Acoustic Structure Inverse Design and Optimization Using Deep

Learning
Xuecong Sun1,2, Yuzhen Yang1*, Han Jia2,3, Han Zhao1,2, Yafeng Bi1, Zhaoyong Sun1
and Jun Yang1,2*
1 Key Laboratory of Noise and Vibration Research, Institute of Acoustics, Chinese Academy of
Sciences, Beijing 100190, People’s Republic of China
2 University of Chinese Academy of Sciences, Beijing 100049, People’s Republic of China
3
State Key Laboratory of Acoustics, Institute of Acoustics, Chinese Academy of Sciences,
Beijing 100190, People’s Republic of China

Correspondence should be addressed to Yuzhen Yang, yangyuzhen@mail.ioa.ac.cn, Jun Yang;

jyang@mail.ioa.ac.cn

Keywords: acoustics structure design, deep learning, multi-order Helmholtz resonator, sound
insulation

Abstract

From ancient to modern times, acoustic structures have been employed to manage the

spread of acoustic waves. Nevertheless, designing these structures traditionally

remains a laborious and computationally intensive iterative process. Recognizing that

complex acoustic systems can be effectively analyzed using the lumped-parameter

method, we introduce a deep learning model that learns the correlation between the

equivalent electrical parameters and the acoustic properties of these structures. As an

illustration, we consider the design of multi-order Helmholtz resonators, showing

experimentally that our model can predict structures with high precision that closely

align with the specified design criteria. Furthermore, our model can seek multiple

solutions in conjunction with dimensionality reduction algorithms and support

1
evolutionary algorithms in optimization tasks. Compared to traditional numerical

methods, our approach offers greater efficiency, flexibility, and universality. The

designed acoustic structures hold broad potential for applications including speech

enhancement, sound absorption, and insulation.

1. Introduction

Acoustic structures have been used for centuries to control acoustic waves in

terms of amplitude and phase. With extensive achievements in fabrication technology,

unprecedented functionalities can be obtained by designing and engineering artificial

structures with more complex properties. Recently, many exotic functionalities, such

as anomalous refraction/reflection [1,2], invisibility [3,4], timbre manipulation [5] and

novel acoustic sensing [6,7], have been realized with fantastic acoustic structures.

However, the development of accurate and computationally efficient design and

optimization approaches for the acoustic structures is still in the early stages. During

the design process, the forward calculation, i.e., predicting the acoustic properties

based on the geometric structures, is well understood with analytical and numerical

approaches, such as the lumped-parameter technique (LPT), transfer matrix method

(TMM) and finite element method (FEM) and modal displacement method [8].

Nevertheless, the inverse problem, i.e., inferring acoustic structures from on-demand

acoustic properties, is currently a prohibitive task even with the most advanced

numerical tools. To search the formidably large design space efficiently, the inverse

design procedure is usually guided by optimization algorithms, such as gradient-based

2
approaches and evolutionary approaches. As structural complexity grows, the above

methods will take a prohibitive amount of time, which seriously limits the usefulness

of such approaches. Therefore, it is of great significance to identify an efficient,

flexible and universal acoustic structure design method.

In recent years, deep learning has emerged as a very powerful computational

method. In light of its exceptional success in domains related to computer science and

engineering [9–13], deep learning has attracted increasing attention from researchers

in other disciplines, including materials science [14], chemistry [15], physics [16–18],

computational imaging and microscopy [19–21]. Moreover, deep learning has become

a radically new approach in the context of photonic and electromagnetic design, such

as approximation of light scattering from plasmonic nanostructures and inverse design

of the electromagnetic metasurface structure, over the past few years [22–27].

Recently, deep learning has also been used to solve the inverse problem of the

nonlocal metasurface, variable cross-sectional acoustic structure and two-dimensional

acoustic cloaking [28–31]. However, both the researches trained the deep neural

networks (DNNs) for specific structures, which are hard to be extended to other

acoustic structures.

In view of the current problems and previously reported designs, here we

propose an efficient, flexible and universal acoustic structure design method to solve

the inverse problem. In our method, we analyze acoustic structures by LPT and

develop connections between geometric parameters (GPs) and equivalent electrical

3
parameters (EEPs). Then, a DNN is set up and trained to determine the corresponding

relation between the EEPs of the acoustic structures and their acoustic properties. To

evaluate the effectiveness of the proposed design method, a multi-order Helmholtz

resonator is designed to realize acoustic insulation at specific frequencies (see Figure

1a). Once design requirements are input into the trained model, the EEPs of the

structures will be generated quickly and automatically. Then, the GPs can be

calculated through the LPT. In this example, it has been proven that the trained DNN

can not only give a satisfied solution of the inverse design problems more quickly

than its numerical counterpart, but also be used to search for multiple solutions

combined with the dimension reduction algorithm. Moreover, it can also be used to

improve the performance of evolutionary approaches in optimization for a desired

property. Finally, we design an acoustic filter to decrease the line-spectrum noise

using the proposed design method. This deep learning approach is an effective design

tool for acoustic structure on-demand design and optimization. Considering that

various acoustic structures can be analyzed by the LPT exactly in the low frequency

range, such as Mie resonant structure [32], membrane-cavity acoustic absorber [33]

and acoustic structure of microphone [34], the proposed approach has a strong

versatility and scalability, which can be further extended to other acoustic structures.

2. Method

2.1 Physical Model

4
Figure 1: Acoustic structure design method based on the deep learning. (a) The proposed deep
learning approach can realize the on-demand design of the acoustic structures. In this work, the
deep learning approach is used to design the THRs. The THRs are placed as a side branch of a
tube, which can be used to filter out the background noise and make the speech clearer. (b)
Schematic view of the THR: The THR is constructed by neck-and-cavity substructures of two
elements arranged in a cascade way. When the THR is placed as a side branch of a tube, there are
two STL peaks corresponding to the two resonant frequencies 𝑓1 and 𝑓2 . (c) Schematic view of the
DNN model. The THRs are designed by a trained DNN. The DNN has a cascaded structure of
three hidden layers of nonlinear processing units, where each layer uses the output from the
previous layer as input.

Owing to the excellent characteristics of manipulating low-frequency sound

waves with subwavelength dimensions, acoustic structures utilizing Helmholtz

resonators have emerged as an attractive option in various fields such as sound

proofing [35,36], asymmetric sound transmission [37], sound metadiffusers [38], and

acoustic superlens [39,40] et al. Moreover, Helmholtz resonators offer significant

flexibility and scalability, often being modified and combined for practical

5
applications. For example, Liu et al. proposed a metasurface muffling coating for

pipeline ventilation noise reduction, utilizing a Fabry-Perot channel composite design

to form a Helmholtz resonance cavity, achieving low-frequency broadband noise

reduction [41]. However, conventional Helmholtz resonators are limited to supporting

a single monopolar resonant mode with a narrow bandwidth, which restricts their use

in functional device design. To address this limitation, recent studies have introduced

multi-order Helmholtz resonators capable of producing multiple resonances [42–45].

Compared to conventional Helmholtz resonators, multi-order Helmholtz resonators

can provide additional resonances at higher frequencies without significantly

increasing volume. For example, Liu et al. achieve multi-order sound absorption by

inserting one or more separating plates with small holes into the interior of a

Helmholtz resonator, effectively broadening the absorption bandwidth significantly

[46]. Figure 1b illustrates a schematic of a two-order Helmholtz resonator (THR),

which is assembled from two neck-and-cavity substructures arranged sequentially.

The lower substructure is the 1st-order element, and the upper substructure is the 2nd-

order element. Here, 𝑎𝑖 and 𝑟𝑖 are the radii of the cylindrical necks and cavities,

respectively; 𝑙𝑖 and ℎ𝑖 are the lengths of the cylindrical necks and cavities,

respectively.

The acoustic properties of the THR can be analyzed by the LPT exactly. The ith-

order cavity is equivalent to the acoustic compliance 𝐶𝑖 . The ith-order neck is

equivalent to the acoustic inertance 𝑀𝑖 and the acoustic resistance 𝑅𝑖 .The relationship
6
between GPs 𝐠𝐩 = [𝑎1 , 𝑙1 , 𝑟1 , ℎ1 , 𝑎2 , 𝑙2 , 𝑟2 , ℎ2 ] and EEPs 𝐞𝐞𝐩 =

[𝑅1 , 𝑀1 , 𝐶1 , 𝑅2 , 𝑀2 , 𝐶2 ] is shown in Table 1. Here, 𝑉𝑖 = 𝜋𝑟𝑖2 ℎ𝑖 is the volume of the ith-

order cavity; 𝜌0 is the static air density; 𝑐0 is the sound speed, and 𝜂 is the viscosity of

8𝑎𝑖 𝑎
air. 𝛿𝑖 = (2 − 𝛽𝑖 𝑟 𝑖 ) is the end correction associated with the inner and outer
3𝜋 𝑖

openings of the ith-order neck, where 𝛽1 = 0.75 and 𝛽2 = 1.05 are the correction

factors (𝑖 = 1,2). The values of the physical constants can be seen in Table S1 of the

Supplementary Materials. Based on these EEPs, the acoustic impedance of the THR

can be written as follows:

1
𝑍𝑇𝐻𝑅 = 𝑅1 √𝜔 + j𝜔𝑀1 + , (1)
1
j𝜔𝐶1 +
𝑅2 √𝜔 + j𝜔𝑀2 + 1⁄j𝜔𝐶2

where 𝜔 = 2𝜋𝑓 is the angular frequency. When the THR is placed as a side branch of

a tube as shown in Figure 1a, the sound transmission loss (STL) can be expressed as:
2
𝑍
𝑋𝑏2 + (2𝑆0 + 𝑅𝑏 )
𝑡 = 10 log , (2)
𝑅𝑏2 + 𝑋𝑏2

where 𝑅𝑏 and 𝑋𝑏 are the real part and imaginary part of 𝑍𝑇𝐻𝑅 , respectively; 𝑍0 = 𝜌0 𝑐0

is the acoustic impedance of the air; 𝑆 is the cross section of the tube [47].

Considering that the THR with two neck-and-cavity substructures can induce two

discrete resonant modes, there are two peaks in the STL curve as shown in Figure 1b.

The frequency corresponding to the low-frequency peak is 1st-order resonant

frequency 𝑓1 , and the frequency corresponding to the high-frequency peak is 2nd-order

resonant frequency 𝑓2 . Therefore, the forward problem of the THR model can be

solved satisfactorily through the LPT.

7
 𝑮𝑷 → 𝑬𝑬𝑷 𝑬𝑬𝑷 → 𝑮𝑷
𝑹𝒊 = 𝒍𝒊 √𝟐𝜼𝝆𝟎 ⁄𝝅𝒂𝟑𝒊 𝜌0 𝑅𝑖 2 8𝜌0 𝛽𝑖 16𝜌0
𝑎𝑖 − ( 2
+ 𝑀𝑖 ) 𝑎𝑖 + =0
√2𝜂𝜌0 3𝑟𝑖 𝜋 3𝜋 2
𝑴𝒊 = 𝝆𝟎 (𝒍𝒊 + 𝜹𝒊 )⁄𝝅𝒂𝟐𝒊 3𝜋𝑅𝑖 𝑎𝑖
𝑙𝑖 =
√2𝜂𝜌0
𝑪𝒊 = 𝑽𝒊 ⁄𝝆𝟎 𝒄𝟐𝟎 𝑉𝑖 = 𝐶𝑖 𝜌0 𝑐02
Table 1. Relationship between GPs and EEPs.

2.2 Challenge of the inverse design problem

In practical applications such as sound insulation, acoustic structures are usually

designed according to noise spectra. And we want to find a structure whose STL

values are high enough at the noise frequencies. Equation (2) shows that STL reaches

its maximum value when the imaginary part 𝑋𝑏 = 0 , and the corresponding

frequencies are the resonant frequencies of THR. Therefore, the resonant frequencies

are generally designed closer to the noise frequencies to have good performance of

sound isolation. If we let 𝑋𝑏 = 0 and simplify the equation, a six-degree equation can

be obtained:

𝐶1 +𝐶2
𝑀1 𝐶12 𝑀22 𝜔06 + 𝑀1 𝐶12 𝑅22 𝜔05 − 𝐶1 𝑀2 (2𝑀1 + 𝑀2 ) 𝜔04 − 𝐶1 𝑅22 𝜔03
𝐶2
2 (3)
𝐶 𝐶 𝐶1 +𝐶2
+ [𝑀1 (𝐶1 ) + 2(𝑀1 + 𝑀2 ) 𝐶1 + 𝑀1 + 𝑀2 ] 𝜔02 − = 0,
2 2 𝐶22

where 𝜔0 is the resonant angular frequency. According to Table 1 and Eq. (3), it can

be seen that all the six GPs [𝑎1 , 𝑙1 , ℎ1 , 𝑎2 , 𝑙2 , ℎ2 ] are strong-coupling and affect the

resonant characteristics of the THR together, which bring difficulties to accurate

inverse design. When we adjust multiple GPs at the same time, the effect on the

acoustic functionality of THR is irregular. Therefore, it is difficult to design THR for a

target acoustic functionality through manual design method.

8
 Considering of the difficulty of inverse design, DNN is suitable to assist in the

design process for avoiding time-consuming numerical methods. However,

establishing a relationship between target acoustic functionalities and acoustic

structures is often a one-to-many multivalued problem, posing a significant challenge

for DNN construction and training. First of all, there can be multiple STL spectra that

meet target acoustic functionalities, each corresponding to a different set of EEPs.

This implies that multiple sets of EEPs can meet the design requirements. Secondly, a

single set of satisfactory EEPs also corresponds to more than one acoustic structures.

Figure 2 shows this multivalued challenge. Here, the design target is to find a

structure whose resonant frequencies are 150 Hz and 250 Hz, where the values of the

STL are at least 10 dB. Four satisfied structures are shown in Figure 2. All of the STL

spectra of these four structures meet the requirements, which are shown in Figure 2a

and c. Structure 1 and Structure 2 correspond to the same group of EEPs, whose

schematic views and GPs are shown in Figure 2b. Structure 3 and Structure 4

correspond to another group of EEPs, whose schematic views and GPs are shown in

Figure 2d. It can be seen that the actual THR structure can vary quite significantly for

the same design requirement. If the target acoustic functionalities and GPs of the

acoustic structures are treated as inputs and outputs for the DNN model, respectively,

the training dataset may contain numerous conflicting samples with identical inputs

but varied outputs. This scenario, where the DNN is intended to create a one-to-one

mapping between inputs and outputs, can complicate the convergence of the training.
9
Figure 2: Different acoustic structures corresponding to the same target acoustic
functionality. (a) STL spectrum calculated by LPT, whose corresponding EEPs are shown in the
inset. (b) The two THR structures corresponding to the EEPs shown in (a). (c) STL spectrum
calculated by LPT, whose corresponding EEPs are shown in the inset. (d) The two THR structures
corresponding to the EEPs shown in (c).

2.3 DNN model

Our goal is to use a DNN model to solve the inverse design problem for the

target acoustic functionality. As mentioned in the previous section, the mapping from

target acoustic functionalities to acoustic structures is typically one-to-many, which

brings difficulties to model training. Therefore, some strategies need to be introduced

to minimize the one-to-many effect. Firstly, the DNN model is trained to learn the

relationship between STL spectra and EEPs, instead of the relationship between STL

spectra and GPs. In this way, the one-to-many multivalued problem, which is caused

by the one-to-many mapping between EEPs and GPs, can be avoided. Moreover, we

make the STL spectrum, instead of the target resonant frequencies, be the input of the
10
DNN model. Two different groups of EEPs may have the same values of the STL at

some frequency points, but it is unlikely to be the same at all of the frequency points.

With the help of the above two strategies, the model training can converge

successfully. And the EEPs, which would most closely produce the target STL

spectrum, can be predicted based on the trained DNN model. The GPs can be

calculated based on the EEPs through the LPT (see Table 1).

The datasets were generated using LPT. Here, the ranges of the GPs are set as

0.1 cm < 𝑎𝑖 < 2.5 cm , 0.1 cm < 𝑙𝑖 < 5 cm and 0.1 cm < ℎ𝑖 < 12.7 cm ( 𝑖 = 1,2 ),

and the radii of the cavities are set as 𝑟1 = 𝑟2 = 5𝑐𝑚. The cross section of the tube is

𝑆 = 100 cm2 . To generate a sample, first, we sampled randomly within the given

ranges of the GPs and obtained a group of GPs. Next, the EEPs and the STL spectrum

could be calculated through the LPT. If the STL at the resonant frequencies exceeded

10 dB, the sample was retained. The above selection guarantees that all samples in the

dataset have strong sound isolation performance at the resonant frequencies, which

contributes to developing a more effective model. Furthermore, we need to ensure that

the samples were uniformly distributed across the frequency range of interest. To

achieve this, the samples were divided into 35 groups based on their resonant

frequencies, ensuring that the distribution of the samples is as uniform as possible (see

Table S2 in the Supplementary Materials). Figure 3a shows the distribution of the

filtered samples. Here, 𝑓1 and 𝑓2 are the 1st-order and 2nd-order resonant frequencies

of the THR, respectively. The two resonant frequencies 𝑓1 and 𝑓2 of the samples in the
11
datasets cover the frequency range of interest. Each square in this figure is

corresponding to a group of samples in the certain frequency band, where the center

frequencies are 𝑓1 and 𝑓2 , and the bandwidth is 50 Hz. The color of the square

𝑡(𝑓1 )+𝑡(𝑓2 )
represents the average STL at the resonant frequency of the samples.
2

In total, our dataset contains 195,000 samples, which were split into training,

validation, and test sets, with proportions of 80%, 10%, and 10%, respectively. We

trained a fully connected network with three hidden layers based on this dataset. The

network's structure is illustrated in Figure 1c. A desired STL spectrum 𝐭 =

[𝑡1 , 𝑡2 , 𝑡3 , … , 𝑡𝑛 ] is taken as the input, where the STL values are higher than 10 dB at

the target frequencies. The STL spectrum is sampled from 101 Hz to 600 Hz with a

step of 1 Hz. Thus, the number of inputs is 𝑛 = 500. To facilitate model training, the

3
EEPs of the samples were normalized to the following ranges: 1 Pa ∙ s 2 /m3 < 𝑅𝑖 <
3
170 Pa ∙ s 2 /m3 , 1 kg/m4 < 𝑀𝑖 < 300 kg/m4 and 7 × 10−10 m4 s 2 /kg < 𝐶𝑖 <

7 × 10−9 m4 s 2 /kg (𝑖 = 1,2). The model's output is the normalized 6-dimensional

EEPs. The numbers of neurons of each hidden layer are 400, 250 and 220 respectively.

The inputs were normalized, shuffled and then fed into the network, which can

accelerate convergence of the algorithm. The mean square error (MSE) was used to

represent a loss function between the desired and actual output. The train loss was

used to generate the gradients, and the network weights were updated by the Adam

algorithm to minimize the discrepancy [48]. The hyperparameters (for example,

number of hidden layers, neurons and learning rate) were set according to the
12
performance on the validation set through grid search. We utilized the batch

normalization technique to improve the convergence speed of the training [49]. In

addition, the dropout regularization technique was employed to avoid overfitting [50].

The hyperparameters for model training can be found in Table 2. The learning rate is

set as 0.001, and the batch size is set as 256. We stopped training when the validation

loss stopped decreasing, and the learning curve of the DNN on the validation set is

shown in Figure 3b. The rapidly decreasing MSE of validation instances shows that

training is highly effective.

The test set is used to evaluate the generalization capability of the trained DNN.

The average loss of the test set converges to 0.0029. And the prediction results on the

test set can be found in S1 of the Supplementary Materials. Based on these results, it

can be concluded that the predictive error is within an acceptable range, indicating

that the model possesses satisfactory generalization performance. Considering that our

model contains 358,948 trainable parameters and the STL spectrum is highly sensitive

to variations in EEPs, a large amount of high-quality data is essential to ensure the

model's performance and prevent overfitting. Detailed analysis of the high-quality

data selection and the dataset size can be found in S2 and S3 of the Supplementary

Materials, respectively.

13
Figure 3: Distribution of samples and learning curve of the DNN. (a) Distribution of the
filtered samples. (b) The MSE of validation set over the epoch.

Learning Rate Batch Size Dropout Rate Activation function

0.001 256 0.2 ReLu
Table 2. Hyperparameters for model training.

3. Results

3.1 DNN solves the inverse design problem

The concept schematic of the design process is shown in Figure 4. Here, we

show an example using the trained DNN to realize acoustic insulation at specific

target target
frequencies. The target resonant frequencies are set as 𝑓1 = 150Hz and 𝑓2 =

250Hz, where the values of the STL are required to exceed 10 dB. Then we generate

a group of curves as the desired STL spectra according to the requirements.

Considering that the profile of the STL peak is similar to the Lorentzian profile, a

desired STL curve can be created by scaling, translating and overlaying the standard

Lorenz curve. Although it cannot be guaranteed that an STL spectrum generated in

this manner will definitely correspond to a group of EEPs, the proposed DNN model,

with its robust fitting capabilities, will endeavor to predict the EEPs that best meet the

requirements. We feed these spectra into the proposed DNN model and obtain the
14
predictive EEPs. Then, the GPs can be calculated through the LPT. Finally, we can

calculate the real STL spectra by the TMM (see S4 of the Supplementary Materials)

and select the most qualified structure based on their average error of the resonant

frequency (AERF), which can be expressed as

target target
1 |𝑓1 (𝐠𝐩) − 𝑓1 | |𝑓2 (𝐠𝐩) − 𝑓2 |
AERF(𝐠𝐩) = ( target + target ). (4)
2 𝑓1 𝑓2

Here, 𝑓1 (𝐠𝐩) and 𝑓2 (𝐠𝐩) are the real resonant frequencies of the structure whose GPs

are denoted as 𝐠𝐩. By using TMM to refine the predicted structures, we can reduce

the approximation error introduced by the LPT, ensuring that the designed THR better

meet the design requirements. Moreover, if computational resources allow, more

accurate methods such as finite element analysis (FEM) can be used for further

refinement and selection of the best design.

Figure 4: Concept schematic of the design process.

15
 Based on the aforementioned methodology, the selected structure 𝛽0 is produced,

whose GPs and STL spectrum are shown in Figure 5a. There are two STL peaks

higher than 10 dB at 150 Hz and 250 Hz, which is in good agreement with the design

requirements. Based on the proposed model, we can further implement quantitative

design to specify STL values, such as 9dB or 12dB, at target frequencies, which is

highly valuable for scenarios that require more precise designs, such as the

development of acoustic filters [6, 51]. The results can be found in S5 of the

Supplementary Materials.

Building on this foundation, we introduced minor perturbations to the GPs of 𝛽0,

resulting in significant changes to the STL spectrum, which no longer met the design

requirements (see S6 of the Supplementary Materials). This further demonstrates that

the inverse design problem necessitates a refined adjustment of the GPs. The

proposed design scheme based on DNN has proven to be highly effective in

addressing this issue. In addition to the aforementioned, we have also chosen design

objectives that were not encompassed within the training dataset to assess the

generalization capability of the DNN model. The outcomes can be found in S7 of the

Supplementary Materials.

It is noteworthy that the model directly predicts the EEPs of the acoustic

structure without imposing any constraints on the geometric structure itself. The THR

is merely one physical realization of this set of EEPs. Due to the limitations of the

mapping relationship between GPs and EEPs, there may be instances where the EEPs
16
predicted by the model cannot be realized by a specific THR. In such cases, we can

seek alternative acoustic structures capable of realizing this set of EEPs within the

framework of LPT. In S8 of the Supplementary Materials, we further demonstrate the

flexibility and transferability of the proposed model using the double-neck THR as an

illustrative example.

Figure 5: Results of the inverse design. (a) STL spectrum of the selected THR structure 𝛽0
calculated by the TMM (blue solid line). The target resonant frequencies 150 Hz and 250 Hz are
marked by the red dotted lines. The GPs of the selected THR structure are shown in the table. (b)
AERF varies with the two most important principal components 𝑝1 and 𝑝2 in PCA space. The
selected structure 𝛽0 is marked with the yellow star. (c) 𝐻1 varies with the two most important
principal components 𝑝1 and 𝑝2 . The best structure 𝛽1 is marked with the red star, whose STL at
150 Hz 14.26 dB. (d) STL spectrum of the structure 𝛽1 , whose GPs are shown in the table.

3.2 Search for multiple solutions

As mentioned before, the corresponding relation between target acoustic

17
functionality and acoustic structures is usually not a deterministic one-to-one mapping.

In the previous section, the design target is to find a structure whose resonant

frequencies are 150 Hz and 250 Hz, where the values of the STL are at least 10 dB.

This is a one-to-many multivalued problem. Considering that the multiple solutions

distribute in the whole GP space, it will take a prohibitive amount of time to find them

through searching in GP space. Here, the proposed DNN model and the principal

component analysis (PCA) can be combined to search for multiple solutions quickly.

PCA is a linear transformation algorithm, which can find a sequence of best

orthogonal linear projections (called principal components) that maximizes the

corresponding variances. Therefore, PCA can be used to map and characterize the

multi-parameter design space, which can transform a set of correlated variables into a

set of new uncorrelated variables [52].

The PCA variable space can be defined by six orthogonal basis vectors 𝐕𝟏 ,

𝐕𝟐 , …, 𝐕𝟔 . In the PCA variable space, the selected structure 𝛽0 shown in Figure 5a

𝛽 𝛽 𝛽
can be described as 𝐏 𝛽0 = [𝑝1 0 , 𝑝2 0 , … , 𝑝6 0 ], and the structures with similar acoustic

properties will gather together. Therefore, we can search around the 𝐏 𝛽0 to find other

satisfied solutions. In order to visualize the results, we generate a 2D hyperplane

defined by the two most important principal components 𝐕𝟏 and 𝐕2 . For each point

[𝑝1𝑖 , 𝑝2𝑖 ] on the hyperplane, the corresponding GPs 𝐠𝐩𝑖 = [𝑎1𝑖 , 𝑙1𝑖 , 𝑟1𝑖 , ℎ1𝑖 , 𝑎2𝑖 , 𝑙2𝑖 , 𝑟2𝑖 , ℎ2𝑖 ]

𝛽
in the original design space can be computed as 𝐠𝐩𝑖 = 𝐕0 + 𝑝1𝑖 𝐕1 + 𝑝2𝑖 𝐕2 + 𝑝3 0 𝐕3 +

𝛽 𝛽 𝛽
𝑝4 0 𝐕4 + 𝑝5 0 𝐕5 + 𝑝6 0 𝐕6 , where 𝑝1𝑖 varies from -5 to 5 and 𝑝2𝑖 varies from -12 to 5.
18
And the corresponding AERFs are computed and shown in Figure 5b. It is clear that

the selected structure 𝛽0 (marked with a yellow star) is one of the optimal solutions

within the range of observations, which demonstrates the accuracy of the proposed

design technology.

In fact, there are many points surround 𝛽0 in PCA space whose corresponding

acoustic properties are also in good agreement with the design requirements. However,

the corresponding structures of these points may vary considerably. Therefore, we can

make further selections based on different properties in other aspects. For example, if

the target is to obtain a structure whose STL in 150 Hz can be as high as possible and

STL in 250 Hz can higher than 10 dB, we can define 𝐻1 as

target target target

𝐻1 = {𝑡(𝑓1 , 𝐠𝐩) 𝑖𝑓 𝑡(𝑓1 , 𝐠𝐩), 𝑡(𝑓2 , 𝐠𝐩) > 10𝑑𝐵 , (5)
0 𝑒𝑙𝑠𝑒

target target target

where 𝑓1 = 150 𝐻𝑧, 𝑓2 = 250 𝐻𝑧; 𝑡(𝑓1 , 𝑔𝑝) are the values of the STL

at 150 Hz of the THR, whose GPs are denoted as 𝐠𝐩 = [𝑎1 , 𝑙1 , 𝑟1 , ℎ1 , 𝑎2 , 𝑙2 , 𝑟2 , ℎ2 ]. It

can be seen that the higher 𝐻1 is, the better the corresponding structure meet the

design requirements. Then, 𝐻1 of the points surround 𝛽0 can be calculated. The results

are shown in Figure 5c. We pick the point with highest 𝐻1 (marked with red star)

within the range of observations, which corresponding structure is 𝛽1 . The STL

spectrum and GPs of the structure 𝛽1 are shown in Figure 5d. It can be seen that the

STL at 150 Hz is improved to 14.26 dB compared with the structure 𝛽0. Details about

the PCA and another example with high performance in 250 Hz are provided in S9 of

19
the Supplementary Materials.

Consequently, a lower-dimensional subspace that exhibits high performance can

be discerned utilizing dimensionality reduction methodologies. Within this reduced

subspace, the predictive outcomes furnished by the DNN model serve as an anchor

point. By conducting a localized search in the vicinity of this anchor, we are enabled

not only to conduct a more efficient search for multiple solutions but also to make

additional selections with a reduced computational cost when various performance

criteria are taken into account.

3.3 DNN aided optimization method

In the previous section, our goal was to design a structure with STL values that

surpassed 10 dB at the target frequencies. Yet, in real-world scenarios, the higher

these values, the better, often requiring the application of optimization algorithms to

adjust the structural parameters effectively [53]. A popular approach involves the use

of evolutionary algorithms, such as genetic algorithms (GA) and particle swarm

optimization [54–56]. However, these algorithms can be quite sensitive to the starting

conditions, which may result in challenges like slow convergence or getting stuck at

local optima. Now that we can easily design an arbitrary STL spectrum with the

proposed DNN, we can leverage it to offer a solid starting point for the evolutionary

algorithms. Specifically, the initial population could consist of a set of high-

performing individuals generated by the DNN. This step can be seen as introducing

prior knowledge to the optimization process, enhancing the algorithms' effectiveness

20
from the outset.

Here, we want to maximize the average value of the STL at the target

frequencies, which can be recast as

target target
min 𝐽 = min −[𝑡(𝑓1 , 𝐠𝐩) + 𝑡(𝑓2 , 𝐠𝐩)]
𝐠𝐩 𝐠𝐩 (6)
target target
𝑠. 𝑡. 𝑡(𝑓1 , 𝐠𝐩) > 10 dB and 𝑡(𝑓2 , 𝐠𝐩) > 10 dB,
target target
where 𝑓1 and 𝑓2 are still set as 150 Hz and 250 Hz, respectively;

target target
𝑡(𝑓1 , 𝐠𝐩) and 𝑡(𝑓2 , 𝐠𝐩) are the values of the STL at the target frequencies of

the THR, whose GPs are denoted as 𝐠𝐩 = [𝑎1 , 𝑙1 , 𝑟1 , ℎ1 , 𝑎2 , 𝑙2 , 𝑟2 , ℎ2 ]. The ranges of

GPs are same with those of the datasets for model training. The objective function

target target
𝐹(𝑓1 , 𝑓2 , 𝐠𝐩) can be expressed as:

target target target target

−[𝑡(𝑓1 , 𝐠𝐩) + 𝑡(𝑓2 , 𝐠𝐩)] 𝑖𝑓 𝑡(𝑓1 , 𝐠𝐩), 𝑡(𝑓2 , 𝐠𝐩) > 10𝑑𝐵
𝐹={ target target , (7)
𝜆− [𝑡(𝑓1 , 𝒈𝒑) + 𝑡(𝑓2 , 𝐠𝐩)] 𝑒𝑙𝑠𝑒

where 𝜆 is the penalty parameter, which is expected to help the real resonant

target target
frequencies of the individuals 𝑓1 and 𝑓2 to get closer to 𝑓1 and 𝑓2 during the

evolution. Therefore, the penalty parameter 𝜆 can be set as:

target target
10 + |𝑓1 − 𝑓1 | + |𝑓2 − 𝑓2 |, 𝑖𝑓 101 ≤ 𝑓1 , 𝑓2 ≤ 600
𝜆={ . (8)
1000, 𝑒𝑙𝑠𝑒
target target
The objective function 𝐹(𝑓1 , 𝑓2 , 𝐠𝐩) was minimized through the GA, which

was realized by an open-source evolutionary algorithm toolbox and framework for

Python, Geatpy [57]. The soea_SEGA module of Geatpy with the default algorithmic

parameters was applied in this task.

21
Figure 6: The optimization effect through the GA with and without prior knowledge. (a) Comparison of the
decreasing process of the objective function 𝐹 under the two conditions (with prior knowledge, blue solid line;
without prior knowledge, yellow dotted line). (b) Comparison of the STL spectra of the optimal structures under
the two conditions (with prior knowledge, blue solid line; without prior knowledge, yellow dotted line).

The optimizations were conducted using GA under two different initial

conditions. In the first scenario, the initial population comprised 50 individuals, all

randomly generated, indicating that we started with no prior knowledge about this

task. In the second scenario, the initial population was a mix of 45 randomly

generated individuals and 5 high-performing elitist individuals. These elitist

individuals, with resonant frequencies of 150 Hz and 250 Hz, were generated by our

proposed DNN model. Each elitist individual can be obtained through a simple query

based on the DNN that takes just a few milliseconds. To ensure a fair comparison, all

other algorithmic parameters were kept consistent. The algorithm reached

convergence after 50 generations. The comparative effectiveness of the optimizations

under these two conditions is illustrated in Figure 6. Figure 6a shows the comparison

of the decreasing process of the objective function 𝐹. Considering that the initial

population with prior knowledge includes a certain number of elitist individuals, the

blue solid line converges faster than the yellow dotted line. Figure 6b shows the
22
comparison of the STL spectra of the optimal structures under the two conditions. It

can be seen that the average STL at the target resonant frequencies with prior

knowledge (14.2 dB) is higher than the average STL at the target resonant frequencies

without prior knowledge (9.6 dB). The above results demonstrate that a good initial

condition can not only effectively accelerate convergence velocity, but also improve

the results of the optimization. Therefore, the DNN-aided optimization method is

feasible and more efficient for optimizing the acoustic structure.

3.4 On-demand design of the acoustic filter

In practical applications, we need to filter out the multi-frequency line-spectrum

noise in the background environment, and even achieve broadband sound insulation in

a certain frequency range, requiring us to design a combined filter for specific noise

frequencies. The traditional design method needs to perform the optimization for each

noise frequency, so the efficiency is very low. In comparison, it takes only a few

seconds to complete the design process using our design strategy. Here, we

demonstrate an example of designing an acoustic filter to decrease the line-spectrum

noise at four frequencies: 150 Hz, 200 Hz, 250 Hz and 300 Hz. To realize four

resonant frequencies, the acoustic filter is a combination of two THRs. The two THRs

were designed using the DNN model, with the photo and GPs shown in Figure 7a.

The STL spectra of the acoustic filter are shown in Figure 7b, where the theoretical

(blue dotted line, calculated by the TMM), simulated (orange solid line, calculated by

the FEM) and experimental (green circles) results are consistent (see Appendix A for
23
details). There are four transmission loss peaks at 150 Hz, 200 Hz, 250 Hz and 300

Hz, corresponding to the four resonant modes of the two THRs. These results confirm

the design is very precise.

To evaluate the effectiveness of the acoustic filter, a pure voice mixed with noise

at the above frequencies impinges from the left port of a square tube. A microphone is

used to receive the signal at the right port of the waveguide. We performed the

measurements with and without acoustic filter as a side branch of the tube. The time-

domain waveform and the spectrogram of the signal filtered by the acoustic filter are

shown in Figure 7d and f, respectively. For comparison, the time-domain waveform

and the spectrogram of the unfiltered signal are shown in Figure 7c and e, respectively.

It can be seen that the proposed acoustic filter significantly decreases the noise energy

and improves the speech clarity of the original signal (the sound before and after

filtering can be heard in the Supporting Video). And the experimental results prove

that our approach can achieve the on-demand design of acoustic filter for many

applications, such as noise reduction of the engine, helicopter and UHV transformers.

24
Figure 7: Structure and experimental results of the acoustic filter. (a) Photo and GPs of the
acoustic filter. Left inset: The GPs of the THR whose resonant frequencies are 150 Hz and 250 Hz.
Right inset: The GPs of the THR whose resonant frequencies are 200 Hz and 300 Hz. (b) STL spectrum
of the acoustic filter. (c) and (d) are the time-domain waveforms of the test signal before and after
filtering, respectively. (e) and (f) are spectrograms of the test signal before and after filtering,
respectively.

4. Discussion

In this paper, we demonstrate a DNN model for acoustic structure design based

solely on desired acoustic properties. The trained model can determine the

corresponding relation between the EEPs and acoustic properties, which shows a very
25
accurate prediction of the geometry of acoustic structures with multiple strong-

coupling parameters. For the multivalued inverse design problem, we can find one of

the satisfied solutions very quickly using the proposed model. Combined with the

dimension reduction algorithm, the proposed model also can be used to search for

multiple solutions. Moreover, the proposed model is able to aid evolutionary

algorithms in completing the optimization task more efficiently. The effectiveness of

the structures designed by our approach was evaluated through acoustic experiments.

As mentioned before, the one-to-many relationship between target acoustic

functionality and acoustic structures presents a big challenge to DNN building and

training. To overcome this, LPT is introduced to the model building, and the proposed

model is trained to learn EEPs instead of GPs of the acoustic structures. The

combination with LPT is an effective way to avoid using complex network, which

needs more training samples and is more difficult to be trained well, to overcome the

one-to-many effect [22,58]. Moreover, the proposed model is not limited to solve the

inverse design problem of THR. It is able to be easily extended to other acoustic

structures, whose equivalent circuit is similar to the THR. Notwithstanding this

limitation, the proposed method is certainly more flexible compared with learning the

GPs directly as the previous studies.

In conclusion, the proposed deep learning approach shows a significant

improvement in efficiency, an acceleration of the design process and an obvious

reduction in both computational and man-powered resources. The trained model

26
would be an effective design tool for designers, especially for layman users who have

less professional knowledge about acoustics. It should be noted that although this

work focuses on inverse design of acoustic structures, the proposed method, which

solves the one-to-many multivalued problem through learning equivalent parameters,

can also be extended to other research fields where the equivalent circuit approach is

frequently used, such as the design of the mechanical, piezoelectric and microwave

devices [59–61]. These good generalization performances of the proposed model

promise a wide range of potential applications.

Conflict of Interest

The authors declare no competing interests.

Authors’ Contributions

J.Y. planned, coordinated, and supervised the project. X.S., Y.Y. and H.J.

conceived the model. X.S. designed the algorithm and performed the theoretical and

numerical calculations. X.S. and H.Z. participated in the experiments and data

analysis. X.S., Y.Y. and H.J. jointly wrote the manuscript. All authors contributed to

discussing the results and reading the manuscript.

Acknowledgements

This work is supported by the Key-Area Research and Development Program of

27
Guangdong Province (Grant No. 2020B010190002), the National Natural Science

Foundation of China (Grant Nos. 11874383 and 12104480), the IACAS Frontier

Exploration Project (Grant No. QYTS202110).

Appendix A. Methods of numerical simulations and experimental measurements

A.1. Numerical simulations.

In this paper, FEM simulations were performed to verify the feasibility of our

design by using the pressure acoustic module and thermoviscous acoustic module of

COMSOL Multiphysics. Plane wave radiation boundary condition is set on the left

side and right side of the calculated fluid domain of the tube. The mesh type is the

tetrahedral mesh, and the largest mesh element size was smaller than 1/6 of the

shortest incident wavelength, and the further refined meshes were applied in the

cylindrical necks. The DNN model is built by PyTorch in Python 3.6. The integrated

development environment is Anaconda 3 in Linux operation system. The GPU

configuration is NVIDIA GeForce GTX 1080Ti.

A.2. Experimental measurements.

The experimental THR samples were fabricated using 3D printing technology

with a wall thickness of 5 mm. The material used for the samples is Lasty-KS, a type

of UV-curable resin, with density of 1.13 g/cm3. The acoustic impedance of the

Lasty-KS is much larger than the acoustic impedance of air, so the wall of the THR

can be regarded as rigid for the sound wave.

28
 The layout of the measurement system is shown in Figure 7a. The two THRs

were placed as a side branch of a square tube, where the connection was sealed by

plasticene. The STL spectrum was measured using four 1/4-inch microphones (Brüel

& Kjær, type-4187). The experimental measurements were conducted by the two-load

method in a square standing wave tube, where the hard wall and acoustic absorbing

sponge (as an anechoic boundary) were separately used at the terminal. When the

mixed signals shown in Figure 7c and d were measured, a pure voice mixed with

noise impinged from the left port of the square tube. A 1/2-inch microphone (Brüel &

Kjær, type-4189) was placed at the right port of the tube to receive the signal, while

an acoustic absorbing sponge was used as the anechoic boundary at the right port.

References

[1] Y. Yang, H. Jia, Y. Bi, H. Zhao, J. Yang, Experimental Demonstration of an

Acoustic Asymmetric Diffraction Grating Based on Passive Parity-Time-Symmetric
Medium, Phys Rev Appl. 12 (2019) 034040.
https://doi.org/10.1103/PhysRevApplied.12.034040.

[2] Y. Yang, H. Jia, S. Wang, P. Zhang, J. Yang, Diffraction control in a non-Hermitian

acoustic grating, Appl. Phys. Lett. 116 (2020) 213501.
https://doi.org/10.1063/5.0004104

[3] Y. Bi, H. Jia, Z. Sun, Y. Yang, H. Zhao, J. Yang, Experimental demonstration of

three-dimensional broadband underwater acoustic carpet cloak, Appl. Phys. Lett. 112
(2018) 223502. https://doi.org/10.1063/1.5026199

[4] Z. Sun, X. Sun, H. Jia, Y. Bi, J. Yang, Quasi-isotropic underwater acoustic carpet
cloak based on latticed pentamode metafluid, Appl. Phys. Lett. 114 (2019) 094101.
https://doi.org/10.1063/1.5085568.

[5] N. Wang, C. Zhou, S. Qiu, S. Huang, B. Jia, S. Liu, J. Cao, Z. Zhou, H. Ding, J.
Zhu, Y. Li, Meta-silencer with designable timbre, Int. J. Extreme Manuf. 5 (2023)
025501. https://doi.org/10.1088/2631-7990/acbd6d.
29
[6] X. Sun, H. Jia, Z. Zhang, Y. Yang, Z. Sun, J. Yang, Sound Localization and
Separation in 3D Space Using a Single Microphone with a Metamaterial Enclosure,
Adv. Sci. 7 (2020) 1902271. https://doi.org/10.1002/advs.201902271.

[7] Z. Sun, Y. Shi, X. Sun, H. Jia, Z. Jin, K. Deng, J. Yang, Underwater acoustic
multiplexing communication by pentamode metasurface, J. Phys. D: Appl. Phys. 54
(2021) 205303. https://doi.org/10.1088/1361-6463/abe43e.

[8] F. Ma, Y. Xu, J.H. Wu, Modal displacement method for extracting the bending
wave bandgap of plate-type acoustic metamaterials, Appl. Phys. Express 12 (2019)
074004. https://doi.org/10.7567/1882-0786/ab27dd.

[9] G. Hinton, L. Deng, D. Yu, G.E. Dahl, A. Mohamed, N. Jaitly, A. Senior, V.

Vanhoucke, P. Nguyen, T.N. Sainath, Deep neural networks for acoustic modeling in
speech recognition: The shared views of four research groups, IEEE Signal Process.
Mag. 29 (2012) 82–97. https://doi.org/10.1109/MSP.2012.2205597

[10] A. Krizhevsky, I. Sutskever, G.E. Hinton, ImageNet Classification with Deep

Convolutional Neural Networks, Commun. ACM 60 (2017) 84–90.
https://doi.org/10.1145/3065386

[11] K. Cho, B. van Merriënboer, C. Gulcehre, D. Bahdanau, F. Bougares, H.

Schwenk, Y. Bengio, Learning Phrase Representations using RNN Encoder–Decoder
for Statistical Machine Translation, Proceedings of the 2014 Conference on Empirical
Methods in Natural Language Processing (EMNLP), 2014: pp. 1724–1734.
https://doi.org/10.3115/v1/D14-1179.

[12] D. Silver, A. Huang, C.J. Maddison, A. Guez, L. Sifre, G. Van Den Driessche, J.
Schrittwieser, I. Antonoglou, V. Panneershelvam, M. Lanctot, Mastering the game of
Go with deep neural networks and tree search, Nature 529 (2016) 484–489.
https://doi.org/10.1038/nature16961

[13] R. Socher, D. Chen, C.D. Manning, A. Ng, Reasoning With Neural Tensor
Networks for Knowledge Base Completion, Advances in Neural Information
Processing Systems, Curran Associates, Inc., 2013.
https://proceedings.neurips.cc/paper_files/paper/2013/file/b337e84de8752b27eda3a12
363109e80-Paper.pdf.

[14] Sanchez-Lengeling Benjamin, Aspuru-Guzik Alán, Inverse molecular design

using machine learning: Generative models for matter engineering, Science 361 (2018)
360–365. https://doi.org/10.1126/science.aat2663.

[15] G.B. Goh, N.O. Hodas, A. Vishnu, Deep learning for computational chemistry, J.
Comput. Chem. 38 (2017) 1291–1307. https://doi.org/10.1002/jcc.24764

[16] T. Zahavy, A. Dikopoltsev, D. Moss, G.I. Haham, O. Cohen, S. Mannor, M.

Segev, Deep learning reconstruction of ultrashort pulses, Optica 5 (2018) 666–673.
30
https://doi.org/10.1364/OPTICA.5.000666

[17] P. Baldi, P. Sadowski, D. Whiteson, Searching for exotic particles in high-energy

physics with deep learning, Nat. Commun. 5 (2014) 4308.
https://doi.org/10.1038/ncomms5308.

[18] J. Carrasquilla, R.G. Melko, Machine learning phases of matter, Nat. Phys. 13
(2017) 431–434. https://doi.org/10.1038/nphys4035

[19] A. White, P. Khial, F. Salehi, B. Hassibi, A. Hajimiri, A Silicon photonics

computational Lensless Active-flat-optics imaging System, Sci. Rep. 10 (2020) 1689.
https://doi.org/10.1038/s41598-020-58027-1

[20] J. Weng, Y. Ding, C. Hu, X.-F. Zhu, B. Liang, J. Yang, J. Cheng, Meta-neural-
network for real-time and passive deep-learning-based object recognition, Nat.
Commun. 11 (2020) 6309. https://doi.org/10.1038/s41467-020-19693-x.

[21] Y. Rivenson, Z. Göröcs, H. Günaydin, Y. Zhang, H. Wang, A. Ozcan, Deep

learning microscopy, Optica 4 (2017) 1437–1443.
https://doi.org/10.1364/OPTICA.4.001437

[22] I. Malkiel, M. Mrejen, A. Nagler, U. Arieli, L. Wolf, H. Suchowski, Plasmonic

nanostructure design and characterization via Deep Learning, Light Sci. Appl. 7 (2018)
60. https://doi.org/10.1038/s41377-018-0060-7.

[23] A. Sheverdin, F. Monticone, C. Valagiannopoulos, Photonic Inverse Design with

Neural Networks: The Case of Invisibility in the Visible, Phys Rev Appl. 14 (2020)
024054. https://doi.org/10.1103/PhysRevApplied.14.024054.

[24] J. Peurifoy, Y. Shen, L. Jing, Y. Yang, F. Cano-Renteria, B.G. DeLacy, J.D.

Joannopoulos, M. Tegmark, M. Soljačić, Nanophotonic particle simulation and
inverse design using artificial neural networks, Sci. Adv. 4 (2018) eaar4206.
https://doi.org/10.1126/sciadv.aar4206

[25] T. Qiu, X. Shi, J. Wang, Y. Li, S. Qu, Q. Cheng, T. Cui, S. Sui, Deep Learning: A
Rapid and Efficient Route to Automatic Metasurface Design, Adv. Sci. 6 (2019)
1900128. https://doi.org/10.1002/advs.201900128.

[26] W. Ma, F. Cheng, Y. Xu, Q. Wen, Y. Liu, Probabilistic Representation and Inverse
Design of Metamaterials Based on a Deep Generative Model with Semi‐Supervised
Learning Strategy, Adv. Mater. 31 (2019) 1901111.
https://doi.org/10.1002/adma.201901111.

[27] W. Ma, Z. Liu, Z.A. Kudyshev, A. Boltasseva, W. Cai, Y. Liu, Deep learning for
the design of photonic structures, Nat. Photonics 15 (2021) 77–90.
https://doi.org/10.1038/s41566-020-0685-y.

[28] Y.-T. Luo, P.-Q. Li, D.-T. Li, Y.-G. Peng, Z.-G. Geng, S.-H. Xie, Y. Li, A. Alù, J.
31
Zhu, X.-F. Zhu, Probability-Density-Based Deep Learning Paradigm for the Fuzzy
Design of Functional Metastructures, Research 2020 (2020) 1–11.
https://doi.org/10.34133/2020/8757403.

[29] M.J. Bianco, P. Gerstoft, J. Traer, E. Ozanich, M.A. Roch, S. Gannot, C.-A.
Deledalle, Machine learning in acoustics: Theory and applications, J. Acoust. Soc. Am.
146 (2019) 3590–3628. https://doi.org/10.1121/1.5133944.

[30] W.W. Ahmed, M. Farhat, X. Zhang, Y. Wu, Deterministic and probabilistic deep
learning models for inverse design of broadband acoustic cloak, Phys. Rev. Res. 3
(2021) 013142. https://doi.org/10.1103/PhysRevResearch.3.013142.

[31] H. Ding, X. Fang, B. Jia, N. Wang, Q. Cheng, Y. Li, Deep Learning Enables
Accurate Sound Redistribution via Nonlocal Metasurfaces, Phys. Rev. Appl. 16 (2021)
064035. https://doi.org/10.1103/PhysRevApplied.16.064035.

[32] Z. Zeng, M. Zhang, C. Li, L. Ren, P. Wang, li jiawei_hust, D.W. Yang, Y. Pan,
Simulation study on characteristics of acoustic metamaterials based on Mie and
Helmholtz resonance for low-frequency acoustic wave control, J. Phys. D: Appl. Phys.
(2021). https://doi.org/10.1088/1361-6463/ac0ad1.

[33] I. Davis, A. McKay, G.J. Bennett, A graph-theory approach to optimisation of an

acoustic absorber targeting a specific noise spectrum that approaches the causal
optimum minimum depth, J. Sound Vib. 505 (2021) 116135.
https://doi.org/10.1016/j.jsv.2021.116135.

[34] J. Esteves, L. Rufer, D. Ekeom, S. Basrour, Lumped-parameters equivalent

circuit for condenser microphones modeling, J. Acoust. Soc. Am. 142 (2017) 2121–
2132. https://doi.org/10.1121/1.5006905.

[35] Y. Cheng, J.Y. Xu, X.J. Liu, One-dimensional structured ultrasonic metamaterials
with simultaneously negative dynamic density and modulus, Phys. Rev. B 77 (2008)
045134. https://doi.org/10.1103/PhysRevB.77.045134

[36] N. Fang, D. Xi, J. Xu, M. Ambati, W. Srituravanich, C. Sun, X. Zhang,

Ultrasonic metamaterials with negative modulus, Nat. Mater. 5 (2006) 452–456.
https://doi.org/10.1038/nmat1644

[37] Y. Li, C. Shen, Y. Xie, J. Li, W. Wang, S.A. Cummer, Y. Jing, Tunable
asymmetric transmission via lossy acoustic metasurfaces, Phys. Rev. Lett. 119 (2017)
035501. https://doi.org/10.1103/PhysRevLett.119.035501

[38] N. Jiménez, T.J. Cox, V. Romero-García, J.-P. Groby, Metadiffusers: Deep-

subwavelength sound diffusers, Sci. Rep. 7 (2017) 5389.
https://doi.org/10.1038/s41598-017-05710-5

[39] X. Yang, J. Yin, G. Yu, L. Peng, N. Wang, Acoustic superlens using Helmholtz-
32
resonator-based metamaterials, Appl. Phys. Lett. 107 (2015) 193505.
https://doi.org/10.1063/1.4935589.

[40] J. Xia, X. Zhang, H. Sun, S. Yuan, J. Qian, Y. Ge, Broadband tunable acoustic
asymmetric focusing lens from dual-layer metasurfaces, Phys. Rev. Appl. 10 (2018)
014016. https://doi.org/10.1103/PhysRevApplied.10.014016

[41] Y. Liu, W. Zhang, G. Cao, G. Zuo, C. Liu, F. Ma, Ultra-thin ventilated

metasurface pipeline coating for broadband noise reduction✰, Thin-Walled Struct.
200 (2024) 111916. https://doi.org/10.1016/j.tws.2024.111916.

[42] H. Long, Y. Cheng, X. Liu, Reconfigurable sound anomalous absorptions in

transparent waveguide with modularized multi-order Helmholtz resonator, Sci. Rep. 8
(2018) 1–9. https://doi.org/10.1038/s41598-018-34117-z

[43] Y. Yu, H. Jia, Y. Yang, H. Zhao, Q. Shi, P. Kong, J. Yang, K. Deng, Multi-order
resonators for acoustic multiband asymmetric absorption and reflection, J. Appl. Phys.
131 (2022) 135102. https://doi.org/10.1063/5.0084450.

[44] X. Sun, H. Jia, Y. Yang, J. Yang, Low-frequency Broadband Sound Insulation

Device Design Method Based on Deep Learning, J. Appl. Acoust. 42 (2023) 611–619.

[45] Q. Shi, Y. Yang, Z. Mei, Y. Lin, X. Li, P. Tian, P. Kong, H. Jia, J. Yang, K. Deng,
Design and demonstration of composite mufflers based on dissipative and reactive
units, Eng. Res. Express 5 (2023) 045029. https://doi.org/10.1088/2631-8695/ad03ae.

[46] C.R. Liu, J.H. Wu, X. Chen, F. Ma, A thin low-frequency broadband metasurface
with multi-order sound absorption, J. Phys. D: Appl. Phys. 52 (2019) 105302.
https://doi.org/10.1088/1361-6463/aafaa3.

[47] A.D. Pierce, Acoustics: An Introduction to Its Physical Principles and

Applications, Springer International Publishing, Cham, 2019.
https://doi.org/10.1007/978-3-030-11214-1.

[48] D.P. Kingma, J. Ba, Adam: A method for stochastic optimization, ArXiv preprint,
ArXiv14126980 (2014). https://doi.org/10.48550/arXiv.1412.6980

[49] S. Ioffe, C. Szegedy, Batch normalization: Accelerating deep network training by

reducing internal covariate shift, ArXiv preprint. ArXiv150203167 (2015).
https://doi.org/10.48550/arXiv.1502.03167

[50] N. Srivastava, G. Hinton, A. Krizhevsky, I. Sutskever, R. Salakhutdinov, Dropout:

a simple way to prevent neural networks from overfitting, J. Mach. Learn. Res. 15
(2014) 1929–1958.

[51] X. Sun, Y. Yang, H. Jia, J. Yang, Physics-aware training for the physical machine
learning model building, The Innovation 3 (2022).
https://doi.org/10.1016/j.xinn.2022.100287.
33
[52] K. Pearson, LIII. On lines and planes of closest fit to systems of points in space,
Lond. Edinb. Dublin Philos. Mag. J. Sci. 2 (1901) 559–572.
https://doi.org/10.1080/14786440109462720

[53] M. Yang, S. Chen, C. Fu, P. Sheng, Optimal sound-absorbing structures, Mater

Horiz 4 (2017) 673–680. https://doi.org/10.1039/C7MH00129K.

[54] K.S. Tang, K.F. Man, S. Kwong, Q. He, Genetic algorithms and their applications,
IEEE Signal Process. Mag. 13 (1996) 22–37. https://doi.org/10.1109/79.543973.

[55] T. Wang, S. Li, S.R. Nutt, Optimal design of acoustical sandwich panels with a
genetic algorithm, Appl. Acoust. 70 (2009) 416–425.
https://doi.org/10.1016/j.apacoust.2008.06.003.

[56] J. Kennedy, R. Eberhart, Particle swarm optimization, in: Proc. ICNN95 - Int.
Conf. Neural Netw., IEEE, 1995, pp. 1942–1948 vol.4.
https://doi.org/10.1109/ICNN.1995.488968.

[57] Jazzbin et al., geatpy: The genetic and evolutionary algorithm toolbox with high
performance in Python, GitHub repository, 2020. https://github.com/geatpy-
dev/geatpy.

[58] D. Liu, Y. Tan, E. Khoram, Z. Yu, Training Deep Neural Networks for the Inverse
Design of Nanophotonic Structures, ACS Photonics 5 (2018) 1365–1369.
https://doi.org/10.1021/acsphotonics.7b01377.

[59] Y. Yang, B. Yang, Equivalent circuit method based on complete magneto-

mechanical coupling magnetostriction parameters for fixed magnetoelectric
composites, Int. J. Mech. Sci. 199 (2021) 106411.
https://doi.org/10.1016/j.ijmecsci.2021.106411.

[60] A. Dorodnyy, S.M. Koepfli, A. Lochbaum, J. Leuthold, Design of CMOS-

compatible metal–insulator–metal metasurfaces via extended equivalent-circuit
analysis, Sci. Rep. 10 (2020) 17941. https://doi.org/10.1038/s41598-020-74849-5.

[61] A. Khajevandi, H. Oraizi, Utilizing interdigital and supershape geometries for the
design of frequency selective surfaces with high angular and polarization stabilities,
Sci. Rep. 12 (2022) 7054. https://doi.org/10.1038/s41598-022-10960-z.

34