Optics Communications 591 (2025) 132171

Contents lists available at ScienceDirect

Optics Communications
journal homepage: [Link]/locate/optcom

Design and performance study of a full-bandwidth mid-wave infrared

metasurface absorber based on deep learning optimization
Fan Shen , Gui Chen , Mingjie Lin , Rendi Luo , Jialiang Zhang , Changshui Chen *
Guangdong Provincial Key Laboratory of Nanophotonic Functional Materials and Devices, Guangdong Basic Research Center of Excellence for Structure and Fundamental
Interactions of Matter, School of Optoelectronic Science and Engineering, South China Normal University, Guangzhou, 510006, China

A B S T R A C T

Conventional metamaterial absorbers often suffer from limited bandwidth, structural complexity, and stringent fabrication tolerances. In response to these chal­
lenges, our work introduces a metal-semiconductor-metal structured metasurface absorber optimized for the mid-wave infrared spectrum. The proposed Ti-GaAs-W
tri-layer architecture achieves full-bandwidth performance in mid-wave infrared through strategically engineered coupling between geometrically symmetric tita­
nium nanostructures and dielectric modes. Leveraging deep neural network-based inverse design, we obtained optimized parameters yielding 90.8 % average ab­
sorptivity across 2750–5450 nm, with peak efficiencies reaching 98.92 % at 4760 nm and 99.23 % at 5050 nm. Angular dependence analysis confirms the structure
maintains >80 % average absorption up to 45◦ incidence, demonstrating superior robustness compared to conventional multi-layer alternatives. This design’s
combination of spectral performance and fabrication simplicity shows significant potential for infrared detection and energy harvesting applications.

1. Introduction increases the overall structural size, potentially leading to high fabri­
cation costs. In contrast, multiple resonant modes in a unitary structure
Metamaterials represent an emerging category of artificially engi­ can also enable broadband absorption [13–15]. Nevertheless, Multiple
neered materials. These synthetic structures exhibit optical properties resonances cannot always be excited directly, which necessitates
that are unattainable in natural materials. The defining characteristic balanced considerations between manufacturing processes and struc­
includes negative refractive index [1], perfect absorption [2], and tural designs. As a result, artificial intelligence methods like deep neural
anomalous reflection [3]. Traditional electromagnetic absorbers rely on networks have gradually become a preferred tool for assisting meta­
material losses to absorb light, they often require long optical paths, surface design [16–23]. Gu et al. proposed a framework combining a
leading to bulky designs and limited flexibility. In contrast, plasmonic primary prediction network and an auxiliary prediction network,
metamaterial absorbers have impedance characteristics adjusted by achieving efficient forward prediction and inverse design, though its
artificially designed ‘meta-atoms’ structures. This enables them to ach­ point-by-point prediction incurs high computational cost [24]. Guo et al.
ieve a good impedance matching with free space [4–6]. This approach employed a recurrent prediction network and variable contribution
not only delivers superior absorption performance but also better meets analysis, incorporating binary encoding of material parameters to
the requirements for the miniaturization of optoelectronic devices. enhance model flexibility and absorption bandwidth, achieving a peak
The three-layer metamaterial absorber [7] exhibits nearly perfect absorption rate of 99.2 %; however, the approach relies heavily on large
narrowband absorption, but its resonant structure confines absorption to datasets and computational resources [25]. Wang et al. introduced a
a narrow wavelength range. To achieve a wider absorption bandwidth, residual fully connected neural network (RFC-NN), which demonstrated
various types of metamaterial absorbers have been proposed and improved accuracy and generalization in inverse design, though the
improved, including dual-band [8] and triple-band [9] metamaterial design parameters were still limited to geometric features [26]. Despite
absorbers. Additionally, the method of coupling multiple metallic res­ their differences, all these methods significantly reduce design time and
onators of different sizes has been introduced to achieve broadband enhance the precision and robustness of the mapping between structural
absorption by covering the electromagnetic spectrum with the reso­ parameters and spectral responses.
nances of the coupled components [10–12]. Although the absorption In addition, recent advances in plasmonic metamaterials have
bandwidth can be easily tuned by modifying structural parameters, highlighted the potential of topological protection to enhance device
integrating multiple resonators within a single unit cell significantly robustness. For instance, SSH-model-based graphene nanoribbons

* Corresponding author.
E-mail address: 20071053@[Link] (C. Chen).

[Link]
Received 28 April 2025; Received in revised form 25 June 2025; Accepted 25 June 2025
Available online 29 June 2025
0030-4018/© 2025 Published by Elsevier B.V.
F. Shen et al. Optics Communications 591 (2025) 132171

√̅̅̅̅ √̅̅
exhibit edge states resistant to disorders [27], while waveguide-coupled polarized waves, Z0 = μ0
and Z = denote the free space impedance
μ
systems demonstrate topological PIT with stable transparency windows
ε0 ε
√̅̅̅̅̅̅̅̅̅
under parametric perturbations [28]. Complementary to these designs, and the metamaterial impedance, n = εr⋅ μr is the effective refractive
double-layered graphene structures offer polarization-insensitive PIT index.
through geometric symmetry [29], though their bandwidth remains Due to the elimination of transmissive modes in the absorbing
limited. structure, the absorption curve simplifies to:
In recent years, driven by the demand for high-temperature detec­ ⃒ ⃒ ⃒√̅̅̅̅̅ √⃒̅̅̅̅2
⃒Z − Z0 ⃒2 ⃒ εr ⃒⃒
tion, gas analysis, and military stealth, extensive research has been A = 1 − R = 1 − ⃒⃒ ⃒ = 1 − ⃒⃒√μ r −
̅̅̅̅̅ √⃒̅̅̅̅ (6)
conducted on broadband absorber designs in the mid-wave infrared Z + Z0 ⃒ ⃒ μr + εr ⃒
(MWIR) band. Currently, only a limited number of broadband meta­
Eq. 6 demonstrates that μr = εr or Z = Z0 constitutes the condition for
surface absorbers operate effectively within the MWIR region [14,30],
perfect absorption. It is noteworthy that impedance matching in MTMA
and their bandwidth remains limited, failing to cover the entire MWIR
requires simultaneous fulfillment of both electric and magnetic reso­
spectrum.
nance conditions.
To address this challenge, we present a deep neural network-assisted
As a result, a metal-semiconductor-metal (MSM) type mid-wave
design for a full-bandwidth MWIR broadband absorber. We employ a
infrared super-surface absorber has been proposed in our paper, and
deep neural network to optimize the metasurface absorber structure,
its periodic repeating structure is presented in Fig. 1(a). The unit cell
achieving the desired absorption performance. We configure the struc­
structure of the metasurface absorber is presented in Fig. 1(b), consisting
tural parameters of the metasurface absorber as input data, while uti­
of four titanium (Ti) cylinders symmetrically distributed around the cell
lizing the corresponding optical absorptance as output data. The
center, gallium arsenide (GaAs), and tungsten (W). The five parameters
absorber’s ideal structural characteristic parameters were determined
D, P, H, G, and L determine the structural parameters of the unit cell. We
through inverse prediction. The average absorption of the absorber in
selected them as input data for training a fully connected deep neural
the wavelength range of 2750–5450 nm is more than 90 %, with peak
network in Python, as shown in Fig. 2. The network adopts a five-layer
absorption reaching 98.92 % at 4760 nm and 99.23 % at 5050 nm. Even
fully connected structure (200-800-800-800-300 neurons), including
at a 45◦ oblique incidence, it maintains an average absorption rate above
batch normalization, LeakyReLU activation and Dropout regularization.
80 %. This absorber offers advantages such as full-band MWIR absorp­
The output layer uses the Sigmoid function to confine the predicted
tion, high absorption efficiency, relatively simple structural design, and
values within the interval of [0,1]. The training adopted the MSE loss
insensitivity to incident angle variations. These characteristics indicate
function and Adam optimizer, combined with cosine annealing learning
strong potential for deployment in thermal imaging, high-temperature
rate scheduling. 9600 iterations were carried out on 12,000 samples (80
detection, gas sensing, and chemical analysis.
% of the training set), with a batch size of 200. The code implements
functions such as automatic GPU-accelerated training, periodic test
2. Principles and design
verification, and visualization of the training process. It can efficiently
complete the forward prediction from structural parameters to spectral
Within the impedance matching framework, the metamaterial’s ab­
characteristics and provides a basic framework for inverse design. Using
sorption characteristics become theoretically predictable [31]. In the
the input structural parameters, the deep neural network predicted the
impedance matching theory, a metamaterial absorber (MTMA) perfor­
MWIR light absorption rate of the MSM MWIR metasurface absorber at
mance is controlled by tailoring the material’s electric and magnetic
wavelengths ranging from 2500 to 5500 nm. In the inverse design, the
response characteristics in order to achieve impedance matching with
MWIR absorptance spectrum of the MSM metasurface absorber serves as
free space. [32]. Under these conditions, no reflection occurs at the
input data, while five structural parameters are defined as output vari­
interface, allowing all incident waves to be absorbed by the MTMA.
ables. Through inverse prediction, we ultimately obtain the unit cell’s
MTMA is typically a sandwich structure, with one side supported by
characteristic parameters that achieve the target mid-wave infrared
a metal ground plane substrate and the other side composed of a metal
absorption performance.
pattern. The magnetic permeability and electric permittivity of the
The optimized parameters D, P, H, G, and L predicted by the deep
material are given by:
neural network for an average absorbance of 90.8 % in the 2750–5500
μ = μ0⋅ μr (ω), (1) nm range are 190 nm, 435 nm, 245 nm, 230 nm, and 125 nm. The finite-
difference time-domain (FDTD) method was used to calculate the light
ε = ε0⋅ εr (ω), (2) absorption rate of the unit cell. In the calculation of optical absorption
using the FDTD method, we set the boundary condition along the z-axis
where μ0 and ε0 are the permeability and permittivity of the free space, as the PML, while applying periodic boundary conditions along both the
μr (ω) and εr (ω) are the frequency dependent permeability and permit­ x-axes and y-axes. The fundamental expression for optical absorptance is
tivity of the medium. As a result of the presence of the ground plane, given by A(ω) = 1− R(ω)− T(ω). With the implementation of a 200-nm-
transmission is eliminated on the bottom boundary of the metamaterial.
Consequently, the reflectance (R) of the MTMA can be determined using
the Fresnel reflection equation [7]:
⃒ ⃒̅2
⃒ε cos θ − √̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅
n2 − sin θ⃒⃒
⃒ r
RTM = |rTM| = ⃒ √⃒̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅ , (3)
⃒εr cos θ + n2 − sin θ⃒

⃒ ⃒̅2
⃒μ cos θ − √̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅
n2 − sin θ⃒⃒
⃒
RTE = |rTE| = ⃒ r √⃒̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅ , (4)
⃒μr cos θ + n2 − sin θ⃒

⃒ ⃒ ⃒ ̅ √⃒̅̅̅̅2
⃒Z − Z0 ⃒2 ⃒⃒√̅̅̅̅ εr ⃒⃒
R = ⃒⃒ ⃒ = ⃒√μ r −
̅̅̅̅̅ √⃒̅̅̅̅ , (5)
Z + Z0 ⃒ ⃒ μr + εr ⃒
Fig. 1. (a) Schematic diagram of the designed MWIR metasurface absorber
where TM and TE are transverse magnetic and transverse electric with periodic repetition; (b) Three-dimensional structure diagram of the ab­
sorber’s unit cell; (c) Front view of the absorber.

2
F. Shen et al. Optics Communications 591 (2025) 132171

Fig. 2. Overview of the forward prediction and inverse design process by the deep neural network. Five parameters (D, P, H, G and L) for the unit cell of meta­
material absorber are set as the training data. The input and output layers interchange for the forward prediction and the inverse design.

thick W layer at the absorber’s substrate, the transmittance T(ω) be­ distinctive absorption peaks reaching 98.92 % at 4760 nm and 99.23 %
comes negligible (T(ω)≈0) across the 2500–5500 nm wavelength at 5050 nm, exhibiting outstanding absorption characteristics.
regime. Consequently, the absorptance formulation can be reduced to To compare the absorption characteristics of different designs, Fig. 4
Eq. (6). (b) presents the absorption spectra of both the metasurface structure
In this paper, we chose Ti as the material for the top metal layer of composed of Ti, GaAs, and W substrates and the bare Ti-GaAs-W thin
the absorber due to its high loss characteristics in the infrared region, film (Fig. 4(a)). Notably, the unpatterned Ti-GaAs-W film exhibits a
which can be used to generate surface oscillating currents [33]. GaAs maximum absorptance of less than 45 % across the entire 2500–5500
was selected as the material for constructing the dielectric layer of the nm wavelength range, significantly lower than metasurface structure
absorber due to its excellent optoelectronic properties and compatibility incorporating Ti, GaAs, and W substrates.
with large-scale manufacturing [34–36]. From visible to terahertz In the optimization of metasurface absorbers configurations,the
wavelengths, GaAs micro-nano structures can be considered as dielectric layer constitutes a critical component in addition to the top
high-refractive-index dielectric resonators supporting various plasmonic metal layer. Under identical structural parameters for the absorber
resonance modes [37–39]. configuration, we compared the absorption spectra of two conventional
dielectric materials (Al2O3 and SiO2) against GaAs, as presented in Fig. 5
3. Data analysis and discussion (a). Throughout the 2500–5500 nm spectral range, the GaAs dielectric
layer exhibits significantly enhanced performance relative to both SiO2
Fig. 3(a) shows the absorption spectrum obtained when plane-wave and Al2O3. The overall optical performance of SiO2 and Al2O3 dielectric
light with wavelengths of 2500–5500 nm is normally incident on the layers proves inferior to that of the GaAs layer. Consequently, GaAs was
absorber. Fig. 3(b) and (c) show the cross-sectional electric field and selected as the optimal dielectric layer for our absorber design.
magnetic field profiles along the cylinder’s central axis. The spectrum Furthermore, after determining the dielectric layer material, we
demonstrates that the absorber achieves an average absorptance conducted a comparative analysis of the absorption spectra among three
exceeding 90 % within the 2750–5450 nm wavelength range, with commonly used top metal materials (Ni, Cr, and Ti) for infrared

Fig. 3. (a) Simulated absorption spectrum of the absorber under normal incidence in the wavelength range of 2500–5500 nm; (b) Electric field distribution in the
vertical cross-section through the cylinder center; (c) Magnetic field distribution in the vertical cross-section through the cylinder center.

3
F. Shen et al. Optics Communications 591 (2025) 132171

Fig. 4. (a) the structure of bare Ti-GaAs-W thin film; (b) Comparison of absorption spectra between the metasurface structure and the unpatterned Ti-GaAs-W film.

Fig. 5. (a) Influence of different dielectric materials on the light absorption efficiency of the absorber; (b) Influence of different top metal layer materials on the light
absorption efficiency of the absorber.

absorption structures, as shown in Fig. 5(b). In the spectral range of Therefore, we further studied the effect of incident light angle on the
2500–5500 nm, Ti demonstrates significantly enhanced performance absorption spectrum. Fig. 6 presents the absorption spectra under obli­
compared to both Ni and Cr. The overall optical performance of Ni and que incidence of plane-wave illumination, with incident angles
Cr was found to be inferior to that of Ti. Consequently, Ti was selected as increasing from 0◦ to 45◦ in steps of 15◦ . Across the 0–45◦ incidence
the optimal top metal material for our absorber design. angle range, the average absorptance of the spectrum shows minimal
In practical applications of metasurface absorbers, the impact of variation. At a 45◦ incident angle, the absorber’s performance degrades
multi-angle incidence on absorption performance must be considered. noticeably, particularly in the 4250–5500 nm range, yet its average
absorption rate stays above 80 %. As shown in Fig. 6, although the ab­
sorption rate decreases slightly, the resonance frequency remains
essentially unchanged. This phenomenon originates from the interaction
mechanism between incident light and resonant modes: their magnetic
coupling efficiency is directly governed by the in-plane projection
component of the incident magnetic field. As the incidence angle θ in­
creases, this horizontal magnetic field component decays following a
cosθ relationship, leading to synchronous reduction in both mode exci­
tation efficiency and absorptance. At the same time, as the incident
angle changes, mode coupling also causes a blue shift in the resonance
wavelength, with the absorption peaks shifting toward shorter
wavelengths.
Another important challenge in the design of metasurface absorbers
is the low sensitivity of the absorption spectrum to inevitable deviations
in design parameters. These deviations may occur in micro-nano fabri­
cation processes including thin film deposition, lithography, and
etching, where the film thickness may vary with deposition and etching
rates, and the structural dimensions of the metamaterial surface may
deviate from the design values due to defects generated during lithog­
raphy. Therefore, the change in the absorption rate curve should be
checked according to possible variations in design parameters. Fig. 7
illustrates the effect of changes in some important structural parameters
Fig. 6. Influence of incident angle on the light absorption efficiency of on the absorption performance of the metasurface absorber, where Fig. 7
the absorber. (a)–(e) show the absorption curves under different metal cylinder radii

4
F. Shen et al. Optics Communications 591 (2025) 132171

Fig. 7. Effect of structural parameters on absorption performance: (a) Metal cylinder radius (D); (b) Distance between metal cylinder centers (G); (c) Unit cell side
length (P); (d) Metal cylinder thickness (H); (e) Dielectric layer thickness (L).

(D), distances between metal cylinder centers (G), unit cell side lengths the shorter-wavelength regime. The dramatic spectral transition occurs
(P), top metal cylinder thicknesses (H), and dielectric layer thicknesses with all other parameters held constant when the metallic nanocylinder
(L). radius reaches 200 nm, characterized by significantly reduced absorp­
Fig. 7(a) and (c) demonstrate that variations in the metallic cylinder tion bandwidth and decreased average absorptance. Notably, identical
radius and unit cell size within a certain range have negligible impact on spectral features emerge in Fig. 7(b) and (c) at G = 240 nm and P = 425
the resonance peak at shorter wavelengths. However, the resonance nm, where the absorption profiles closely resemble that of the bare Ti-
peak at the longer wavelengths undergoes a redshift with either GaAs-W thin film in Fig. 4(b). This correlation suggests the phenome­
increasing metallic cylinder radius or decreasing unit cell size. This can non relates to the nanocylinder-to-boundary distance, with our mea­
be explained by the origin of the resonance peaks. The shorter- surements confirming that absorption spectra adopt this characteristic
wavelength resonance peak mainly originates from the plasmonic sur­ shape when the cylinder-boundary separation falls below 2.5 nm.
face polariton (PSP) mode generated at the metal-dielectric interface at Fig. 7(b) and (d) demonstrate that the inter-cylinder spacing within a
the bottom, which is less affected by the top metal radius and unit cell certain range exhibits negligible influence on the absorber’s perfor­
size. Therefore, changing the top metal radius and unit cell size has little mance, whereas the cylinder height significantly governs its absorption
effect on the position of the shorter-wavelength resonance peak. How­ characteristics. The absorptance can be calculated using Eq. (6), where
ever, the longer-wavelength resonance peak mainly originates from the the impedance Z can be described by Ref. [40]:
cavity mode, which is generated by phase-coherent superposition of √̅̅̅̅̅̅
ωμ 1
multiply reflected waves between the top and bottom metal surfaces. Z ≅ (1 + j) = (1 + j) , (7)
Thus, the metal cylinder radius and unit cell size have a significant effect 2σ σ δs
on the cavity mode. As the metal cylinder radius increases and the unit
where μ, σ and ω represent the magnetic permeability, electrical con­
cell size decreases, the MSM resonator tends to form a Ti-GaAs-W film
ductivity and angular frequency of the metasurface, and δs represents
sandwich structure. This structural evolution leads to progressively
the electromagnetic penetration depth in metallic media. Eq. (7)
broadened absorption bandwidth but reduced average absorptance in
demonstrate that the absorption performance is primarily determined

5
F. Shen et al. Optics Communications 591 (2025) 132171

by the electromagnetic penetration depth in metallic media, while Table 1

exhibiting negligible dependence on the inter-cylinder spacing of the Summary of some reported absorbers.
metasurface. The system demonstrates sensitivity exclusively to varia­ Reference Layer Absorption bandwidth Average absorptivity
tions in the top metal layer thickness.
Our work 3 2.75–5 μm 90.8 %
Fig. 7(e) demonstrates that the absorption efficiency is found to be [42] 6 0.3–3 μm >90 %
critically governed by the dielectric spacer thickness. Interference the­ [43] 8 0.28–4.5 μm ~94.5 %
ory can well explain this phenomenon. The dielectric layer thickness [25] 9 0.2–4.2 μm ~90 %
determines the effective cavity length, serving as a critical factor gov­
erning both the peak absorptance and reflection coefficient of the
– review & editing, Supervision.
absorber. If the dielectric layer thickness for maximum absorption at a
specific frequency is d, according to interference theory [41]:
( )⎞ Declaration of competing interest
⎛
nr n
a− b+exp 2π m nri
The authors declare that they have no known competing financial
ni
⎜ ⎟
arctan⎝ b+nnr a ⎠
i interests or personal relationships that could have appeared to influence
mπ
d= + , (8) the work reported in this paper.
2nr k nr k
Data availability
where a and b represent the real and imaginary parts of the reflection
coefficient, nr and ni represent the real and imaginary parts of the
Data will be made available on request.
refractive index of the dielectric layer, and m is an integer greater than
or equal to zero. With increasing dielectric layer thickness, higher-order
References
modes (m = 1,2,3 …) are successively excited, enhancing optical ab­
sorption. In dielectric materials, ni is usually much smaller than nr, so [1] S. Anantha Ramakrishna, Physics of negative refractive index materials, Rep. Prog.
Eq. (8) can be simplified to: Phys. 68 (2005) 449.
( ) [2] Y. Zhang, P. Wu, Z. Zhou, X. Chen, Z. Yi, J. Zhu, T. Zhang, H. Jile, Study on
temperature adjustable terahertz metamaterial absorber based on vanadium
arctan − ab
mπ dioxide, IEEE Access 8 (2020) 85154–85161.
d≈ + . (9) [3] S. Sun, K.-Y. Yang, C.-M. Wang, T.-K. Juan, W.T. Chen, C.Y. Liao, Q. He, S. Xiao,
2nr k nr k W.-T. Kung, G.-Y. Guo, L. Zhou, D.P. Tsai, High-efficiency broadband anomalous
reflection by gradient meta-surfaces, Nano Lett. 12 (2012) 6223–6229.
For the lowest-order (m = 0) absorption band, the thickness of the
[4] X. Fang, K.F. MacDonald, N.I. Zheludev, Controlling light with light using coherent
dielectric layer is directly related to the phase of the reflection coeffi­ metadevices: all-optical transistor, summator and invertor, Light Sci. Appl. 4
( )
(2015) e292, e292.
cient (arctan − ab ) and the wavenumber (k). When m > 0, the absorp­ [5] C. Tian, Q. Tu, W. Liu, J. Wang, Recent advances in microfluidic technologies for
organ-on-a-chip, TrAC, Trends Anal. Chem. 117 (2019) 146–156.
tion peaks appear at different values of m as the thickness of the [6] J. Grant, Y. Ma, S. Saha, A. Khalid, D.R.S. Cumming, Polarization insensitive,
dielectric layer increases. According to Eq. (9), this occurs because a 2π broadband terahertz metamaterial absorber, Opt. Lett. 36 (2011) 3476–3478.
phase delay develops across the entire dielectric layer. Consequently, [7] N.I. Landy, S. Sajuyigbe, J.J. Mock, D.R. Smith, W.J. Padilla, Perfect metamaterial
absorber, Phys. Rev. Lett. 100 (2008) 207402.
partial absorption peaks exhibit slight redshifts as the dielectric layer [8] Z. Liu, L. Guo, Q. Zhang, Design of dual-band terahertz perfect metamaterial
thickness increases. absorber based on circuit theory, Molecules (2020).
In a word, the absorber demonstrates excellent broadband absorp­ [9] X. Shen, Y. Yang, Y. Zang, J. Gu, J. Han, W. Zhang, T. Jun Cui, Triple-band
terahertz metamaterial absorber: design, experiment, and physical interpretation,
tion performance. Table 1 presents a comparison between this design Appl. Phys. Lett. 101 (2012) 154102.
and several previously reported absorbers. It is evident that achieving a [10] F. Ding, Y. Cui, X. Ge, Y. Jin, S. He, Ultra-broadband microwave metamaterial
similar level of absorption in other works typically requires more absorber, Appl. Phys. Lett. 100 (2012) 103506.
[11] H. Cai, Y. Sun, X. Wang, S. Zhan, Design of an ultra-broadband near-perfect bilayer
structural layers, whereas the absorber in this study benefits from its grating metamaterial absorber based on genetic algorithm, Opt. Express 28 (2020)
fewer layers, exhibiting enhanced robustness and angular tolerance. 15347–15359.
[12] P. Yu, H. Yang, X. Chen, Z. Yi, W. Yao, J. Chen, Y. Yi, P. Wu, Ultra-wideband solar
absorber based on refractory titanium metal, Renew. Energy 158 (2020) 227–235.
4. Conclusion [13] Y. Luo, L. Huang, J. Ding, W. Liu, B. Sun, C. Xie, H. Yang, J. Wu, Flexible and
transparent broadband microwave metasurface absorber based on multipolar
In summary, this paper proposes a wideband, wide-angle mid- interference engineering, Opt. Express 30 (2022) 7694–7707.
[14] X. shi, E. Hou, Z. Liang, S. Zhang, R. Dai, W. Xin, D. Meng, H. Liu, H. Xu, Y. Liu,
infrared metasurface absorber based on a metal-semiconductor-metal
Broadband metamaterial absorber based on hybrid multi-mode resonance in mid-
(MSM) configuration, with optimal parameters determined via deep wave and long-wave infrared region, Results Phys. 42 (2022) 105972.
learning algorithms. The absorber demonstrates an average absorption [15] Q. Feng, M. Pu, C. Hu, X. Luo, Engineering the dispersion of metamaterial surface
exceeding 90 % across the 2750–5450 nm spectral range, achieving peak for broadband infrared absorption, Opt. Lett. 37 (2012) 2133–2135.
[16] G. Lan, Y. Wang, J.-Y. Ou, Optimization of metamaterials and metamaterial-
absorption rates of 98.92 % at 4760 nm and 99.23 % at 5050 nm. microcavity based on deep neural networks, Nanoscale Adv. 4 (2022) 5137–5143.
Remarkably, it maintains over 80 % average absorption within the same [17] Z. Ding, W. Su, Y. Luo, L. Ye, W. Li, Y. Zhou, B. Tang, H. Yao, Artificial neural
bandwidth even at 45◦ oblique incidence. This work bridges a critical network-based inverse design of metasurface absorber with tunable absorption
window, Mater. Des. 234 (2023) 112331.
gap in full-bandwidth metasurface absorbers for mid-infrared applica­ [18] S.K. Patel, J. Surve, R. Jadeja, V. Katkar, J. Parmar, K. Ahmed, Ultra-wideband,
tions, offering promising potential in thermal imaging, solar energy polarization-independent, wide-angle multilayer swastika-shaped metamaterial
harvesting, gas detection, and chemical sensing. solar energy absorber with absorption prediction using machine learning,
Advanced Theory and Simulations 5 (2022) 2100604.
[19] C. Ma, Y. Chang, S. Wu, R.R. Zhao, Deep learning-accelerated designs of tunable
CRediT authorship contribution statement magneto-mechanical metamaterials, ACS Appl. Mater. Interfaces 14 (2022)
33892–33902.
[20] S. Zhou, Y. Gai, G. Lan, Optimization of tunable terahertz metamaterial absorber
Fan Shen: Writing – original draft, Visualization, Validation, Soft­ based on deep learning method, Opt. Commun. 562 (2024) 130561.
ware, Methodology, Conceptualization. Gui Chen: Resources, Project [21] Z.A. Kudyshev, A.V. Kildishev, V.M. Shalaev, A. Boltasseva, Machine
administration. Mingjie Lin: Investigation, Data curation. Rendi Luo: learning–assisted global optimization of photonic devices, Nanophotonics 10
(2021) 371–383.
Project administration, Formal analysis, Data curation. Jialiang Zhang:
[22] H. Xia, S.-L. Chen, Y. Wang, Y. Zhao, H. Jia, R. Yang, Y. Jay Guo, Deep-learning-
Writing – review & editing, Formal analysis. Changshui Chen: Writing assisted intelligent design of terahertz hybrid-functional metasurfaces with
freeform patterns, Opt Laser. Technol. 181 (2025) 112041.

6
F. Shen et al. Optics Communications 591 (2025) 132171

[23] H. Zhan, X. Yuan, S. Wang, X. Chen, Z. Feng, H. Cui, J. Guo, Deep reinforcement [34] M. Konagai, M. Sugimoto, K. Takahashi, High efficiency GaAs thin film solar cells
learning hybrid models enabled broadband metamaterial solar absorbers by peeled film technology, J. Cryst. Growth 45 (1978) 277–280.
optimization and design, Opt. Commun. 584 (2025) 131824. [35] I. Massiot, N. Vandamme, N. Bardou, C. Dupuis, A. Lemaître, J.-F. Guillemoles,
[24] L. Gu, H. Liu, Z. Wei, R. Wu, J. Guo, Optimized design of plasma metamaterial S. Collin, Metal nanogrid for broadband multiresonant light-harvesting in ultrathin
absorber based on machine learning, Photonics 10 (2023) 874. GaAs layers, ACS Photonics 1 (2014) 878–884.
[25] J. Guo, L. Gu, X. Yuan, R. Wu, Q. Ma, On demand design of multi-freedom [36] J. Ruan, S. Ji, Z. Tao, F. Lan, Ultra-wideband metamaterial absorber doped GaAs in
metamaterial absorbers based on recurrent prediction networks, Opt. Mater. the infrared region, J. Electromagn. Waves Appl. 35 (2021) 1088–1098.
Express 14 (2024) 1822. [37] B. Wang, P. Liu, Z. Liu, H. Pan, X. Xu, R. Tang, Biomimetic Construction of Cellular
[26] S. Wang, X. Yuan, L. Gu, S. Xie, Q. Ma, Z. Wei, J. Guo, Innovative design of Shell by Adjusting the Interfacial Energy, 111, 2014, pp. 386–395.
metamaterial perfect absorbers via residual fully connected neural network [38] R. Kumar, N. Ranjan, R. Singh, V. Kumar, On development of thermoplastics based
modeling, Opt. Commun. 545 (2023) 129732. meta-structures by FFF based 3D printing, in: M.S.J. Hashmi (Ed.), Encyclopedia of
[27] S.-X. Xia, D. Zhang, X. Zhai, L.-L. Wang, S.-C. Wen, Phase-controlled topological Materials: Plastics and Polymers, Elsevier, Oxford, 2022, pp. 416–424.
plasmons in 1D graphene nanoribbon array, Appl. Phys. Lett. 123 (2023) 101102. [39] Y. Cheng, Y. Qian, H. Luo, F. Chen, Z. Cheng, Terahertz narrowband perfect
[28] D. Zhang, S. Xia, W. Xu, X. Zhai, L. Wang, Topological plasmonically induced metasurface absorber based on micro-ring-shaped GaAs array for enhanced
transparency in a graphene waveguide system, Phys. Rev. B 109 (2024) 245420. refractive index sensing, Phys. E Low-dimens. Syst. Nanostruct. 146 (2023)
[29] S.-X. Xia, X. Zhai, L.-L. Wang, S.-C. Wen, Plasmonically induced transparency in 115527.
double-layered graphene nanoribbons, Photon. Res. 6 (2018) 692. [40] L.D. Hai, V.D. Qui, N.H. Tung, T.V. Huynh, N.D. Dung, N.T. Binh, L.D. Tuyen, V.
[30] Y. Li, B. An, L. Li, J. Gao, Broadband LWIR and MWIR absorber by trapezoid D. Lam, Conductive polymer for ultra-broadband, wide-angle, and polarization-
multilayered grating and SiO2 hybrid structures, Opt. Quant. Electron. 50 (2018) insensitive metamaterial perfect absorber, Opt. Express 26 (2018) 33253–33262.
459. [41] H.J.O.e. Chen, Interference Theory of Metamaterial Perfect Absorbers, 20 7, 2011,
[31] H. Tao, N.I. Landy, C.M. Bingham, X. Zhang, R.D. Averitt, W.J. Padilla, pp. 7165–7172.
A metamaterial absorber for the terahertz regime: design, fabrication and [42] F. Qin, X. Chen, Z. Yi, W. Yao, H. Yang, Y. Tang, Y. Yi, H. Li, Y. Yi, Ultra-broadband
characterization, Opt. Express 16 (2008) 7181–7188. and wide-angle perfect solar absorber based on TiN nanodisk and Ti thin film
[32] J. Li, F. Wang, G. Wen, Y. Huang, W. Zhu, Planar Metamaterial for Matched structure, Sol. Energy Mater. Sol. Cell. 211 (2020) 110535.
Waveguide Termination, vol. 28, 2013, pp. 1236–1243. [43] H. Yuan, Z. Wang, T. Sun, Q. Song, Z. Yi, C. Tang, Q. Zeng, S. Cheng, P. Wu, Ultra-
[33] S. Duan, J. Hu, T. Liu, Z. Wang, J. Shi, X. Zhang, Cascaded nanocavity-shaped broadband absorber and near-perfect thermal emitter based on multi-layered
metasurfaces for realizing an intensive radiation absorption in the mid-to-long grating structure design, Energy 316 (2025) 134594.
infrared region based on electromagnetic wavefield resonance response and
accumulation, Opt. Mater. Express 14 (2024) 2303.