cao, L.; Wan, S.;
(6) Pendry, J. B. Negative Refraction Makes a Perfect Lens. Phys. Assouar, B. Ultrathin acoustic absorbing metasurface based on deep
Rev. Lett. 2000, 85 (18), No. 3966. learning approach. Smart Mater. Struct. 2021, 30 (8), No. 085003.

1686 https://doi.org/10.1021/acsaom.3c00246
ACS Appl. Opt. Mater. 2023, 1, 1679−1687
ACS Applied Optical Materials pubs.acs.org/acsaom Article

(26) Lv, Z. Y.; Li, J. Q.; Hou, Z. W.; Ding, Y. S.; Xu, W. D.; Pei, Y.
M. Design method and machine learning application of acoustic
holographic computational metamaterials. Sci. China Technol. Sci.
2021, 65 (1), 238−243.
(27) Wu, R. T.; Liu, T. W.; Jahanshahi, M. R.; Semperlotti, F. Design
of one-dimensional acoustic metamaterials using machine learning
and cell concatenation. Struct. Multidiscip. Optim. 2021, 63 (5),
2399−2423.
(28) Zheng, B.; Yang, J.; Liang, B.; Cheng, J. C. Inverse design of
acoustic metamaterials based on machine learning using a Gauss-
Bayesian model. J. Appl. Phys. 2020, 128 (13), No. 134902.
(29) Chen, Y.; Zhu, J.; Xie, Y.; Feng, N.; Liu, Q. H. Smart inverse
design of graphene-based photonic metamaterials by an adaptive
artificial neural network. Nanoscale 2019, 11 (19), 9749−9755.
(30) Brunton, S. L.; Kutz, J. N. Methods for data-driven multiscale
model discovery for materials. J. Phys. Mater. 2019, 2 (4), No. 044002.
(31) Wang, C.; Tan, X. P.; Tor, S. B.; Lim, C. S. Machine learning in
additive manufacturing: State-of-the-art and perspectives. Addit.
Manuf. 2020, 36, No. 101538.
(32) Wilt, J. K.; Yang, C.; Gu, G. X. Accelerating Auxetic
Metamaterial Design with Deep Learning,. Adv. Eng. Mater. 2020,
22 (5), No. 1901266.
(33) Waskom, M. Seaborn: statistical data visualization. J. Open
Source Software 2021, 6 (60), No. 3021.

1687 https://doi.org/10.1021/acsaom.3c00246
ACS Appl. Opt. Mater. 2023, 1, 1679−1687