RESEARCH ARTICLE

Transcranial Acoustic Metamaterial Parameters

Inverse Designed by Neural Networks Citation: Yang Y, Jiang D, Zhang Q,
Yuming Yang1, Dong Jiang1, Qiongwen Zhang1, Xiaoxia Le2, Tao Chen2*, Le X, Chen T, Duan H, Zheng Y.
Transcranial Acoustic Metamaterial
Huilong Duan1*, and Yinfei Zheng 1* Parameters Inverse Designed
by Neural Networks. BME Front.
1
College of Biomedical Engineering and Instrument Science, Zhejiang University, Hangzhou, Zhejiang 2023;4:Article 0030. https://doi.
310027, China. 2Key Laboratory of Marine Materials and Related Technologies, Zhejiang Key Laboratory of org/10.34133/bmef.0030

Marine Materials and Protective Technologies, Ningbo Institute of Material Technology and Engineering, Submitted 11 April 2023
Chinese Academy of Sciences, Ningbo 315201, China. Accepted 4 September 2023
Published 25 September 2023
*Address correspondence to: duanhl@zju.edu.cn (H.D.); tao.chen@nimte.ac.cn (T.C.); zyfnjupt@zju.edu.
Copyright © 2023 Yuming Yang et al.
cn (Y.Z.) Exclusive licensee Suzhou Institute
of Biomedical Engineering and
Objective: The objective of this work is to investigate the mapping relationship between transcranial Technology, CAS. No claim to original

Downloaded from https://spj.science.org on September 29, 2023

U.S. Government Works. Distributed
ultrasound image quality and transcranial acoustic metamaterial parameters using inverse design under a Creative Commons
methods. Impact Statement: Our study provides insights into inverse design methods and opens the Attribution License 4.0 (CC BY 4.0).
route to guide the preparation of transcranial acoustic metamaterials. Introduction: The development
of acoustic metamaterials has enabled the exploration of cranial ultrasound, and it has been found that
the influence of the skull distortion layer on acoustic waves can be effectively eliminated by adjusting
the parameters of the acoustic metamaterial. However, the interaction mechanism between transcranial
ultrasound images and transcranial acoustic metamaterial parameters is unknown. Methods: In this
study, 1,456 transcranial ultrasound image datasets were used to explore the mapping relationship
between the quality of transcranial ultrasound images and the parameters of transcranial acoustic
metamaterials. Results: The multioutput parameter prediction model of transcranial metamaterials
based on deep back-propagation neural network was built, and metamaterial parameters under
transcranial image evaluation indices are predicted using the prediction model. Conclusion: This inverse
big data design approach paves the way for guiding the preparation of transcranial metamaterials.

Introduction complex numerical calculations on the structure [15]. This pro-

cess consumes tremendous time and computing resources owing
Manipulation and design of acoustic and electromagnetic meta- to its inherent iterative nature and requirement for full-field
materials are the study of controlling waves by designing delicate numerical simulation [16]. As the design becomes more complex
structures. Some exotic acoustic properties can be obtained in and the changes in the structure become more diverse, problems
acoustic metamaterials (AMMs) through ordered modulation of will require more time and resources to solve. Moreover, owing
sound waves, such as negative reactance [1,2], acoustic cloak [3], to the incomplete theoretical model of transcranial AMM, the
acoustic focusing [4,5], abnormal Doppler characteristics [6,7], interaction mechanism between transcranial imaging and AMM
and canceling out the distortion layer such as skull [8–12]. Some is unclear. It is challenging to design the structure of transcranial
scholars have found that by adjusting the parameters of the AMM, AMM with incomplete numerical models only by numerical
the influence of the skull distortion layer on the acoustic wave calculations. Therefore, it is urgent to find new ways to simplify
can be effectively eliminated [8,11]. By adding microstructure or even replace traditional design methods [17].
AMM before the distorted layer, the transmitted acoustic energy The advancement in the field of neural networks has shed
is substantially enhanced, and the reflected acoustic energy is light on developing new methods to overcome the challenges
reduced considerably. From a macroscopic point of view, the mentioned above. Material design incorporates structure as
transcranial AMM “offsets” the distorted layer in the acoustics. an additional design variable and thus offers new routes for
While ultrasound as a biologically beneficial mechanical wave- the development of materials with desired properties. It then
form has demonstrated practical implications in the biomedical allows for top-down reverse design approaches that map the
field [13,14], the transcranial AMM used in this application has required system performance onto materials structure and
a characteristic response to ultrasound stimulation. Therefore, it properties [18]. Currently, neural networks driven by big data
is theoretically feasible for transcranial AMM to compensate for have been used in the design and optimization of metamate-
the attenuation and distortion effect of the skull. rial parameters and achieved remarkable results [19–22]. In
However, designing a matching AMM structure for a spe- AMM research, neural-­network-based approaches have been
cific acoustic effect requires designers to repeatedly perform proposed for the inverse design of photonic structures [22],

Yang et al. 2023 | https://doi.org/10.34133/bmef.0030 1

acoustic cloak structures [23], acoustic absorber structures Results
[24,25], and so on. The inverse design process allows fast and
accurate prediction of the metamaterial structure and prop- Designing of transcranial AMM parameter prediction
erties with neural networks. neural network
This study proposes a neural-network-decision-based param- In this study, we use a back-propagation (BP) neural network
eters prediction method for transcranial AMM shown in Fig. 1. to find the mapping relationship between the parameters of
The reverse design approaches map the required transcranial transcranial AMM and the quality of the transcranial ultra-
ultrasound imaging system performance onto AMM parameters. sound image. The BP neural network is one of the most widely
The big data model was used to find the mapping relationship used neural network models, suitable for finding the mapping
between the parameters of transcranial AMM (particle size, fill- or matching relations between data [26]. There is no relation-
ing ratio, and thickness) and the quality of the transcranial image ship in time or space between the data in the dataset we used,
(lateral resolution, axial resolution, and imaging depth). Then, and the dataset has few dimensions. Therefore, we choose a BP
material parameters were deduced from the imaging experiment neural network to predict the parameters of transcranial AMM. The
to guide AMM preparation. Notably, this framework allows us flow chart of the BP neural network is presented in Fig. 2A.
to design transcranial AMM using realistic transcranial ultra- We took the imaging depth, lateral resolution, and axial res-
sound imaging results, which is novel for metamaterial design olution as the input of the prediction model. For the output, we
techniques. took the average particle size, filling ratio, and thickness of the

Downloaded from https://spj.science.org on September 29, 2023

Fig. 1. Neural-network-decision-based parameter prediction method for transcranial AMM. The inverse model inputs imaging parameters and predicts the corresponding
material parameters. (A) The transcranial AMM preparation and the main parameter of the transcranial AMM. (B) The imaging parameter data acquisition from the transcranial
ultrasound imaging system. (C) The reverse design approaches map the required transcranial ultrasound imaging system performance onto AMM parameters.

Yang et al. 2023 | https://doi.org/10.34133/bmef.0030 2

AMM. We determined that the number of nodes in the input where Yi represents the expected output of the model, yi rep-
layer should be 3, and the number of nodes in the output layer resents the actual output of the model, and N represents the
should also be 3. In this paper, we chose the Adam (adaptive number of samples.
moment estimation) [27] optimizer and ReLU (rectified linear To determine the number of hidden layers, we fixed the
units) [28] activation function, as shown in Fig. 2B. The other remaining hyperparameters and set the number of hidden
hyperparameters were determined by experiments. The mean layers to 3, 4, 5, and 6. The learning rate is set to 0.05, ReLU
square error (MSE) and mean absolute error (MAE) were used function is used as the network activation function, Adam
to evaluate the reliability of the model, while hyperparameters algorithm is selected as network optimization function, and
were adjusted. The following equations are used to express MSE the number of nodes in each hidden layer is fixed at 4. After
and MAE: 500 iterations, the number of hidden layers was fixed at 4,
which minimizes prediction error. The experimental results
1 ∑N ( )2
(1)
MSE = Yi − yi are shown in Fig. 3A.
N i=1
To determine the number of nodes in each hidden layer, we
first changed the number of nodes in the first hidden layer
1 ∑N | (4, 5, 6, 7, and 8) while fixing the number of nodes in the fol-
MAE = Y − yi ||
N i=1 | i (2) lowing hidden layer to 4 and then chose the number of nodes

Downloaded from https://spj.science.org on September 29, 2023

Fig. 2. Results for the transcranial AMM parameters prediction neural network. (A) The chart flow of the BP algorithm. (B) Variation diagram of loss in the training process of
different activation functions. (C) Variation diagram of loss in the training process of the BP neural network.

Yang et al. 2023 | https://doi.org/10.34133/bmef.0030 3

 Downloaded from https://spj.science.org on September 29, 2023
Fig. 3. Experimental results of hyperparameters comparison. (A) MSE and MAE of the different number of hidden layers. (B) MSE of the different number of nodes in different
hidden layers. (C) MSE of the different number of nodes in different hidden layers. (D) MSE and MAE of different learning rates. (E) MSE and MAE of different decay rates.

that produced the optimal results. The optimal nodes of the experimental results, the decay factor of the learning rate was
first hidden layer were fixed, and the number of nodes in other set to 0.999.
hidden layers was determined similarly. Each experiment was On the basis of the experiments, the hyperparameters of the
iterated 500 times. The experimental results of the first layer BP neural network used in this paper are shown in Table S1.
are shown in Fig. 3B and C. According to the experimental Briefly, the optimization algorithm is Adam method, the acti-
results of the prediction model, the number of nodes in hidden vation function is ReLU function, the number of hidden layers
layers was set to 7-4-4-4, and the overall network structure was is 4, the number of hidden layer nodes is 7-4-4-4, the learning
3-7-4-4-4-3. rate is 0.05, the learning rate attenuation factor is 0.999, and
The fixed learning rate method cannot effectively achieve the dropout probability value is 0.97.
fast network convergence with a small dataset, so we chose the
exponential decay learning rate method to dynamically and The preparation of hydrogel-based transcranial AMM
flexibly set the learning rate. The following equations are used A series of hydrogel-based metamaterials with adjustable bub-
to express the exponential decay learning rate: ble apertures were fabricated via a 2-step method. First, hydro-
philic monomer acrylamide (AAm) was polymerized at room
global_step
(3) temperature in the presence of N,N′-methylene bis(acrylamide)
Lr = Lrbase × 𝜏 decay_steps as a cross-linker, ammonium persulfate as an initiator, and
N,N,N′,N′-tetramethylethylenediamine as an accelerator. After
where Lr represents the learning rate, Lrbase represents the base reaction for 24 h, the prepared polyacrylamide (PAAm) hydro-
learning rate, τ represents the decay factor (taken from a con- gels were crushed and swollen in the water to remove unreacted
stant between 0 and 1), global_step represents the current num- monomers. Then, the fully swollen PAAm hydrogel fragments
ber of training steps, and decay_steps represents the decay rate were freeze-dried and refined into smaller powders. Through
(the decay rate is set to 1, meaning that the learning rate is the freeze-thawing method, the final hydrogel-based metama-
updated every iteration). We implemented several common terials were obtained by doping different amounts of the above
choices of learning rate; the experimental results are shown in powder into poly(vinyl alcohol) (PVA) solution (10 wt%) with
Fig. 3D. According to the experimental results, the basic learn- the assistance of different stirring rates and sealed in a sand-
ing rate was set to the optimal value (0.05). Then, we experi- wiched mode. By adding different amounts of PAAm, the bub-
mented to determine the decay factor of the learning rate; the ble apertures can be adjusted. Transcranial AMM with different
experimental results are shown in Fig. 3E. According to the amounts of PAAm is characterized in Fig. 4.

Yang et al. 2023 | https://doi.org/10.34133/bmef.0030 4

 Downloaded from https://spj.science.org on September 29, 2023
Fig. 4. Characterizations of transcranial AMM. (A) A sample of transcranial AMM with 2 mm in thickness. Scale bar, 2 mm. Hydrogel-based metamaterials with adjustable bubble
apertures, which were achieved by adding different amounts of PAAm [e.g., (B) 0.1 wt%, (C) 0.5 wt%, (D) 1 wt%, (E) 2 wt%, and (F) 3 wt%]. Scale bars, 200 μm.

Training of transcranial AMM parameter The main parameters (mean particle size, filling ratio, and
prediction neural network thickness) were changed to prepare the transcranial AMM with
Data were collected through the transcranial ultrasound imag- different parameter ratios. The average particle size ranges from
ing experimental system, as shown in Fig. 1B. The experimen- 10 to 400 μm, the filling ratio ranges from 0.1% to 5%, and the
tal system was mainly composed of 5 parts: ultrasonic imaging material thickness varies from 2 to 8 mm. On the basis of the
system, transcranial AMM, skull, water tank, and resolution transcranial AMM, ultrasonic craniotomy images were col-
phantom [29]. The ultrasonic imaging system was used to col- lected. The evaluation indices of ultrasonic image quality (lateral
lect the transcranial images. Transcranial AMM was used to resolution, axial resolution, and imaging depth) were obtained
enhance the energy of ultrasonic penetration through the skull. by offline calculation of the relevant image data. This prepared
The resolution phantom was the imaging target of the system. the data for the subsequent prediction of the parameters of the
A resolution phantom for ultrasonic craniotomy was designed transcranial AMM by the neural network.
and manufactured to measure imaging resolution parameters. The ultrasonic signals sent by the ultrasonic probe are sequen-
Figure 1 shows the schematic diagram and the picture of the tially passing through the transcranial AMM and the skull to
phantom with a resolution of 3 mm. Transcranial ultrasound image the resolution phantom placed inside the skull. Ultrasonic
imaging was carried out using the prepared transcranial AMM. images were collected at imaging depths of 15, 17.5, 20, 22.5, 25,
Figure 5 shows the schematic diagram and effect of the tran- 27.5, 30, 32.5, 35, 37.5, 40, 42.5, and 45 mm, and the correspond-
scranial ultrasound imaging experiment. In Fig. 5A and C, the ing lateral resolution and axial resolution resolutions were cal-
ultrasound probe penetrated the human skull to image the culated offline. The full width at half maximum (FWHM) is a
target phantom with and without transcranial AMM. Figure standard resolution evaluation method in the imaging field,
5D and F shows ultrasound images obtained from Fig. 5A and which is usually obtained by calculating the point spread function
C, respectively. Figure 5B and E is the control group in water of the system. This study adopted an energy drop of 8 dB as the
without the skull. As shown in Fig. 5F, we can see that without standard FWHM to measure the resolution of ultrasonic images.
the transcranial AMM, the target phantom cannot be distin- On the basis of the above design of the ultrasonic craniotomy
guished. However, with the transcranial AMM in Fig. 5D, the experiment, a total of 1,456 sets of data were collected. The col-
targets phantom can be clearly imaged. In Fig. 5D and E, the lected data were further preprocessed to ensure the effectiveness
horizontal hyperechoic lines represent the echo signal from of network training. The first step in this process is data cleaning.
the upper and lower edges of the resolution phantom, and the Data cleaning principles are as follows: The image data items with
hyperechoic dots represent the cross-sectional imaging of the poor effect are directly deleted. The offline calculation achieves
wire wrapped in the resolution phantom, which means that the resolution using the average peak FWHM of the 3 lines inside
the targets can be clearly distinguished. In contrast to Fig. 5F, the phantom. The FWHM of the central peak is 0.5 times away
in which the presence of the cranium results in attenuation of from the average value, indicating that this value is abnormally
the signal, only the blurry top and bottom edges of the mold higher or lower than the 2 values before and after it, and this group
can be seen, and the internal information from the thin metal of data is discarded from consideration. After data cleaning, 915
lines is absent. sets of available compelling data were screened out and split into

Yang et al. 2023 | https://doi.org/10.34133/bmef.0030 5

 Downloaded from https://spj.science.org on September 29, 2023
Fig. 5. Schematic diagram and renderings of the transcranial ultrasound imaging experiment. (A) Transcranial ultrasound imaging with the presence of transcranial AMM.
(B) Direct imaging with the 3-mm resolution phantom. (C) Transcranial ultrasound imaging without the presence of transcranial AMM. (D to F) Corresponding ultrasonic
images of (A), (B), and (C), respectively.

3 distinct groups: 60% of the data were for training, 20% was for as shown in Fig. 6A. The ultrasound probe was used to collect
validation, and 20% was for final testing. The adequate data were the transcranial images through the cerebellum fossa (Fig. 6B).
normalized for deviation. By scaling the data according to the Through computed tomography scan (Philips/iCT 256), the
scale, the sample data fell in a range between 0 and 1. thickness of skull at Z1 is about 1.53 mm. Transcranial AMM
As shown in Fig. 2C, the loss values of the BP neural network was used to enhance the energy of ultrasonic penetration through
for the test datasets can converge to a low-value region after the skull. The 2-mm resolution phantom was the imaging target
around 100 iterations. Regarding error evaluation indices, the of the system (Fig. 6C). It can be seen from Fig. 6E that the
MSE and MAE of the BP neural network the test datasets are signal amplitude of the 3 wires in the resolution phantom is
0.02479 and 0.08195, respectively. Furthermore, the training not uniform because of the nonuniformity of the skull. At 20-,
time is 2.11396 s. Briefly, the BP neural network can obtain 30-, and 40-mm imaging depth, different transcranial AMMs
relatively good prediction results after the iterative training can distinguish the 2-mm resolution phantom. The quality of
once the hyperparameters are set to the optimal value. transcranial images was further verified through the actual
The BP network model with high prediction accuracy was preparation of transcranial AMM, and the practicality of the
further used to predict transcranial AMM parameters in the network model was demonstrated.
region with a rich cerebrovascular distribution (20 to 40 mm in
depth). The lateral and axial resolutions were fixed at 2 mm, and Conclusions
the predicted results are shown in Table S2. It can be seen that
the network model built in this study can establish the mapping The skull’s strong attenuation and distortion effect on ultrasound
and matching relationship between the quality of transcranial make it challenging to realize transcranial ultrasound imaging. It
images and the parameters of AMM and then guide the actual is of great theoretical and clinical importance to establish the
preparation of transcranial AMM. mapping relationship between the quality of transcranial ultra-
After using the BP network model to predict transcranial sound images and the parameters of transcranial AMM. In this
AMM parameters for 2 mm resolution at different imaging study, image data were collected using a skull body model and the
depths, we prepared different transcranial AMMs to verify their resolution phantom, and the dataset was constructed. Prediction
ultrasound imaging qualities. Data were collected through models based on deep BP neural networks were designed, and
the transcranial ultrasound imaging experimental system, the actual prediction of the parameters of a metamaterial was

Yang et al. 2023 | https://doi.org/10.34133/bmef.0030 6

 Downloaded from https://spj.science.org on September 29, 2023
Fig. 6. Lateral profile of energy distribution for different transcranial AMMs at different imaging depths. (A) Transcranial ultrasound imaging experiment device. (B) Data
acquisition through the cerebellum fossa (Z1). (C) The photo of the resolution phantom. (D) Transcranial ultrasound image of 2-mm resolution phantom. (E) Lateral energy
distribution profile at different imaging depths for different transcranial AMMs.

carried out, which is of great importance for guiding the prepa- The resolution phantom was made of polydimethylsiloxane
ration of transcranial AMM. However, this study still has some wrapped in 3 fine metal wires, and the spacing of the metal wires
limitations, and improvements can be made in the following ways: was the measurement index of resolution. The diameter of the
(a) by further improving the quality of transcranial ultrasound gold filament was 0.3 mm, and the length, width, and thickness
image acquisition; (b) by further expanding the scale of neural of the separation rate die were 50, 15, and 4 mm, respectively.
network datasets (such as increasing the number of parameters In this study, a SonixTouch (Ultrasonix, Richmond, Canada)
of the transcranial AMM); (c) by further optimizing the model ultrasonic imaging system was used to control the L14-5 linear
and improving the accuracy of predicting the parameters of tran- array transducer (plane size of 45 mm × 10 mm) with a transmis-
scranial AMM (for example, including the introduction of a phys- sion center frequency of 5-MHz ultrasonic signals. A SonixDAQ
ics guided inverse design model [30]) or the use of multilayer (Ultrasonix, Richmond, Canada) multichannel echo data acqui-
perceptron network to mine the complex relation between the sition module was used to collect ultrasonic echo signals, obtain
transcranial AMM parameters and the imaging quality charac- images at different imaging depths, and conduct offline processing
teristics [31]; (d) by improving the preparation method of the on the echo signals to calculate image resolution, contrast, and
transcranial AMM (for example, through microfluidic technology contrast noise ratio.
[32,33]) and improving the preparation of the transcranial AMM
predicted by the network model to verify the accuracy of the
model prediction. Acknowledgments
Funding: This work was supported by the National Key R&D
Materials and Methods Program of China (2022YFB3204300), the Zhejiang Provincial
Key R&D Program of China (2022C01002), and the National
Experiment system Major Scientific Research Instrument Development Project
The experimental system was mainly composed of 5 parts: ultra- (81827804). Author contributions: Y.Y. and D.J. designed and
sonic imaging system, transcranial AMM, skull, water tank, and carried out the experiments. Y.Y., Y.Z., D.J., and Q.Z. wrote the
resolution phantom. The ultrasonic imaging system was used to paper. T.C. and X.L. guided the preparation of AMMs. H.D.
collect the transcranial images. Transcranial AMM was used to and Y.Z. guided the project. All authors contributed to the data
enhance the energy of ultrasonic penetration through the skull. analysis and discussions. Competing interests: The authors
The resolution phantom was the imaging target of the system. declare that they have no competing interests.

Yang et al. 2023 | https://doi.org/10.34133/bmef.0030 7

Data Availability 16. Zangeneh-Nejad F, Sounas DL, Alù A, Fleury R. Analogue
computing with metamaterials. Nat Rev Mater. 2021;6(3):207–225.
The data are available from the authors upon a reasonable request. 17. de Pablo JJ, Jackson NE, Webb MA, Chen L-Q, Moore JE,
Morgan D, Jacobs R, Pollock T, Schlom DG, Toberer ES, et al.
Supplementary Materials New frontiers for the materials genome initiative. npj Comput
Mater. 2019;5(1):1–23.
Table S1. Hyperparameters of BP neural network prediction 18. Tan RK, Zhang NL, Ye W. A deep learning–based method for
model. the design of microstructural materials. Struct Multidiscip
Table S2. Prediction results of 2-mm resolution network at dif- Optim. 2020;61(4):1417–1438.
ferent imaging depths. 19. Wu L, Liu L, Wang Y, Zhai Z, Zhuang H, Krishnaraju D,
Wang Q, Jiang H. A machine learning-based method to design
References modular metamaterials. Extreme Mech Lett. 2020;36:100657.
20. Hou Z, Tang T, Shen J, Li C, Li F. Prediction network of
1. He H, Chunyin Q, Liping Y, Xiangxi C, Xiying F, Manzhu K, metamaterial with Split ring resonator based on deep learning.
Fan Z, Zhengyou L. Topological negative refraction of Nanoscale Res Lett. 2020;15(1):1–8.
surface acoustic waves in a Weyl phononic crystal. Nature. 21. Kollmann HT, Abueidda DW, Koric S, Guleryuz E, Sobh NA.
2018;560(7716):61–64. Deep learning for topology optimization of 2D metamaterials.
2. Park CM, Lee SH. Zero-reflection acoustic metamaterial with a Mater Des. 2020;196:109098.
negative refractive index. Sci Rep. 2019;9(1):1–7. 22. Ma W, Liu Z, Kudyshev ZA, Boltasseva A, Cai W, Liu Y. Deep
3. Zhang S, Xia C, Fang N. Broadband acoustic cloak for learning for the design of photonic structures. Nat Photonics.
ultrasound waves. Phys Rev Lett. 2011;106(2):024301. 2021;15(2):77–90.
4. Zhang S, Yin L, Fang N. Focusing ultrasound with 23. Ahmed WW, Farhat M, Zhang X, Wu Y. Deterministic

Downloaded from https://spj.science.org on September 29, 2023

an acoustic metamaterial network. Phys Rev Lett. and probabilistic deep learning models for inverse design of
2009;102(19):194301. broadband acoustic cloak. Phys Rev Res. 2021;3(1):013142.
5. Song GY, Huang B, Dong HY, Cheng Q, Cui TJ. Broadband 24. Mahesh K, Kumar Ranjith S, Mini R. Inverse design of a
focusing acoustic lens based on fractal metamaterials. Sci Rep. Helmholtz resonator based low-frequency acoustic absorber
2016;6(1):1–7. using deep neural network. J Appl Phys. 2021;129(17):174901.
6. Zhai S, Zhao X, Liu S, Shen F, Li L, Luo C. Inverse doppler 25. Gurbuz C, Kronowetter F, Dietz C, Eser M, Schmid J, Marburg S.
effects in broadband acoustic metamaterials. Sci Rep. Generative adversarial networks for the design of acoustic
2016;6(1):1–10. metamaterials. J Acoust Soc Am. 2021;149(2):1162–1174.
7. Liu C, Long H, Zhou C, Cheng Y, Liu X. Reversed doppler 26. LeCun Y, Boser B, Denker JS, Henderson D, Howard RE,
effect based on hybridized acoustic mie resonances. Sci Rep. Hubbard W, Jackel LD. Backpropagation applied to
2020;10(1):1–7. handwritten zip code recognition. Neural Comput.
8. Shen C, Xu J, Fang NX, Jing Y. Anisotropic complementary 1989;1(4):541–551.
acoustic metamaterial for canceling out aberrating layers. 27. Kingma DP, Ba J. Adam: A method for stochastic optimization.
Phys Rev X. 2014;4(4):041033. arXiv. 2014. https://doi.org/10.48550/arXiv.1412.698
9. Craig SR, Welch PJ, Shi C. Non-Hermitian complementary 28. Nair V, Hinton GE. Rectified linear units improve restricted
acoustic metamaterials for lossy barriers. Appl Phys Lett. boltzmann machines. Paper presented at: Proceedings of the
2019;115(5):051903. 27th International Conference on Machine Learning (ICML-10);
10. Craig SR, Welch PJ, Shi C. Non-Hermitian complementary 2010 Jun 21; Madison, WI, USA.
acoustic metamaterials for imaging through skull with 29. Du B, Wang J, Zheng H, Xiao C, Fang S, Lu M, Mao R. A novel
imperfections. Front Mech Eng. 2020;6:55. transcranial ultrasound imaging method with diverging wave
11. Wang J, Allein F, Boechler N, Friend J, Vazquez-Mena O. transmission and deep learning approach. Comput Methods
Design and fabrication of negative-refractive-index Prog Biomed. 2020;186:105308.
metamaterial unit cells for near-megahertz enhanced acoustic 30. Chen J, Chen Y, Xu X, Zhou W, Huang G. A physics-guided
transmission in biomedical ultrasound applications. Phys Rev machine learning for multifunctional wave control in active
Appl. 2021;15(2):024025. metabeams. Extreme Mech Lett. 2022;55:101827.
12. Park CI, Piao C, Lee H, Kim YY. Elastic complementary 31. Zhou W, Wang S, Wu Q, Xu X, Huang X, Huang G, Liu Y,
meta-layer for ultrasound penetration through solid/liquid/gas Fan Z. An inverse design paradigm of multi-functional elastic
barriers. Int J Mech Sci. 2021;206:106619. metasurface via data-driven machine learning. Mater Des.
13. Yuk H, Wu J, Zhao X. Hydrogel interfaces for merging humans 2023;226:111560.
and machines. Nat Rev Mater. 2022;7(12):935–952. 32. Lukyanova L, Séon L, Aradian A, Mondain-Monval O, Leng J,
14. Huang D, Wang J, Song C, Zhao Y. Ultrasound-responsive Wunenburger R. Millifluidic synthesis of polymer core-shell
matters for biomedical applications. Innovation. micromechanical particles: Toward micromechanical resonators for
2023;4(3):100421. acoustic metamaterials. J Appl Polym Sci. 2013;128(6):3512–3521.
15. Elsawy MM, Lanteri S, Duvigneau R, Brière G, Mohamed MS, 33. Galogahi FM, Zhu Y, An H, Nguyen N-T. Core-shell
Genevet P. Global optimization of metasurface designs using microparticles: Generation approaches and applications.
statistical learning methods. Sci Rep. 2019;9(1):1–15. J Sci: Adv Mater Dev. 2020;5(4):417–435.

Yang et al. 2023 | https://doi.org/10.34133/bmef.0030 8