Professional Documents
Culture Documents
Optimal Design OfMultiresonant Layered Acoustic Metamaterials - Cante - Oliver
Optimal Design OfMultiresonant Layered Acoustic Metamaterials - Cante - Oliver
Engineering Structures
journal homepage: www.elsevier.com/locate/engstruct
Keywords: Broadband sound attenuation at low frequency ranges (below 500 Hz) has been a challenge in the acoustics
Acoustic metamaterials field which cannot be solved, via conventional materials, unless impractical amounts of mass are employed.
Coupled resonances Multiresonant Layered Acoustic Metamaterials (MLAM) offer exceptional attenuating properties at lower
MLAM
frequencies, through novel coupled resonances mechanisms, in a layered configuration that make them
Computational homogenization
amenable for large-scale manufacturing. To show the potential capabilities of MLAM, a novel computational
Genetic algorithm optimization
design strategy has been developed to optimize the metamaterials’ performance in terms of their Sound
Transmission Loss (STL). First, a multiscale homogenization framework specifically derived for MLAM allows
an accurate and extremely fast evaluation of their STL response to normal-incidence acoustic waves in the
frequency range of interest. Then, the MLAM design is parameterized into a set of relevant geometric features,
which are optimized by means of an optimization scheme based on standard genetic algorithms combined
with the homogenization model. The results demonstrate how this design strategy is a powerful tool to obtain
optimal MLAM panel designs subject to constraints imposed by the application, for instance, in terms of weight
or thickness of the panel, or the manufacturing process (e.g. geometric tolerances).
∗ Correspondence to: Escola Superior d’Enginyeries Industrial Aeroespacial i Audiovisuals de Terrassa (ESEIAAT), Universitat Politècnica de Catalunya -
BarcelonaTech (UPC), Campus Terrassa UPC C/ Colom 11, 08222 Terrassa, Spain.
E-mail addresses: david.roca.cazorla@upc.edu, droca@cimne.upc.edu (D. Roca).
https://doi.org/10.1016/j.engstruct.2023.116555
Received 17 November 2022; Received in revised form 10 June 2023; Accepted 30 June 2023
Available online 22 July 2023
0141-0296/© 2023 The Author(s). Published by Elsevier Ltd. This is an open access article under the CC BY license (http://creativecommons.org/licenses/by/4.0/).
G. Sal-Anglada et al. Engineering Structures 293 (2023) 116555
acoustic metamaterial configuration consists of attaching resonating that provides an optimized MLAM panel design is presented in Sec-
cylindrical structures (typically of a soft material) to a plate [10– tion 4. Results show the potential of MLAM-based panels to improve,
13]. Other works have exploited mechanical features of the resonators more than 20 dB, the sound attenuation in broadband frequency ranges
to produce active metamaterials capable of controlling their response below 500 Hz, when compared to conventional (homogeneous) panels
via band gaps tuning [14,15]. More recently, given the advances in of the same weight.
additive manufacturing techniques, 3D-printed acoustic metamaterial
configurations based on purely geometric features have gained popular- 2. MLAM computational model
ity [16–20]. In general, while these are suitable to produce prototypes,
the intricate geometries involved, and/or the required tolerances (to 2.1. MLAM design concept
tackle low-frequency ranges) in some design features, present an issue
for scaling-up production, thus limiting their feasibility for practical The MLAM is configured by a stack of different layers, each one
applications. characterized by (1) a single material, (2) a thickness that remains
Nowadays, the acoustic metamaterials technology has reached a constant in the out-of-plane direction, and (3) a pattern repeated pe-
level of maturity high enough on paper, but there is still a lack of riodically in the in-plane directions (see Fig. 1-b). The specific choice
feasible solutions that enable its introduction into a large-scale in- of these features for each layer gives them different functions within
the whole system. In its core, the MLAM is composed of two resonat-
dustrial market. This is due to two main challenges. The first one
ing layers joined by a connecting layer in-between. The role of the
is related to manufacturing issues —namely, most of the discussed
resonating layers is to trigger local resonance effects, at two different
realizations are effective in terms of performance for their intended
frequencies, through the vibration of internal features of the unit cell
applications, but their fabrication relies on either complex processes
geometric design. The connecting layer is responsible for coupling these
(for instance, assembling several different material parts into small
resonances, which typically requires a small enough stiffness in the
unit cells), or additive manufacturing technologies (which are still far
wave propagation direction. Additionally, in order to provide space for
from competing in the mass-production industry due to time and cost
vibration to the resonating layers, separating layers are also part of the
limitations). The second challenge is related to performance. Resonance
overall MLAM design. Finally, the MLAM core needs to be attached to a
phenomena are, intrinsically, localized to narrow frequency bands—
wall-like structure or, alternatively, be covered by additional skin layers
especially in the low-frequency range. Standard mechanisms to increase
at both ends (creating a standalone insulation panel).
these bandwidths using metamaterials typically imply increasing the
resonating masses, which impacts the range of applicability of the 2.2. STL evaluation
design. This issue has also been tackled by manipulating the stiffness
component instead of the mass, aiming at increasing the frequency The STL response to normal-incidence acoustic waves is typically
bandwidths without changing the dimensions. In Ref. [21], the stiffness given as a measure of the attenuating capabilities of a material and,
ratio between two local resonators connected in series on an elastic therefore, it will be used to evaluate the performance of MLAM panel
plate is exploited to produce wider bandgaps based on negative stiff- designs. The analysis at oblique angles of incidence is not considered in
ness. Other works have proposed different mechanisms for tuning such this work because the main attenuation features (which are used from
bandgaps, for instance, by incorporating hollow platelets inside a soft a design point of view) arise in all cases, regardless of the angle [26].
matrix under a prestress state [22], or by exploiting thermal properties To evaluate the performance of MLAM panel designs, a plane wave
of the different components inside the metamaterial structure [23]. is assumed to propagate through air (with density 𝜌a = 1.25 kg/m3
In an attempt to overcome these two challenges, a Multiresonant and velocity 𝑐a = 340 m∕s) in the normal direction to the material’s
Layered Acoustic Metamaterial (MLAM) solution was proposed in a incidence surface (𝑥-direction). Part of this wave is reflected back to
previous work [24]. The concept relies on its layered configuration the incidence acoustic domain, while the other is transmitted to the
to make it potentially easy to manufacture through well-established other side of the material (transmission acoustic domain), preserving
techniques, for instance, a process involving lamination and die cutting. the propagation direction (see Fig. 1-a). In this context, because the skin
Additionally, it introduces a coupling mechanism (derived from the layers’ surfaces (or, equivalently, the wall-structure where the MLAM
notion of double-periodicity provided by Gao et al. [25]) that pro- core is attached) are flat, the pressure fields, 𝑝i and 𝑝t , in the acoustic
duces a double-peak response in terms of the Sound Transmission Loss (air) domains are assumed:
(STL) of the material. Essentially, this coupling effect allows to join
• In the incidence side:
two local resonance bandwidths in a whole range of uninterrupted
( )
additional attenuation (i.e., with no transmission peaks in-between). 𝑝i (𝑥, 𝑡) = 𝑃i 𝑒i(−𝜔𝑡+𝜅𝑥) + 𝑅𝑒i(−𝜔𝑡−𝜅𝑥) (1)
This double-peak STL response has been observed also in a hollow
where 𝑃i is the wave’s amplitude, 𝑡 refers to the time variable,
pillar-hollow plate configuration for sound transmission perpendicular
𝜔 is the frequency, 𝜅 = 𝜔∕𝑐a is the wavenumber, and 𝑅 is the
to the direction of periodicity [26], and the underlying physics of
reflection coefficient (fraction of the wave’s amplitude that is
its formation have been recently explored by Stein et al. [27]. While
reflected back).
the concept of MLAM was already demonstrated in Ref. [24] (both
• In the transmission side:
numerically and experimentally), in the present work, a computational
design strategy is proposed as a tool to obtain optimal MLAM-based 𝑝t (𝑥, 𝑡) = 𝑃i 𝑇 𝑒i(−𝜔𝑡+𝜅𝑥) (2)
panel configurations for different target performance and application
where 𝑇 refers to the transmission coefficient (fraction of the
constraints. First, a specific homogenization model capable of perfectly
wave’s amplitude which is transmitted through the material).
capturing the particular MLAM behaviour in terms of the STL response
to normal-incidence acoustic waves is proposed. This simple model In the solid domain, the response is driven by the balance of linear
allows extremely fast STL evaluations that are used to build a database momentum equations:
for the effective mass and stiffness properties in terms of a parametriza-
∇ ⋅ 𝝈 = 𝐩̇ (3)
tion of a proposed MLAM panel design. Finally, this database feeds an
optimization algorithm, from which an optimal MLAM panel configu- where 𝝈 denotes the stress tensor, and 𝐩̇ = 𝜌̈𝐮 is the inertial force
ration is obtained, meeting the target properties and constraints. The density (with 𝜌 being the material density and 𝐮 referring to the
computational model is described in Section 2, and the overall design ̇ is used to refer to the time derivative
displacement field). Notice that (∙)
strategy is detailed in Section 3. An example application of the model of (∙).
2
G. Sal-Anglada et al. Engineering Structures 293 (2023) 116555
Fig. 1. (a) Sound transmission loss evaluation set-up. Part of the normal-incidence acoustic waves propagating on the left side of the panel (incidence acoustic domain) is reflected
back, while the other is transmitted through the right side of the panel (transmission acoustic domain). (b) Representation of the MLAM layered structure. The MLAM core is
composed of two distinct resonating layers (layers 3 and 5), joined by a connecting layer (layer 4). Two separating layers (layers 2 and 6) are attached at both ends to allow
space for the resonators out-of-plane vibration, and the whole assembly is enclosed with skin layers (layers 1 and 7) that can be replaced by conventional wall structures.
Defining 𝛤i and 𝛤t as the interfaces between the solid and the in- (1) Considering an energetic equivalence —also known as generalized
cidence and transmission acoustic domains, respectively, compatibility Hill–Mandel principle— between the total mechanical power of
conditions in terms of displacements and external surface forces in the a point in the macroscale and the volume average value of its
𝑥-direction (denoted here by 𝑢 and 𝑓 ), will be imposed, namely corresponding microscale representative volume element (RVE)
{ counterpart
(𝑒i𝜅𝑥 − 𝑅𝑒−i𝜅𝑥 )𝑒−i𝜔𝑡 , ∀𝐱 ∈ 𝛤i
𝑢(𝐱, 𝑡) = (4) 𝐩̇ ⋅ 𝐮̇ + 𝝈 ∶ 𝜺̇ = ⟨𝐩̇ 𝜇 ⋅ 𝐮̇ 𝜇 + 𝝈 𝜇 ∶ 𝜺̇ 𝜇 ⟩𝛺𝜇 (7)
𝑇 𝑒i𝜅𝑥 𝑒−i𝜔𝑡 , ∀𝐱 ∈ 𝛤t
{
−i𝜌a 𝑐a 𝜔(𝑒i𝜅𝑥 + 𝑅𝑒−i𝜅𝑥 )𝑒−i𝜔𝑡 , ∀𝐱 ∈ 𝛤i where 𝜺 refers to the strain field. Notice that the subscript (∙)𝜇
𝑓 (𝐱, 𝑡) = (5) is used to refer to the microscale counterparts of the magni-
i𝜌a 𝑐a 𝜔𝑇 𝑒i𝜅𝑥 𝑒−i𝜔𝑡 , ∀𝐱 ∈ 𝛤t
tude (∙), and the angle brackets are considered to denote the
Notice that, for convenience, the value for the pressure waves’ ampli- volume average over the whole RVE domain, 𝛺𝜇 , i.e., ⟨(∙)⟩𝛺𝜇 =
tude has been (arbitrarily) chosen 𝑃i = −i𝜌a 𝑐a 𝜔, so that the resulting 1∕|𝛺𝜇 | ∫𝛺 (∙)d𝛺.
𝜇
system is normalized in terms of the displacements’ amplitude. (2) Imposing a set of kinematic restrictions that account for the
Then, the reflection and transmission coefficients, 𝑅 and 𝑇 , can be problem’s dynamic response. In this case
obtained, for a given frequency 𝜔, as a solution of Eq. (3) in the fre-
quency domain, upon considering Eqs. (4) as boundary conditions, and 𝐮𝜇 (𝐱′ , 𝑡) = 𝐮(𝐱, 𝑡) + 𝜺(𝐱, 𝑡) ⋅ (𝐱′ − 𝐱0′ ) + 𝐮̃ 𝜇 (𝐱′ , 𝑡) (8)
Eqs. (5) as external actions in the system. A more detailed description
where 𝐱′ denotes the spatial coordinates in the microscale, 𝐱0′ =
of the steps involved to solve the resulting system can be found in ⟨𝐱′ ⟩𝛺𝜇 is the centroid of the RVE, and 𝐮̃ 𝜇 is the microfluctuations
Roca et al. [28]. Eventually, the transmission loss, TL (in dB), can be field. For problems involving periodic unit cells, 𝐮̃ 𝜇 satisfies peri-
evaluated as odic boundary conditions in the periodicity directions. Denoting
TL(𝜔) = −20 log10 |𝑇 (𝜔)| (6) 𝛤𝜇(−) and 𝛤𝜇(+) the matching periodic boundaries of the unit cell
(see Fig. 2), then
From Eqs. (7) and (8), one can obtain the following expressions for the
To derive a simplified model capable of evaluating the STL response
effective macroscopic stress and inertial force density:
of a given MLAM design, the multiscale homogenization framework
proposed in Roca et al. [29] will be used as basis. This model is 𝝈 = ⟨𝝈 𝜇 + 𝐩̇ 𝜇 ⊗S (𝐱′ − 𝐱0′ )⟩𝛺𝜇 (10)
built upon the assumption of a separation of scales condition, so it
𝐩̇ = ⟨𝐩̇ 𝜇 ⟩𝛺𝜇 (11)
holds for settings where a lower scale (microscale) can be identified
in the domain, the characteristic size of which, 𝓁𝜇 , is significantly In Eq. (10), the symbol ⊗S
denotes the symmetric outer product,
smaller (at least by an order of magnitude) than the wavelength, 𝜆, i.e., 𝐚 ⊗S 𝐛 = (𝐚 ⊗ 𝐛 + 𝐛 ⊗ 𝐚)∕2.
of any wave propagating in the macroscale, i.e. 𝜆 ≫ 𝓁𝜇 . In its general For the problem of STL evaluation in MLAM cases, there is a
statement described in Ref. [29], the multiscale model allows to obtain fundamental distinction between the in-plane domain (where there is
homogenized properties —in terms of effective constitutive (elastic) periodicity of each layer’s unit cells), and the out-plane (thickness) di-
tensor and density— from a given unit cell (see Fig. 2). This is done rection (which is the plane-wave propagation direction). Furthermore,
by: since the coupling mechanism of MLAM is triggered by the distinct
3
G. Sal-Anglada et al. Engineering Structures 293 (2023) 116555
4
G. Sal-Anglada et al. Engineering Structures 293 (2023) 116555
for 𝑗 = {3, 5}. As in the other layers, the interface to the adjacent
layers is obtained via the outer frame of width 𝑤c .
5
G. Sal-Anglada et al. Engineering Structures 293 (2023) 116555
where TLh stands for the STL of the homogeneous acoustic insulator, 4. Results
̃
and TL opt corresponds to the STL of the optimized MLAM structure
capped by the target sound transmission loss, i.e., TL∗ (𝑓 ) = TLh (𝑓 ) + To illustrate an example implementation of the homogenization
𝛥TL, scheme combined with the two-step optimization strategy, a Multireso-
nant Layered Acoustic Metamaterial (MLAM) made of (a) a polyamide
̃ (𝑓 ) = min{TL (𝑓 ), TL∗ (𝑓 )}
TL ∀𝑓 ∈ [𝑓L , 𝑓U ] (29)
opt opt material for the skin, separating, and resonating layers, and (b) a
with TLopt referring to the TL of the optimized MLAM structure. rubber-like material for the connecting layer, has been chosen. The
Notice that the constraint Eqs. (22)–(28) have been included to properties of each material are listed in Table 1, specifying the density
the original objective function (20) in order to impose the interval of value, the Young’s modulus and the Poisson’s ratio.
interest (22)–(24), guarantee the coupling between the two resonant Regarding the MLAM dimensions, the cell size 𝑎 and the parameter
peaks (25), enforce an upper bound to the total reference mass of the 𝑑a have been set to 60 mm and 25 mm, respectively, in order to
MLAM structure (26), set an upper bound to the MLAM core’s thickness properly operate in the interval of interest, i.e., frequencies in the range
(27), and establish upper and lower bounds for each of the parameters, [100, 500] Hz, and be able to maximize the additional attenuation with a
defined by 𝜑𝑖,U and 𝜑𝑖,L , respectively, for each parameter 𝜑𝑖 in (28). coupled, double-peak sound transmission loss (STL) response. The ref-
In particular, the coupling between the resonant peaks 𝑓𝛼 and 𝑓𝛽 is erence total panel thickness, ℎref , has been set to 30 mm. Additionally,
enforced by ensuring that the norm of the second derivative of the each parameter of the chromosome 𝜳 has been accordingly bounded in
transmission loss with respect to the frequency, i.e., the curvature of the the optimization, leaving an extensive margin in the resultant effective
STL —numerically computed using a central finite difference method— properties to guarantee the coupling between the two resonant peaks
is always positive in this interval. In this norm, the angle brackets with the specified materials and dimensions. The limits for each of the
⟨(∙)⟩mc are used to express the macaulay’s function of (∙), i.e., ⟨(∙)⟩mc parameters are also specified in Table 1.
is equal to (∙) if it is positive, and 0 otherwise. Based on this set of parameters, a database of effective parameters
The resultant size optimization problem (21) represents a highly of each cell type is computed for the two-step optimization strategy,
computational expensive problem to be solved via traditional opti- as explained in the previous section. The resultant data, including
mization techniques. To overcome this computational issue, a two-step each tested chromosome and the corresponding effective properties, is
strategy has been selected: (1) an offline-step to create a database of interpolated using Matlab’s built-in griddedInterpolant function, which
(𝑗)
effective parameters (i.e., 𝐸eff , 𝜌(𝑗)
eff
, 𝑚̃ (𝑗,𝑖) , and 𝑘̃ (𝑗,𝑖) ) using the layer- provides the linear interpolant for each effective property. This step is
by-layer homogenization model from a set of parameters 𝜳 for each computed only once at the offline-step for each of the cell types, allow-
cell configuration (see Fig. 3), and (2) an online-step to optimize the ing to significantly reduce the computational cost of the optimization.
design of the MLAM structure using both (a) the database to create Consequently, multiple optimizations can be performed with different
an interpolation function and (b) a genetic algorithm method. In the bounds of frequencies 𝑓L and 𝑓U , reference surface densities 𝜇ref , total
6
G. Sal-Anglada et al. Engineering Structures 293 (2023) 116555
Fig. 5. Two-step size optimization strategy: (Offline-step) a database is first created for a set of parameters 𝜳 (listed in Eq. (19)) in the interval of interest using a structured
(𝑗)
hypercube sampling, thus obtaining the corresponding effective properties (i.e., 𝐸eff , 𝜌(𝑗)
eff
, 𝑚̃ (𝑗,𝑖) , and 𝑘̃ (𝑗,𝑖) ), and (Online-step) the optimization of the parameters is then performed
using a genetic algorithm in combination with interpolation functions of the effective parameters.
Table 1
Material properties and geometrical parameters considered for the optimization. Each material is assigned to the corresponding layer types indicated in the last column (Layers).
The (upper and lower) bounds of each parameter 𝜑𝑖 (in Eq. (19)) and the corresponding constant spacing are used to generate the interpolation function of each effective property
using Matlab’s built-in griddedInterpolant function at the offline-step.
Materials Density 𝜌 (kg/m3 ) Young’s modulus 𝐸 (MPa) Poisson’s ratio 𝜈 Layers
Polyamide (PA) 1050 1650 0.40 1, 2, 3, 5, 6, 7
Rubber-like 1060 0.15 0.47 4
𝑤(3)
𝑓
(mm) 𝑤(3)
2
(mm) 𝑤(3)
a (mm) ℎ(3) (mm) ℎ(4) (mm) 𝑤(5)
f
(mm) 𝑤(5)
2
(mm) 𝑤(5)
a (mm) ℎ(5) (mm)
Lower bound 1.5 1 1 0.5 0.5 1.5 1 1 0.5
Upper bound 5 12 8 10 10 5 12 8 10
Spacing 0.5 0.25 0.2 0.5 0.1 0.5 0.25 0.2 0.5
thicknesses ℎref , and penalty factors 𝜁, among others, at a low compu- hence the double-peak STL response. It is worth noticing that the first
tational cost. The online step starts with an initial configuration with resonant peak corresponds to the resonance of the first resonating
an arbitrarily set of parameters that satisfy the constraint equations. layer (i.e., layer 3) and the second peak is produced by the second
To simplify the selection of parameters for this initial configuration, resonating layer (i.e., layer 5). In both cases, the resonating modes are
both resonating layers are set equal, with a single resonance frequency consistent with the one-dimensional propagation of the wave, as shown
inside the target range. This allows the optimization algorithm to find in Fig. 6-b.
solutions in the frequency range of interest. The same geometrical parameters (detailed in Table 2) have been
used to assess the performance of the layer-by-layer homogenization,
4.1. MLAM optimization detailed in Section 2.3. In this sense, the optimized STL response
(obtained via the interpolation step in the optimization scheme) has
The results concerning the size optimization of the MLAM core
been compared with the one obtained from directly applying the 1D
for 𝛥TL = 20 dB are illustrated in Fig. 6-a, including the STL of
homogenization scheme for the exact optimized parameters. The good
the initial configuration (dotted blue line), the optimized one (solid
agreement between both curves shown in Fig. 6-b is considered as
blue line). The STL response is also compared with the corresponding
a validation of the interpolation step involved in the optimization.
results for the reference homogeneous material of same surface density
Additionally, both STL responses have been compared with the one ob-
of 15 kg/m2 (dash-dotted green line). From the results, it can be
easily noticed that the optimized solution corresponds to a high-band tained from a direct numerical simulation (DNS) of the corresponding
gap coupled response which attenuates more than 20 dB for almost complete 3D MLAM (also in Fig. 6-b). In this case, the results show
300 Hz. Despite starting from the same reference surface density 𝜇ref , an exceptional correlation for frequencies lower than 600 Hz. Above
the surface density of the optimized MLAM structure is smaller than this value, additional 3D effects neglected in the homogenization model
the one for the homogeneous material while attenuating a significantly assumptions perturb the overall behaviour of the MLAM, thus limiting
higher quantity in the target frequency range. Consequently, one can its applicability at higher frequency ranges. This loss in accuracy is
be confident that the ga optimization is performing as expected. In this related to the separation of scales hypothesis, which starts to fail at
context, the optimizer attempts to adjust the second resonant frequency higher frequencies. In general, for typical unit cell sizes and material
𝑓𝛽 as high as possible, i.e., near the upper frequency bound 𝑓U , while properties (in the same order of magnitude of the ones considered
pulling the lowest frequency 𝑓𝛼 as far away (towards lower frequen- here) enough accuracy is achieved for the target frequency range below
cies) as possible maintaining the coupling between both resonances, 500 Hz.
7
G. Sal-Anglada et al. Engineering Structures 293 (2023) 116555
Table 2
Numerical values used for the initial MLAM configuration and the corresponding values obtained for the optimized one once the ga optimization has been performed, specifying
the surface density and the total thickness of each structure. In this sense, the surface density of the optimized structure decreases from 15 kg/m2 up to 13.65 kg/m2 , while
achieving higher additional attenuation areas with larger frequency bandwidths.
Configuration 𝑤(3)
f
(mm) 𝑤(3)
2
(mm) 𝑤(3)
a (mm) ℎ(3) (mm) ℎ(4) (mm) 𝑤(5)
f
(mm) 𝑤(5)
2
(mm) 𝑤(5)
a (mm) ℎ(5) (mm) 𝜇ref (kg/m2 ) ℎcore (mm)
Initial 1.5 2.1 1 7.3 1 1.5 2.2 1 7.3 15 19.6
Optimized 1.57 1 1.1 4.95 9.98 5 3.72 1.29 8 13.65 26.9
(solid blue line), the homogenized version obtained from applying the
layer-by-layer homogenization directly on the optimal set of parameters
(dashed blue line), and the DNS simulation of the same case (blue
dots). As aforementioned, the length of the attachments, 𝑑a , depends
on the cell size 𝑎, being set to 22.5 mm, 25 mm, 27.5 mm, and 30 mm,
respectively. Regarding the penalized term, the same previously speci-
fied initial penalty factor 𝜁0 and penalty ratio 𝐶𝜁 values are considered
regardless the cell size 𝑎.
At first look, one can observe that the four optimized MLAM config-
urations tend to a similar optimized solution, obtaining almost the same
MLAM attenuation window area 𝐴𝛥TL while showing higher (though
small) discrepancies in the double-peak frequency bandwidth 𝛥𝑓 . In
particular, 𝛥𝑓 presents a concave character with respect to the cell size
𝑎, showing a maximum of 242 Hz for 𝑎 = 70 mm. As already discussed in
the previous section for 𝑎 = 60 mm, the second resonant frequency 𝑓𝛽 is
designed to be near the upper frequency bound 𝑓U = 500 Hz, while the
first one is minimized so that the second term in Eq. (20) is maximized.
This may be due to the fact that higher resonant frequencies are easier
to achieve than lower ones while satisfying all the constraint equations.
In this context, lower frequencies may be much more constrained by
mass and geometrical dimensions, thus making it more difficult for
the optimizer to find alternative coupled solutions at lower frequencies
with better objective function values. Nevertheless, up to a lower cell
size limit for which local resonances in the target range can no longer
be achieved with the material properties considered, similar optimized
solutions can be found regardless of the cell size, thus proving that the
optimizer is properly searching the optimal set of parameters for this
application.
In addition, it is worth to stress out that all optimized solutions
exhibit a lower surface density than the reference homogeneous insu-
lator (of 15 kg/m2 ), in the range between 12.5 and 13.7 kg/m2 for
the different cell sizes. Despite the lower surface density, the optimized
MLAM configurations achieve an additional attenuation of 𝛥TL = 20 dB
for more than 300 Hz, and an attenuation of 𝛥TL = 10 dB for almost
350 Hz. Furthermore, the total thickness of the optimized MLAM panel
also varies with the cell size 𝑎, leading to solutions with a thicknesses
that range from 27 to 30 mm.
This set of optimizations is also used to proof the correlation of the
Fig. 6. (a) Sound transmission loss curves obtained for the MLAM panel with a cell layer-by-layer homogenization and the interpolation step with the cor-
size 𝑎 = 60 mm. The optimized STL response (solid blue line) is compared with the responding DNS for different cell size values. For the lengths considered
ones corresponding to the initial configuration (dashed dark blue line) and to the in this work, and frequencies up to 600 Hz, one can be more than
homogeneous acoustic insulator (dashed–dotted green line). The initial and optimized convinced that the homogenization scheme is capturing the double-
MLAM attenuation window areas are shaded in darker and lighter blue, respectively.
The resulting resonating layers topology for the optimized MLAM are also depicted near
peak resonance phenomena without major problems and within a small
the corresponding resonant frequencies 𝑓𝛼 = 263 Hz and 𝑓𝛽 = 495 Hz. (b) Optimized margin of error with respect to the DNS computation. Additionally, the
STL response obtained from the interpolation step (solid blue line) compared to the interpolated STL (solid blue line) almost resembles the homogenized
one obtained from directly applying the homogenization scheme to the optimal set of response (dashed blue line) for all these cases, only noting marginal
parameters (dashed blue line). The blue circles show the STL response obtained from
differences among both curves. These discrepancies are proportional
the corresponding DNS analysis. The modes corresponding to the resonance peaks are
depicted as well. (For interpretation of the references to colour in this figure legend, to the discretization size of the structured hypercube, which has been
the reader is referred to the web version of this article.) thoroughly sampled to minimize the error in the range of interest (read
Section 3.2 for more details).
A set of ga optimizations with different cell sizes 𝑎 have been In a similar manner to the previous study, the mass effect in the
performed in order to study its effect in the MLAM attenuation window MLAM response has been assessed via a set of optimizations limiting
area 𝐴𝛥TL and the double-peak frequency bandwidth 𝛥𝑓 . The results the maximum reference surface density 𝜇ref . In this case, Eq. (26) has
can be seen in Fig. 7 for 𝑎 = {50, 60, 70, 80} mm, each one including been rewritten as an equality constraint to impose the exact value of the
the optimized MLAM response obtained from the database interpolation target surface density. The results obtained for different 𝜇ref values are
8
G. Sal-Anglada et al. Engineering Structures 293 (2023) 116555
Fig. 7. STL results obtained for the MLAM structure with different cell size 𝑎 values. In each case, the optimized STL response (solid blue lines) is compared with corresponding
STL curves obtained from applying the 1D homogenization scheme (dashed blue lines) and 3D DNS (blue circles) with the exact set of optimized parameters. The resulting MLAM
attenuation windows (shaded blue areas), double-peak frequency bandwidths 𝛥𝑓 , as well as the surface areas and total thicknesses for each case are indicated. For reference, the
STL response of an homogeneous acoustic insulator (of 𝜇ref = 15 kg/m2 ) is also included in each figure (dash-dotted green lines). Additionally, the resulting optimized topologies
for the resonating layers’ unit cells are included considering the same relative scale for each cell size. (For interpretation of the references to colour in this figure legend, the
reader is referred to the web version of this article.)
displayed in Fig. 8. As in the previous study, the optimization approach Fig. 8-c shows that the optimized MLAM panels practically double this
tries to adjust the highest resonance frequency 𝑓𝛽 with the upper performance metric regardless of the target reference surface density,
frequency bound 𝑓U , while minimizing the first resonant frequency 𝑓𝛼 when compared to the corresponding case of a homogeneous panel of
in order to maximize both terms of Eq. (20). In this context, coupled, the same mass per unit area. Interestingly, it appears that the additional
double-peak solutions are achieved for all the optimizations regardless specific attenuation achieved by the optimized MLAM is higher for
the target surface density value. Nevertheless, the MLAM attenuation smaller reference surface densities, showing the optimization strategy
window area 𝐴𝛥TL (first term in Eq. (20)) monotonically increases with is even more efficient in these more challenging cases.
the target density, as expected (see Fig. 8-b). This tendency, however,
is not observed for the double-peak frequency bandwidth 𝛥𝑓 (second 5. Conclusions
term in Eq. (20)), which behaves concavely with respect to the surface
density and presents a maximum at 𝜇ref = 12.5 kg/m2 . Despite this, Multiresonant Layered Acoustic Metamaterials (MLAM) proposed in
if one looks at the effective attenuation bandwidth —defined as the this work can be considered practical solutions with great potential
frequency range in the interval on interest with a MLAM STL response to effectively attenuate sound at low frequency ranges, in particular
over the target 𝛥TL = 20 dB with respect to the reference homogeneous between 100 Hz and 500 Hz. The layer-based design (with two resonat-
insulator case (orange shaded areas in Fig. 8)— it increases also upon ing layers) exploits coupled-resonances phenomena among the different
increasing the target surface density, from 200 Hz (for 𝜇ref = 7.5 kg/m2 ) layers to significantly increase the attenuation in a wide frequency
to over 300 Hz (for 𝜇ref = 17.5 kg/m2 ). bandwidth, resulting in higher sound transmission losses (STL) than
To evaluate the overall panel attenuation capacity in the frequency those obtained via homogeneous acoustic insulators of the same, or
range of interest per unit of surface density, a performance metric, even higher, surface densities. Such additional attenuation can be
further increased in intensity and frequency range by means of size and
defined as the specific panel attenuation, is considered
topology optimization techniques.
1 𝑓
𝛥𝑓 ∗
∫𝑓 U TL(𝑓 )d𝑓 Although the design of MLAM panel structures could be directly op-
L
𝛱= (in dB m2 ∕kg) (31) timized with conventional techniques, the computational effort would
𝜇
9
G. Sal-Anglada et al. Engineering Structures 293 (2023) 116555
Fig. 8. Effect of the target surface density 𝜇ref on the transmission loss response for a MLAM of cell size 𝑎 = 60 mm. (a) STL response for the 𝜇ref = 12.5 kg/m2 case (solid blue line)
compared with corresponding STL responses of homogeneous acoustic insulators of different surface densities (dash-dotted green lines). The inset shows the different optimized
STL curves for the different target surface density values. (b) On the left axis (blue curve), effective additional attenuation level achieved by the optimized MLAM panels as a
function of the target surface density parameter. On the right axis (orange curve), effective bandwidth for attenuation over target 𝛥TL achieved by the optimized MLAM panel
(shaded region), and location of the double-peak frequencies. (c) Specific panel attenuation performance metric for the optimized MLAM compared to corresponding parameter for
equivalent surface density homogeneous panel. (For interpretation of the references to colour in this figure legend, the reader is referred to the web version of this article.)
be extremely high, requiring several days to complete a single opti- than 10 times reduction in weight to achieve these same levels of
mization. To overcome this computational cost, a specific layer-by-layer attenuation (in this frequency range) via conventional (mass-based)
homogenization technique has been proposed here, allowing to pre- acoustic insulators (notice the corresponding STL curves in Fig. 8-a).
cisely compute the STL in a matter of seconds. The effective properties These results reveal the potential for optimized MLAM-based panels,
that characterize each layer of the representative unit cell are obtained along with the proposed design strategy, to become a highly efficient,
through the proposed homogenization technique, along with a set of customizable and practical solution to sound insulation for several
attached masses and springs that represent the internal resonators. As a applications.
result, each layer can be represented by an equivalent 1D finite element
model with internal resonators, thus providing an extremely reduced CRediT authorship contribution statement
system of equations that can be solved practically instantaneously
for the desired frequency range. In this context, the results displayed G. Sal-Anglada: Visualization, Software, Validation, Writing – orig-
throughout this work reveal a robust correlation between DNS and inal draft. D. Yago: Software, Methodology. J. Cante: Funding acquisi-
homogenized solutions for MLAM configurations in the frequency range tion, Supervision. J. Oliver: Funding acquisition, Supervision. D. Roca:
of interest, up to 600 Hz. Software, Methodology, Conceptualization, Writing – review & editing,
Aiming to maximize the additional attenuation with respect to the Supervision.
homogeneous acoustic insulator while increasing at the same time the
Declaration of competing interest
frequency distance between the two resonant frequencies in a coupled
double-peak STL response, a two-step size optimization strategy has
The authors declare that they have no known competing finan-
been proposed in combination with the layer-by-layer homogenization
cial interests or personal relationships that could have appeared to
technique. To this end, the representative unit cell has been char-
influence the work reported in this paper.
acterized by a set of geometrical parameters, which have been then
optimized by a standard genetic algorithm. By using the proposed Data availability
homogenization framework to create a database of effective properties
of each layer (offline step), one can afford the high computational cost No data was used for the research described in the article.
of the ga optimization (online step), allowing to perform multiple size
optimizations with different parameters. Acknowledgements
In this regard, the proposed methodology provides a tool for de-
signing application-specific RVE cells based on the employed materials, The authors acknowledge the Spanish Ministry of Science and In-
desired frequency range, target additional attenuation, and geomet- novation, Spain (MCIN/AEI/10.13039/501100011033) for supporting
rical limitations (imposed by the manufacturing procedures or the G. Sal-Anglada with the PhD Grant PRE2019-088777 under the FPI
application settings), among others, with superior performance to the programme, and for funding this research through the ‘‘Severo Ochoa
one obtained via homogeneous acoustic insulators. As an example, Programme for Centres of Excellence in R&D’’ (Grant CEX2018-000797-
optimized MLAM panels have been obtained for which additional S-19-1) and the AC-METATECH project (TED2021-129413B-C21). This
attenuation of more than 20 dB have been achieved over a 300 Hz research has also been funded by the Ministry of Research and Uni-
range, between 200 Hz and 500 Hz, with a surface densities around versities of the Government of Catalonia, Spain, through the research
13 kg/m2 and thicknesses smaller than 30 mm. This represents a more grant 2021-PROD-00016 for the project METACOUSTECH.
10
G. Sal-Anglada et al. Engineering Structures 293 (2023) 116555
( T )T
(𝑗) ̂ ∗(𝑗,𝑖)
Appendix A. Derivation of the homogenized properties 𝛼̈ 𝜇∗(𝑗,𝑖) + 𝜆∗(𝑗,𝑖)
𝜇 𝛼𝜇∗(𝑗,𝑖) = − I(𝑗) M(𝑗)
𝜇 P𝜇 𝝓𝜇 𝑢̄̈ (𝑗) (A.11)
The FE-discretized version of Eq. (13) can be written in terms of where 𝛼𝜇∗(𝑗,𝑖) are the modal amplitudes corresponding to the mass-
normalized modes. To make the units consistent, we will define
𝐮̂ (𝑗) (𝑗) (𝑗)
𝜇 = I 𝑢̄ + X(𝑗) 𝜀(𝑗) + 𝐮̃̂ (𝑗) (A.1)
𝜇 ( T )
∗(𝑗,𝑖) 2
̂ denote the nodal values for the degrees of freedom (∙), and 𝑚̃ (𝑗,𝑖) = I(𝑗) M(𝑗) P(𝑗) 𝝓̂ (A.12)
where (∙) 𝜇 𝜇 𝜇
⎧⋮ ⎫ ⎧ ⋮ ⎫ 𝑘̃ (𝑗,𝑖) = 𝜆∗(𝑗,𝑖)
𝜇 𝑚̃ (𝑗,𝑖) (A.13)
⎪1⎪ ⎪𝑥̂ ′(𝑗,𝑎) − 𝑥′(𝑗) ⎪
⎪ ⎪ ⎪ 0 ⎪ 𝛼𝜇∗(𝑗,𝑖)
I(𝑗) = ⎨0⎬ , X(𝑗) =⎨ 0 ⎬ (A.2) 𝑞 (𝑗,𝑖) = 𝑢̄ (𝑗) + √ (A.14)
⎪0⎪ ⎪ 0 ⎪ 𝑚̃ (𝑗,𝑖)
⎪ ⎪ ⎪ ⎪
⎩⋮ ⎭ ⎩ ⋮ ⎭ Note that each 𝑞 (𝑗,𝑖) refers to the total displacement amplitude for
the 𝑖th internal resonator. These definitions allow us to obtain the
with 𝑥̂ ′(𝑗,𝑎) being the 𝑥′ -coordinate of the 𝑎th node. Fixed degrees
following expressions from Eqs. (A.10) and (A.11), after some algebraic
of freedom and periodic boundary conditions are applied on the mi-
manipulation:
crofluctuation field through a matrix P(𝑗)
𝜇 , such that
(𝑗)
∑
𝑁 √
⎧ (𝑗,r) ⎫ 𝐮̂ ∗(𝑗) =
∗(𝑗,𝑖)
𝝓̂ 𝜇 𝑚̃ (𝑗,𝑖) (𝑞 (𝑗,𝑖) − 𝑢̄ (𝑗) ) (A.15)
⎪ 𝐮̃̂ 𝜇 ⎪ ⎡𝟎 𝟎⎤ 𝜇
⎪ 𝐮̃̂ (𝑗,f) ⎪ ⎢𝟎 𝟎⎥ { ̂ (𝑗,−) }
𝑖=1
⎪ 𝜇 ⎪ ⎢ ⎥ 𝐮̃ 𝜇 𝑚̃ (𝑗,𝑖) 𝑞̈(𝑗,𝑖) + 𝑘̃ (𝑗,𝑖) (𝑞 (𝑗,𝑖) − 𝑢̄ (𝑗) ) = 0
𝐮̃̂ (𝑗) = ⎨𝐮̃̂ 𝜇(𝑗,+) ⎬ = ⎢𝟏 𝟎⎥ = P(𝑗)
𝜇 𝐮𝜇
̂ ∗(𝑗) (A.3) (A.16)
𝜇
⎪𝐮̃̂ (𝑗,−) ⎪ ⎢𝟏 𝐮̃̂ (𝑗,i) √
𝟎⎥ 𝜇
⎪ 𝜇(𝑗,i) ⎪ ⎢ ⎥ Note that Eq. (A.11) has been multiplied by 𝑚̃ (𝑗,𝑖) to obtain Eq. (A.16).
⎪ 𝐮̃̂ 𝜇 ⎪ ⎣𝟎 𝟏⎦
Using Eq. (A.10), it is possible to obtain the expression for the effective
⎩ ⎭
inertial force density in terms of 𝑞 (𝑗,𝑖) as
where 𝐮̃̂ 𝜇(𝑗,r) , 𝐮̃̂ (𝑗,f)
𝜇 , 𝐮 ̃̂ (𝑗,+)
𝜇 and 𝐮̃̂ (𝑗,−)
𝜇 are the nodal displacements referred
(𝑗)
to the boundaries 𝛤𝜇 , 𝛤𝜇(𝑗,f) , 𝛤𝜇(𝑗,+) and 𝛤𝜇(𝑗,-) , respectively, 𝐮̃̂ (𝑗,i)
(𝑗,r)
are the 1 ∑
𝑁
𝜇
𝑝̇ (𝑗) = 𝜌(𝑗) 𝑢̄̈ (𝑗) + 𝑚̃ (𝑗,𝑖) 𝑞̈(𝑗,𝑖) (A.17)
remaining nodal displacements, and 𝟏 denotes the identity matrix. The eff
|𝛺𝜇(𝑗) | 𝑖=1
unit cell’s system of equations for the 𝑗th layer is given by
⎛ ∑
𝑁 (𝑗)
⎞
1 ⎜I(𝑗) T M(𝑗) I(𝑗) − (𝑗,𝑖) ⎟
M(𝑗) ̂̈ (𝑗) (𝑗) ̂ (𝑗)
𝜇 𝐮𝜇 + K𝜇 𝐮𝜇 = 𝟎 (A.4) 𝜌(𝑗) = 𝑚
̃ (A.18)
|𝛺𝜇(𝑗) | ⎜⎝ ⎟
eff 𝜇
𝑖=1 ⎠
where M𝜇(𝑗) and K𝜇(𝑗) are the mass and stiffness matrices for the cor-
Again, since the only relevant component of the macroscopic dis-
responding layers, respectively. Following the procedure explained in
placement imposed is in the 𝑥-direction, only the 𝑥-component of the
Roca et al. [29], the resolution proceeds by splitting the RVE system
macroscopic inertial force density is obtained.
into a quasi-static and an inertial components.
(a) Quasi-static system:
For the quasi-static subsystem, 𝑢̄ (𝑗) = 0 and the resulting effective Appendix B. STL evaluation for the homogenized MLAM
inertial force density is neglected, i.e., 𝐩̇ ≈ 𝟎. In this context, from sub-
stituting Eq. (A.1) into Eq. (A.4), the solution for the microfluctuation The 1D version of Eq. (3) applied to the 𝑗th layer domain reads
field can be obtained in terms of the macroscopic strain as 𝜕𝜎 (𝑗)
= 𝑝̇ (𝑗) (B1)
T 𝜕𝑥
𝐮̂ 𝜇∗(𝑗) = −(K∗(𝑗) −1 (𝑗) (𝑗) (𝑗) (𝑗)
𝜇 ) P𝜇 K𝜇 X 𝜀 (A.5)
Now, introducing Eqs. (A.6) and (A.17) into Eq. (B1), and coupling it
T with Eqs. (A.16), the following system is obtained
where K∗(𝑗)
𝜇 = P(𝑗) (𝑗) (𝑗)
𝜇 K𝜇 P𝜇 . This allows us to obtain the effective
( )
macroscopic stress as ⎧ (𝑗) (𝑗) 1 ∑𝑁 (𝑗) ̃ (𝑗,𝑖) (𝑗) 𝜕 (𝑗) 𝜕 𝑢̄
(𝑗)
⎪𝜌eff 𝑢̄̈ + 𝑖=1 𝑘 (𝑢̄ − 𝑞 (𝑗,𝑖) ) − 𝐸eff =0
(𝑗) (𝑗)
𝜎 (𝑗) = 𝐸eff 𝜀 (A.6) ⎨ |𝛺𝜇(𝑗) | 𝜕𝑥 𝜕𝑥 (B2)
⎪𝑚̃ (𝑗,𝑖) 𝑞̈(𝑗,𝑖) + 𝑘̃ (𝑗,𝑖) (𝑞 (𝑗,𝑖) − 𝑢̄ (𝑗) ) = 0
(𝑗)
=
1 T
X(𝑗) (𝟏 − K𝜇(𝑗) P(𝑗) ∗(𝑗) −1 (𝑗) T (𝑗) (𝑗)
(A.7) ⎩
𝐸eff
|𝛺𝜇 | 𝜇 (K𝜇 ) P𝜇 )K𝜇 X
Upon a standard 1D FEM integration of Eqs. (B2), considering just one
Notice that because we only imposed a macroscopic strain in the 𝑥- element for the whole layer, the resulting system reads
direction, only the 𝑥-component of the macroscopic stress is obtained [ ]{ } [ ]{ } { }
as a result. M(𝑗)
𝑢𝑢 𝟎 𝐮̂̈ (𝑗) K(𝑗)
𝑢𝑢 K(𝑗)
𝑢𝑞 𝐮̂ (𝑗) 𝐟̂ (𝑗)
̈ (𝑗) + K(𝑗) K(𝑗) = (B3)
(b) Inertial system: 𝟎 M(𝑗)
𝑞𝑞 𝐪 ̂ 𝑞𝑢 𝑞𝑞 𝐪̂ (𝑗) 𝟎
For the inertial subsystem, 𝜀(𝑗) = 0 and it is assumed that the with
resulting effective stress can be neglected, i.e. 𝝈 ≈ 𝟎. Therefore, after { }
𝑢𝑗
substituting Eq. (A.1), Eq. (A.4) yields 𝐮̂ (𝑗) = (B4)
𝑢𝑗+1
M𝜇∗(𝑗) 𝐮̂̈ ∗(𝑗) + K∗(𝑗) ̂ ∗(𝑗) = −P(𝑗)
T (𝑗) (𝑗) (𝑗)
̈ (A.8) { }
𝜇 𝜇 𝐮𝜇 𝜇 M𝜇 I 𝑢̄ 𝑓𝑗
𝐟̂ (𝑗) = (B5)
T (𝑗) (𝑗) ∗(𝑗,𝑖) 𝑓𝑗+1
where M∗(𝑗)
𝜇 = P(𝑗)
Denoting 𝜆∗(𝑗,𝑖)
𝜇 M𝜇 P 𝜇 .
𝜇 and the eigenval- 𝝓̂ 𝜇
ues and mass-normalized eigenvectors, respectively, of the following ⎧ ⋮ ⎫
⎪ ⎪
equation ̂ (𝑗)
𝐪 = ⎨𝑞 (𝑗,𝑖) ⎬ (B6)
⎪ ⋮ ⎪
(K𝜇∗(𝑗) − 𝜆∗(𝑗,𝑖)
∗(𝑗,𝑖)
M𝜇∗(𝑗) )𝝓̂ 𝜇 =𝟎 (A.9) ⎩ ⎭
𝜇
(𝑗) (𝑗) [ ] 𝑁 (𝑗) (𝑗,𝑖) [ ]
∗(𝑗,𝑖) 𝐸eff 𝐴 1 −1 ∑ 𝑘̃ 1−𝛽 𝛽
projecting the system given by Eq. (A.8) into the eigenvectors 𝝓̂ 𝜇 K(𝑗)
𝑢𝑢 = + (B7)
ℎ(𝑗) −1 1 2 𝛽 1−𝛽
yields 𝑖=1
(𝑗) ⎡⋱ 0⎤
∑
𝑁
∗(𝑗,𝑖) K(𝑗) ⎢ 𝑘̃ (𝑗,𝑖) ⎥ (B8)
𝐮̂ ∗(𝑗) = 𝝓̂ 𝜇 𝛼𝜇∗(𝑗,𝑖) (A.10) 𝑞𝑞 = ⎢ ⎥
𝜇
𝑖=1 ⎣0 ⋱⎦
11
G. Sal-Anglada et al. Engineering Structures 293 (2023) 116555
12