Robust-to-Uncertainties Optimal Design of

Seismic Metamaterials
Paul-Remo Wagner 1; Vasilis K. Dertimanis, Ph.D. 2;
Eleni N. Chatzi, Ph.D., A.M.ASCE 3; and James L. Beck, Ph.D., M.ASCE 4
Downloaded from ascelibrary.org by MICHIGAN TECHNOLOGICAL UNIVERSITY on 09/07/19. Copyright ASCE. For personal use only; all rights reserved.

Abstract: Metamaterials, which draw their origin from a special class of structured (periodic) materials characterized by a dynamic filtering
effect, have recently emerged as a prospective means for structural seismic protection. This paper explores such a periodic arrangement in the
form of local adaptive resonators buried in the soil, serving as a seismic protection barrier. As a starting point, a simplistic representation is
chosen herein that comprises chains of mass-in-mass unit cells. A robust-to-uncertainties optimization of such a chain, addressing uncer-
tainties at the level of the excitation, the system properties and the model structure itself, is conducted. The optimization problem is for-
mulated within the context of reliability assessment, where the objective function is the failure probability of the structure to be protected
against seismic input. The problem is solved through adoption of the subset optimization algorithm enhanced through the simultaneous
implementation of a stochastic approximation algorithm. It is demonstrated that not all parameters of the chain model require optimization,
because the failure probability proves to be a monotonic function of a subset of the parameters. A primary objective herein lies in optimizing
the internal unit-cell stiffness properties. It is further demonstrated that the effectiveness of the protection offered by the metamaterial is
improved for spatially varying unit-cell properties. The optimization procedure is carried out in the frequency domain, with an example
application confirming that a time domain optimization is expected to yield similar optimal configurations. A parametric study using a
nonlinear model is also presented, offering a starting point for more refined future investigations. DOI: 10.1061/(ASCE)EM.1943-
7889.0001404. © 2017 American Society of Civil Engineers.
Author keywords: Seismic isolation; Periodic lattice; Metamaterials; Optimal design; Reliability.

Introduction frequencies of the building. This decrease in overall horizontal

stiffness results in an increase of the fundamental period and,
Seismic events, while relatively rare, may bear devastating conse- when designed properly, places the building in a more favorable
quences on a human, fiscal, and societal level. From a civil engineer- position with respect to the predominant ground motion frequen-
ing standpoint, seismic hazard mitigation poses a great challenge, cies (Harvey and Kelly 2016). The effectiveness of the isolation
as traditional deterministic design methods may not adequately depends on numerous factors, but is especially difficult with light-
anticipate or account for such extreme events. In ensuring seismic weight structures, soft soil sites, and with respect to near fault
resilience, i.e., the ability to recover against earthquakes, different effects.
approaches have been conceived to reduce the impact of seismic For TMD, a mass is attached to the primary structure through a
events on structures, such as base isolation (BI) (Skinner et al. 1975; spring. By properly tuning this attached oscillator, kinetic energy of
Harvey and Kelly 2016) and tuned mass damping (TMD) (Kaynia the main structure can be partially transferred to the TMD. This
et al. 1981; Chey et al. 2010). reduces the movement in the primary structure, albeit only over
The basic idea behind BI is the horizontal decoupling of the a relatively narrow frequency range, typically near the fundamental
ground from a protected structure. These BI systems are typically frequency of the structure. Therefore TMD is most effective, if the
implemented by introducing low horizontal stiffness bearings be- excitation characteristics lie within a certain frequency range. If this
tween the ground and the structure, thereby modifying the natural is not guaranteed, the TMD may in fact result in an increase of the
1
primary structural response. In structural engineering applications,
Ph.D. Student, Dept. of Civil, Environmental and Geomatic Engineer- this leads to a major concern because the dynamic excitation (seis-
ing, Institution of Structural Engineering, ETH Zürich, Stefano-Franscini-
mic or wind) is highly arbitrary and may not be guaranteed, with
Platz 5, 8093 Zürich, Switzerland. E-mail: wagner@ibk.baug.ethz.ch
2
Researcher, Dept. of Civil, Environmental and Geomatic Engineering,
sufficient certainty, to lie in a prescribed range (Kaynia et al. 1981).
Institution of Structural Engineering, ETH Zürich, Stefano-Franscini-Platz Another concern is related to the choice of the target frequency of
5, 8093 Zürich, Switzerland. E-mail: v.derti@ibk.baug.ethz.ch the TMD because of the uncertainty in the primary structure’s natu-
3 ral frequencies. This can lead to the so-called mistuning effect,
Assistant Professor, Dept. of Civil, Environmental and Geomatic
Engineering, Institution of Structural Engineering, ETH Zürich, Stefano- which ultimately might render a TMD ineffective (Chey et al. 2010).
Franscini-Platz 5, 8093 Zürich, Switzerland (corresponding author). A novel approach to address shortcomings of existing protection
E-mail: chatzi@ibk.baug.ethz.ch systems seeks to mitigate seismic vibrations through the use of
4
Professor, Division of Engineering and Applied Science, California In- metamaterials (Finocchio et al. 2014). These were originally de-
stitute of Technology, Pasadena, CA 91125. E-mail: jimbeck@caltech.edu
rived from research on phononic crystals—physical particles rep-
Note. This manuscript was submitted on April 4, 2017; approved on
August 2, 2017; published online on December 20, 2017. Discussion resenting mechanical vibrations responsible for the transmission of
period open until May 20, 2018; separate discussions must be submitted everyday sound and heat (Maldovan 2013). It has been found that
for individual papers. This paper is part of the Journal of Engineering when arranged in a periodic configuration as in crystals, waves of
Mechanics, © ASCE, ISSN 0733-9399. certain frequencies cannot propagate through the crystal medium.

© ASCE 04017181-1 J. Eng. Mech.

J. Eng. Mech., 2018, 144(3): 04017181

 This frequency range has been called a phononic band gap, or stop similarly to TMD by reducing the structural response in certain
band, in the literature. narrow frequency ranges. Contrary to TMD, in SMs the seismic
Researchers have proposed different types of seismic metama- energy is hindered from reaching the protected structure and so
terials (SMs) to produce band gaps in the low frequency regions does not have to be initially transferred through the structure to
that are relevant to earthquake protection schemes. For a compre- the energy dissipating TMD.
hensive literature review, the reader is referred to Dertimanis et al. This exploratory study contributes to ongoing SM research
(2016). Among others, Cheng et al. (2013) propose mass-in-mass through establishing a needed uncertainty management approach,
concrete unit cells interconnected through silicon rubber. This ap- which is illustrated by deriving the optimal configuration of a
proach produced band gaps in the range between 15 and 49 Hz. By six-unit-cell chain employed for protection of a single degree of
placing pillars in the soil around a structure, Huang and Shi (2013)
Downloaded from ascelibrary.org by MICHIGAN TECHNOLOGICAL UNIVERSITY on 09/07/19. Copyright ASCE. For personal use only; all rights reserved.

freedom oscillator. Frequency domain reliability optimization is

achieved a gap between 32 and 47 Hz. In Achaoui et al. (2016), adopted, with remarks regarding the time domain behavior. It is
band gaps from 8–49 Hz are achieved by placing large spheres in also shown that a choice of spatially varying unit-cell properties,
the soil around and under a structure. under certain conditions, achieves more efficient protection than
Focusing on the attenuation of surface waves, Palermo et al. using identical unit-cell chains. A parametric study is also carried
(2016) propose a seismic metabarrier, consisting of subwavelength out using the optimized structure with nonlinear springs.
resonant structures that are buried under the soil surface, for con- The paper begins with a comprehensive design problem formu-
verting Rayleigh waves into shear bulk waves that propagate away lation, introducing the mass-in-mass SM model with available
from the soil surface. Using a similar principle, Colombi et al. analysis methods and a seismic excitation model. Subsequently,
(2016a) describe a resonant metawedge for effective wave mode the optimal stochastic design approach for reliability is introduced
conversion, providing both analytical and numerical models. using these models. In the next section, the models are investigated
Miniaci et al. (2016) proceed one step further and develop optimal specifically for the case of SM optimization, including an initial
design criteria, by investigating a number of SM configurations. At sensitivity analysis for the most important parameters. The authors
this point, it is worth noting that Colombi et al. (2016b) have also then continue to present an efficient two-stage optimization algo-
investigated the behavior of natural periodic structures, such as for-
rithm for solving the posed design optimization problem. Finally, a
ests, in terms of surface Rayleigh wave attenuation.
parametric study is presented, illustrating the results obtained by
Recently, Dertimanis et al. (2016) demonstrated that lower fre-
the described robust-to-uncertainties reliability design framework.
quency bands, within the active range of earthquake motions and
the fundamental period of structural systems, are indeed feasible.
Using one-dimensional SMs with a small number of unit cells, they
proposed a technologically realizable unit-cell design prototype, Problem Formulation
capable of achieving attenuation already from as low as 0.5 Hz.
An extension of this concept to two-dimensional (e.g., plane) SMs Physical Layout
is presented in Wagner et al. (2016a, b).
Consider the structure of Fig. 1(a), which is excited by a horizontal
In view of the ongoing research on SMs, one aspect that has
ground acceleration component [ẍg ðtÞ] attributable to a near-fault
received considerable less attention thus far pertains to the system-
seismic event. The induced near-fault kinematic effect [e.g., the
atic exploration of the induced design uncertainties. Indeed, SMs
are hard to design through conventional methods, primarily be- ground motion xg ðtÞ] may be considered as the combined result
cause seismic events are inherently uncertain and are thus hard to of the propagation of vertical S waves and surface waves (Love
quantify deterministically. Additionally, the uncertainties related to and/or Rayleigh).
SMs themselves are large, because little experimental data or ex- To attenuate the ground acceleration, a number of, not neces-
perience are available. As some sort of model of a physical system sarily identical, resonators are embedded in the soil, as shown
is adopted, it is also necessary to take into account modeling errors in Fig. 1(b). Each cell comprises an external (mext ) and an internal
that are associated with any model (Beck and Katafygiotis 1998). mass (mint ), which are interconnected by a linear spring of stiffness,
Building on some earlier work (Au and Beck 2003; Taflanidis and kint . A discussion on the design of such resonators falls outside the
Beck 2008b), this paper attempts to address underlying uncertain- scope of this work, and the reader is referred to Dertimanis et al.
ties through a comprehensive robust-to-uncertainties reliability ap- (2016), to Palermo et al. (2016), and to Miniaci et al. (2016), for a
proach. To do this, the authors resort to a two-stage optimization thorough treatment of the engineering aspects of this approach.
framework that was originally conceived in Taflanidis and Beck Herein it is just emphasized that these studies have already con-
(2008a). The authors further define a SM model that consists of firmed the design feasibility, in terms of geometrical, dimensional,
a chain of mass-in-mass unit cells. To model a protected structure, and engineering constraints.
a linear oscillator is attached to this chain. It is then attempted to The unit cells are embedded into a soil of specific shear modulus
optimize the internal unit-cell properties (i.e., internal stiffness), to and of purely elastic behavior, which comprises an integral part of
obtain a minimum failure probability with respect to engineering the SM through the resonator–soil–resonator interaction mecha-
quantities like relative displacement and absolute acceleration of nism. The soil is herein modeled by means of a linear spring of
the oscillator, whose response is to be mitigated. stiffness, kext (Makris and Gazetas 1992). For the purposes of this
In this paper, SMs that are composed of a series of soil em- study, and for ensuring a proper behavior of the SM, the following
bedded resonators are studied. These resonator chains result in re- assumptions are adopted:
duction of the vibrational response amplitude in certain frequency 1. The soil maintains its contact (and thus the same elastic prop-
ranges (Dertimanis et al. 2016), referred to as response reduction erties) along the height of the resonator.
ranges (RRR) in this paper. These SMs can be understood as a com- 2. The resonator floats only along the horizontal direction, im-
bination of BI and TMD, while additionally exploiting structural plying specific boundary conditions on the top and the bottom of
periodicity. The external stiffness, kext , to some degree, decouples the internal and the external masses.
the structure from the ground, and thereby bears a similar effect 3. The SM forms a box–type structure that is excited on
to BI systems. The internal resonators, on the other hand, work its far left side by the ground acceleration and causes a vibration

© ASCE 04017181-2 J. Eng. Mech.

J. Eng. Mech., 2018, 144(3): 04017181

 x g (t )
Downloaded from ascelibrary.org by MICHIGAN TECHNOLOGICAL UNIVERSITY on 09/07/19. Copyright ASCE. For personal use only; all rights reserved.

(a)

(a)

x g (t )

L
(b)

Fig. 2. SM models for soil embedded resonators: (a) system model G1 :

seismic metamaterial; (b) system model G2 : seismic metamaterial with
(b) attached oscillator

Fig. 1. Graphical interpretation of the physical problem: (a) unpro-

tected structure under near fault ground excitation; (b) SM-based
model already offers valuable insight into the metadamping phe-
protected structure
nomenon, and it is computationally cheaper to evaluate, whereas it
can be quite easily extended to circular or dome-like barriers in two
or three dimensions (Wagner et al. 2016a).
acceleration response on the external mass of the far right resonator, The vibration attenuation properties of the SM have been ex-
on which the structure to be protected is adjusted. plored using Bloch’s theory, under the assumption of an infinite
4. The SM is amenable to uncertainties, both in terms of number of unit cells (Dertimanis et al. 2016). In the case of SM
excitation and its structural parameters. composed of identical units, it can be shown that the frequency
On the basis of these assumptions, the purpose herein is to prob- band gap is determined by Finocchio et al. (2014)
abilistically design SMs that perform optimally with respect to sffiffiffiffiffiffiffiffi
some reliability objective. As it will be later shown, the design rffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
m 1 kint
parameter is the internal stiffness, kint , which may be different in f0;int ≤ f ≤ f0;int 1 þ int ; f 0;int ¼ ð1Þ
mext 2π mint
a subset of the SM’s unit cells. To do so, the adopted modeling
approach consists of three interrelated models, the SM model, the
excitation model, and the performance evaluation model, which are and is independent of the external stiffness, kext . In a finite SM,
described in the following. however, Bloch’s theory is not applicable, and quantities from con-
ventional vibration analysis, such as the frequency response func-
tion (FRF), are utilized instead. Fig. 3 illustrates the difference
Seismic Metamaterial Model
between the dispersion relation of a SM with an infinite number
To obtain a mathematical description of the physical layout, the SM- of unit cells and the FRF amplitude of a SM with a finite number
protected structure (e.g., a multistory building), is modeled as a lin- of unit cells. Note that the response in the frequency domain band
ear oscillator corresponding to the fundamental structural vibration gap is not eliminated, but significantly reduced. The response out-
mode, and is attached to the unit cell furthest away from the exci- side the gap is however increased; this is discussed in Dertimanis
tation source, as displayed in Fig. 2(b). The protected structure is et al. (2016). The frequency range for which the amplitudes are
assumed to have a frequency of 1 Hz, as this is near the fundamental reduced (e.g., it retains a value <0 dB) will be henceforth referred
frequency of typical civil engineering structures (Hatzigeorgiou and to as the response reduction range (RRR).
KanapitsasP 2013). The oscillator mass is chosen equal to the total Focusing further on the finite SM, the number of unit cells, N,
SM mass ( mint;i þ mext;i ¼ mSDOF ), whereas its stiffness kSDOF the external stiffness, kext , and the mass ratio between the external
is determined so that the resulting frequency f SDOF coincides with and the internal mass, r ¼ mext =mint , play an important role for the
the protected structure’s fundamental one. It is emphasized that SMs performance of the induced RRR. It must be emphasized that re-
are not restricted to this one-dimensional setup, but this simple sponse reduction occurs only because kext ≠ ∞ (i.e., the soil is not

© ASCE 04017181-3 J. Eng. Mech.

J. Eng. Mech., 2018, 144(3): 04017181

 sffiffiffiffi rffiffiffiffiffiffiffi
1 k̂ kext X kext
f̂ ¼ ¼ ; where m̂ ¼ ðmint;i þ mext;i Þ k̂ ¼
2π m̂ N2 N
ð2Þ

where k̂ = equivalent chain stiffness for N identical springs con-

nected in series, and m̂ = total mass of the chain. This reveals that
the number of unit cells and the external stiffness are interrelated
parameters. The effects of the mass ratio are illustrated in Fig. 5 for
fixed N. It is shown that the ratio, r, primarily affects the upper limit
Downloaded from ascelibrary.org by MICHIGAN TECHNOLOGICAL UNIVERSITY on 09/07/19. Copyright ASCE. For personal use only; all rights reserved.

(a) (b)
of the associated frequency band gap. For lower mass ratios, the
Fig. 3. Comparison of the band gaps in the infinite chain (dispersion band gaps extend into higher frequencies and are wider overall.
relation) and the RRR in the finite chain [FRF HðfÞ] of model G1 : Implementing low mass ratios forms a major engineering chal-
(a) dispersion relation; (b) complex FRF HðfÞ lenge, as the large internal mass ought to be supported by a much
smaller external one.
In addition to the aforementioned parameters and in an effort to
produce wider band gaps in lower frequency ranges, which is criti-
rigid which is always true in practice), if this were the case the cal for earthquake engineering applications, the use of the so called
excitation movement would simply bypass the local resonators rainbow-traps is proposed (Zhu et al. 2013; Krödel et al. 2015),
and transfer the motion directly from the first element of the chain according to which nonidentical mass-in-mass unit cells with dis-
to the last. Similarly, the external stiffness has to be non-null, tributed natural frequencies are utilized. In this setting, the aim is
kext ≠ 0, because null external stiffness would hinder any energy to create a chain of unit cells with individually overlapping band
transport along the SM and again render the latter ineffective. gaps and to thereby increase the overall RRR. Using this approach,
The effect of N and kext on the FRF amplitude inside the RRR a response reduction in a range from 4–7 Hz has been reported
is depicted in Fig. 4. (Krödel et al. 2015), which is close (but not sufficient) to the target
Define now an equivalent natural frequency, ω̂, for a lattice of frequency ranges for civil engineering structures. The use of rain-
mass-in-mass unit cells as bow traps is also followed in the current study, however, instead of

(a) (b) (c) (d)

(e)

Fig. 4. FRF of the absolute acceleration amplitudes ẍ in dB for different external stiffness kext and unit cell numbers N with identical unit cells in
model G1 ; the excitation frequency is given by f ex and the internal frequency by f int ; the output is truncated from (−100 to 0) dB to emphasize the
locations of the band gaps: (a) kext ¼ 10 s−2 , N ¼ 1, ω̂ ¼ 3.16 rad s−1 ; (b) kext ¼ 1,000 s−2 , N ¼ 10, ω̂ ¼ 3.16 rad s−1 ; (c) kext ¼ 100 s−2 , N ¼ 10,
ω̂ ¼ 1 rad s−1 ; (d) kext ¼ 10,000 s−2 , N ¼ 100, ω̂ ¼ 1 rad s−1 ; (e) grayscale for the FRF

(a) (b) (c)

(d)

Fig. 5. FRF of the absolute acceleration amplitudes ẍ in dB for different mass ratios r ¼ m mint and unit cell numbers N with identical unit cells in model
ext

G1 ; the excitation frequency is given by f ex and the internal frequency by f int ; the output is truncated from (−100 to 0) dB to emphasize the locations
of the band gaps; mint ¼ 0.15, mext ¼ 0.015: (a) r ¼ 0.01, N ¼ 10; (b) r ¼ 0.1, N ¼ 10; (c) r ¼ 1, N ¼ 10; (d) grayscale for the FRF

© ASCE 04017181-4 J. Eng. Mech.

J. Eng. Mech., 2018, 144(3): 04017181

 the logarithmic spacing of the internal frequencies (Krödel et al. (Taflanidis and Beck 2008a; Taflanidis et al. 2008). The consider-
2015), the authors herein optimize the internal oscillator natural ation of low frequency near fault ground motions is of particular
frequency properties to achieve optimal SM protection perfor- interest regarding the seismic hazard for civil engineering struc-
mance. It must be noted that, whereas the focus lies on seismic tures, as the predominant frequency of these ground motions typ-
excitation, the underlying methodology can be easily extended ically coincides with the fundamental frequency of the exposed
to other engineering vibration attenuation problems as well. structure.
The adopted SM model incorporates parametric and excitation As illustrated in Fig. 6, the adopted stochastic ground motion
uncertainties, and design variables. The former are arranged in the model E combines high and low frequency counterparts. After cre-
parameter vector, θs , defined as ating individual time histories for these two components [Figs. 6(a
and b)], the combined acceleration time history is calculated. To
Downloaded from ascelibrary.org by MICHIGAN TECHNOLOGICAL UNIVERSITY on 09/07/19. Copyright ASCE. For personal use only; all rights reserved.

θs ¼ ½kext ; mint ; mext ; f SDOF ; ζ ð3Þ this end, the amplitudes of the discrete Fourier spectra of the simu-
lated time histories are subtracted [Figs. 6(c and d)]. A synthetic
for which a probability density function (PDF) pðθs Þ is assigned.
time history is then constructed, with its Fourier amplitude spec-
In Eq. (3), f SDOF refers to the natural frequency of the attached
trum derived from the previous step and its phase determined as
oscillator, whereas ζ corresponds to a damping ratio, herein intro-
the phase of the high frequency component’s discrete Fourier trans-
duced in the form of viscous Rayleigh damping through a Caughey
form (DFT) [Fig. 6(e)]. Finally, the near fault ground motion time
matrix (Chopra 2007), where applicable. The uncertain excitation ϵ
history is obtained as the superposition of this time history with the
is further elaborated on in the next subsection. The design variables,
time history of the low frequency component [Fig. 6(f)].
φi , correspond to the model parameters that optimization is in-
It has to be noted that the subtraction in the above procedure
tended for. Following Eq. (1), the internal-cell properties are the
[Fig. 6(d)] cannot be guaranteed to generally yield positive values.
defining parameters for the location of the band gaps, and are there-
This fact is well known by the authors of the original paper
fore chosen as design variables. Equivalent to the parameter vector,
(Mavroeidis and Papageorgiou 2003) but was observed to not have
the authors define a design vector, φ, containing all φi . The result-
a significant impact on the resulting time histories.
ing stochastic system model has the form
The stochastic excitation model is parameterized by the uncer-
ξ ¼ Gðϵ; θs ; φÞ ð4Þ tain vector, θϵ , defined as

where the uncertain output ξ may be any response quantity of the θϵ ¼ ½Mw ; R; Zw ; γ p ; ν p ; εf ; εv  ð5Þ
SM model, either in its infinite (Bloch theory), or in its finite
in which the seismic magnitude M w , the epicentral distance R and
(classical vibration theory) format.
the white noise sequence Zw correspond to the high frequency com-
ponent [details in Boore (2003)], whereas the remaining parameters
Earthquake Excitation Model γ p , ν p , εf , and εv correspond to the low frequency component
[details in Mavroeidis and Papageorgiou (2003)]. The assigned
To establish a robust-to-uncertainties system design procedure for PDF pðθϵ Þ of θϵ is discussed in the following section. The uncertain
SMs, the probabilistic seismic ground motion model of Taflanidis excitation, ϵ, to be used in the SM is thus the output of the stochas-
(2007) is implemented. The method extends the models initially tic excitation model, that is
developed by Boore (2003) and Mavroeidis and Papageorgiou
(2003) for synthetic ground motion generation, and provides a ϵ ¼ Eðθϵ Þ ð6Þ
stochastic model for near-fault seismic excitations. It is well suited
for stochastic system design, because of the well-defined physical It is noted that the source spectra currently used for the sto-
meaning of its parameters and its relatively low computational cost chastic ground motion model is developed for earthquakes in

(a) (b) (c)

(d) (e) (f)

Fig. 6. Near fault ground motion model used to create acceleration time histories; the shown plots are from a simulation with seismic magnitude
M w ¼ 7 and a fault distance R ¼ 10 km; the sampling rate is Δt ¼ 0.01 s: (a) high frequency component ðM w ; R; Zw Þ; (b) low frequency component
ðγ p ; ν p ; εf ; εv Þ; (c) DFT of time histories; (d) subtraction of spectra; (e) synthetic time history; (f) superposition of Figs. 6(b and e) [ϵðtÞ]

© ASCE 04017181-5 J. Eng. Mech.

J. Eng. Mech., 2018, 144(3): 04017181

 California, but can in general be replaced by any applicable source 2008a). Both techniques require the repeated evaluation of the
spectrum model (Taflanidis 2007; Boore 2003). model class M.

Optimal Stochastic Design Model Evaluation

The problem treated in this study pertains to the optimal design Given a certain design choice, φ, and a single realization of the
of the SM under the presence of uncertainties. In this setting, a uncertain parameter vector, θ, the evaluation of the model class
stochastic model class M is defined, which combines the stochastic M pertains to the determination of the value of the failure indicator
system model, Gðϵ; θs ; φÞ, the stochastic excitation model, Eðθϵ Þ, function, I F . The underlying process is illustrated for one design
and a performance evaluation model, jðξ; θw Þ. The class M incor- choice, φ, and one realization of θ ∼ pðθjφÞ in Fig. 7 for a fre-
Downloaded from ascelibrary.org by MICHIGAN TECHNOLOGICAL UNIVERSITY on 09/07/19. Copyright ASCE. For personal use only; all rights reserved.

porates all the individual model parameterizations into the compact quency domain evaluation. The stochastic excitation model produ-
parameter vector, θ ces a time–series ϵðtÞ ¼ Eðθϵ Þ, which is then transformed to the
frequency domain, ϵðfÞ. The stochastic system model is accord-
θ ¼ ½θs ; θϵ ; θw  ð7Þ ingly evaluated to produce a FRF as ξðfÞ ¼ G½ϵðfÞ; θs ; φ that
is driven to the performance evaluation model j½ξðfÞ; θw  where
with corresponding PDF, pðθjφÞ. Both the uncertain parameter vec-
tor, θ, and the design vector, φ, are bounded by admissible spaces, θw ¼ η ð10Þ
Θ and Φ, respectively, so that θ ∈ Θ ⊂ Rnθ and φ ∈ Φ ⊂ Rnφ .
In defining the objective function for the SM design problem with η introduced below.
considered herein, attention is focused on the class of reliability The evaluation model consists of an augmented limit state func-
objective problems (ROPs) where a failure probability PðFjφÞ is ~
tion gðφ; θÞ, which checks the system model output, ξðfÞ, to evalu-
used as the objective function [e.g., Eq. (9)]. The goal is to find a ate whether the performance is acceptable or not (Taflanidis and
design vector, φ, that results in the lowest failure probability while Beck 2008a)
being bounded by the admissible design space, Φ. The induced 
~
0 gðφ; θÞ ≤ 0
optimization problem is thus defined as I F ðφ; θÞ ¼ ð11Þ
~
1 gðφ; θÞ > 0
φ ¼ arg minPðFjφÞ ð8Þ
φ∈Φ The augmented limit state function g~ is obtained by superimpos-
ing a modeling error, η, to a noise–free limit state function, g, as
~
gðφ; θÞ ¼ gðφ; θÞ þ ηðφ; θÞ ð12Þ
Robust-to-Uncertainties Design of Seismic
Metamaterials where η = zero mean Gaussian distribution with standard deviation,
ση , that retains a certain influence on the outcome of the optimi-
Failure Probability zation process (Taflanidis 2007; Beck and Katafygiotis 1998).
It must be noted that such a modeling error can be alternatively
The failure probability, i.e., the probability that the SM performs in
assigned at any level of the model class, M, as, for example, di-
an unacceptable way, is expressed as a weighted integral of the
rectly at the model response (Papadimitriou et al. 2001).
form
Whereas other choices for gðφ; θÞ are applicable, for the pur-
Z
poses of this study the limit state function is defined either as the
PðFjφÞ ¼ I F ðφ; θÞpðθjφÞdθ ð9Þ exceedance of a threshold value (frequency domain analysis), or as
Θ
the peak response threshold (time domain analysis)
where I F ðφ; θÞ = failure indicator function with Boolean range,    
depending on whether the system performs below accepted levels ξ ðφ; θÞ maxjξðφ; θÞj
gðφ; θÞ ¼ log RMS gðφ; θÞ ¼ log
(value 1), or if all performance objectives are met (value 0). ϑRMS ϑ
In view of Eq. (8), it follows that the solution of the asso- ð13Þ
ciated optimization problem requires the evaluation of the integral
in Eq. (9). This is generally not possible analytically, except for In the case of a frequency domain analysis, the failure proba-
special cases (Papadimitriou et al. 1997), whereas numerical evalu- bility, PðFjφÞ, implies the probability that the root–mean-square
ation also becomes infeasible for high dimensional parameter (RMS) of the system response exceeds the set threshold value
spaces (Taflanidis 2007). A way to solve the integral in a reasonable ϑRMS . It is noted that the RMS value is directly related to the energy
computational time is through stochastic simulation techniques. content of the earthquake excitation, which is propagated through
Another approach is to use stochastic subset optimization to find the SM. As such, the left counterpart of Eq. (13) essentially refers to
a subset of the design variable space that has a high plausibility the overall value of a seismic event, rather than on a specific time-
for containing the optimal design point, φ (Taflanidis and Beck frequency interval of the typically nonstationary excitation, aiming

Fig. 7. Key steps for structural analysis performed in the frequency domain using a system model

© ASCE 04017181-6 J. Eng. Mech.

J. Eng. Mech., 2018, 144(3): 04017181

 at minimizing the total energy that reaches the protected structure. ~ ext
m
In a time domain setting, PðFjφÞ, represents the probability that the mSDOF ¼ 1 ~ int þ m
Nðm ~ ext Þ ¼ mSDOF m
~ int ¼ ð18Þ
r
maximum system response exceeds a set threshold, ϑ. Defining the
limit state function in the time domain comes with the added ad- whereas the adopted numerical values correspond to the means of
vantage of using a threshold, ϑ, that directly corresponds to typical the relative distributions. The sensitivity analysis is performed in
engineering quantities (e.g., absolute acceleration, relative dis- the frequency domain for a SM model with identical unit-cell setup
placement, spectral displacement/velocity/acceleration) as opposed and the results are displayed in Fig. 8.
to their RMS value.
Mass Ratio
Fig. 8 indicates that lower ratios, i.e., higher internal masses, yield a
Downloaded from ascelibrary.org by MICHIGAN TECHNOLOGICAL UNIVERSITY on 09/07/19. Copyright ASCE. For personal use only; all rights reserved.

Design Variables more favorable system response. Most importantly, however, it is

shown that the failure probability seems to be a monotonic function
From Eq. (1) it follows that the natural frequency of the internal
of the ratio. It is therefore not necessary to optimize with respect
resonator of the SM, f0;int , plays the most important role in the
to the unit-cell mass ratio, because lower ratio values will always
determination of the frequency band gap. Each element of the de-
yield a better protection. There is, however, a plateau below
sign vector, φ ∈ Φ
r ≈ 0.1, under which the mass ratio does not yield a significantly
φ ¼ ½φ1 ; φ2 ; : : : ; φnφ T ð14Þ lower failure probability. Based on this observation and the results
presented in Dertimanis et al. (2016), a mass ratio of
is thus defined as the common logarithm of some internal
~ ext
m
frequency r~ ¼ ¼ 0.1 ð19Þ
~ int
m
φj ¼ log10 ð2πf0;int;j Þ ð15Þ
is henceforth selected.
where the conversion is carried out to extend the design space into
higher internal frequencies, while maintaining a higher resolution External Stiffness
around the expected optimal frequencies. The external stiffness, kext , essentially represents the soil stiffness
For a given internal mass, mint;i , stored in the uncertain param- and its accurate and systematic treatment is outside the scope of
eter vector, θ, the natural frequency, f 0;int;j , of the ith internal res- the current study. However, very high external stiffness values es-
onator is directly related to its internal stiffness, kint;j , by sentially render the whole SM ineffective, as the excitation bypasses
the unit-cell resonators that are supposed to dissipate the seismic
kint;j ¼ ð2πf 0;int;j Þ2 mint;i ð16Þ energy. Low external stiffness values are generally desirable because
they decouple the system from the excitation. From our parametric
where j ¼ ½ði − 1Þmodnφ  þ 1 = design vector index, and i ¼
1; : : : ; N = unit-cell index. One further prescribes that
Nmodnφ ¼ 0, i.e., the number of design variables, nφ , has to be
a divisor of the number of cells, N, and that nφ ≤ N. This indexing
ensures that the assignment of the internal resonator’s properties is
taking place in alternating order. For example, a SM with N ¼ 6
unit cells and nφ ¼ 2 design parameters has φ ¼ ½kint;1 ; kint;2  as
the design vector and
kint ¼ ½kint;1 ; kint;2 ; kint;1 ; kint;2 ; kint;1 ; kint;2  ð17Þ

as the internal stiffness vector. The condition that Nmodnφ ¼ 0 can

generally be relaxed, but this will produce chains of unit cells with
some design variables being assigned more often than others. This
will presumably reduce the symmetry in the objective function,
which is not desirable in our case.

Sensitivity Analysis of Failure Probability

To assess the influence of the uncertain parameter vector, θ, on
the failure probability (the objective function), a sensitivity analysis
is performed for some of its key elements. To this end, the SM is
assumed to have a total mass of 1 (no unit), so that the stiffness
and the damping are expressed in units of 1=s2 and 1=s, respec-
tively. To distribute the total structural mass to the unit cells, the
following assumptions are made, loosely based on the work done
by Dertimanis et al. (2016):
1. The mass mSDOF of the protected oscillator equals unity.
2. The total SM mass is equal to the protected oscillator mass
Fig. 8. Failure probability estimator sensitivity to model parameters r,
and is distributed uniformly onto the unit cells.
kext , and ζ for an identical unit-cell chain with φ ¼ 0.25, 0.5, and 0.75
3. The external mass in each unit cell is equal to r times the mass
and an attached oscillator (G2 ); the failure probabilities were estimated
of the internal mass.
using standard Monte Carlo simulation with N ¼ 3,000 samples and
These assumptions can be translated into the following expres-
CRN
sions for the nominal parameters

© ASCE 04017181-7 J. Eng. Mech.

J. Eng. Mech., 2018, 144(3): 04017181

 study (Fig. 8), it follows that the stiffness has to be below ≈300 s−2 Table 1. Summary of the Univariate PDFs That Were Assigned to the
to ensure that the SM is effective by reaching failure probabilities Parameter Vector θ Assuming Independence
below ≈80% of the unprotected oscillator failure probability. This Vector Parameter PDF PDF parameters
can, however, always be achieved by increasing the number of
θs kext Lognormal μ ¼ logðk~ ext Þ, σ ¼ 0.01
unit cells. mint Lognormal μ ¼ logðm~ int Þ, σ ¼ 0.01
Fig. 8 shows that the failure probability estimator, P̂ðFjφÞ, is mext Lognormal μ ¼ logðm~ ext Þ, σ ¼ 0.01
also a monotonic function of the external stiffness. It is therefore f SDOF Lognormal μ ¼ logðf~ SDOF Þ, σ ¼ 0.2
not necessary to optimize this model parameter because the lowest ~ σ ¼ 0.01
ζ Lognormal μ ¼ logðζÞ,
stiffness will yield the lowest failure probability. The optimal de-
sign vector, φ , is also a function of the external stiffness, and this θϵ Mw Truncated exponential Mw;min ¼ 6, M w;max ¼ 8,
Downloaded from ascelibrary.org by MICHIGAN TECHNOLOGICAL UNIVERSITY on 09/07/19. Copyright ASCE. For personal use only; all rights reserved.

can be justified by considering that increasing the external stiffness μ¼1

causes an overall system stiffness increase, which corresponds to R Lognormal μ ¼ logð20Þ, σ ¼ 0.5
higher natural frequencies. Because the optimal unit-cell parame- Zw Gaussian white noise μ ¼ 0, σ ¼ 1
ters are usually clustered around these resonant frequencies, it is γp Normal μ ¼ 1.8, σ ¼ 0.3
π π
thus not surprising that stiffer systems (i.e., high kext ) have a lower νp Uniform ν p;min ¼ − , ν p;max ¼
2 2
failure probability if their unit-cell resonant frequencies are high as εf Normal μ ¼ 0, σ ¼ 0.57
well. The same can be said about soft systems (i.e., lower kext ) and εv Normal μ ¼ 0, σ ¼ 0.49
lower internal unit-cell resonant frequencies. θw η Normal μ ¼ 0, ση ¼ 0.01
These insights lead to the following adoption of the external Note: The individual parameters are explained in the section “Problem
stiffness expected value: Formulation.”
1
k~ ext ¼ 100 2 ð20Þ
s
adopted here is the two-stage approach originally presented in
Taflanidis and Beck (2008a) that is based on the gradient-free sto-
Damping Ratio chastic subset optimization algorithm (SSO). SSO identifies a rel-
Expectedly, higher damping ratios bear a positive influence on all atively small region in the design space, Φ, that is likely to contain
response quantities and reduce the failure probability. In the fre- the optimal value, φ . Yet, whereas the SSO algorithm converges
quency domain, this is evident by the smoothing of the response in quickly to a relatively small subset, I  , with a high likelihood of
the areas of resonant peaks, whereas in the time domain damping containing φ , it generally requires a high number of function eval-
directly influences the response settling times (Dertimanis et al. uations to converge to the actual optimum. This is especially true if
2016). Based on reported typical ranges [e.g., 2–8% (Chopra the objective function is relatively flat in the last identified subset.
2007)], and adopted values in similar structures (Finocchio et al. This is equivalent to a low sensitivity of the objective function and
2014), a mean value of decreases the plausibility of the subsets containing the optimal
design variables.
ζ~ ¼ 5% ð21Þ
It is therefore beneficial to switch to a different optimization al-
is chosen. Following this sensitivity analysis, the presented values gorithm when a sufficiently small region in the initial design space,
as random variables are chosen in what follows. Φ, has been identified. This is accomplished through the implemen-
tation of the simultaneous perturbation stochastic approximation
(SPSA) algorithm (Kleinman et al. 1999; Spall 2003). The SPSA
Model Parameter Probability Density Functions
is based on the observation that during an optimization process,
A probability density function (PDF), pðθjMÞ, conditional on the the same level of accuracy can be achieved when the gradient vector
selected model class M is assigned to the uncertain parameter vec- is the result of a random direction perturbation, as opposed to a per-
tor, θ. All PDFs are assumed conditional on the chosen model class, turbation in each dimension of the design space (Spall 2003).
but for brevity, the explicit expression of M is omitted. In conform- In what follows, a brief outline of the two-stage SM optimiza-
ity with the Bayesian interpretation of probability, the PDFs reflect tion process is described.
the uncertainty in our knowledge of these parameters, and are there-
fore explicitly designated as carriers of information (Jaynes 2003), Shape of Initial Design Space
instead of being considered as some form of randomness. To solve the posed optimal design problem for the SM chain, the
Attributable to the lack of representative experimental data, the chosen design variables have to be limited initially. This is done
PDFs of the SM model parameters, pðθs Þ, are assigned according through the introduction of an initial admissible design space, Φ,
to engineering judgment, supported by the results of the parametric which typically corresponds to some geometrical shape with the
sensitivity analysis previously presented. These values are used as dimension of the design problem (e.g., nφ , see Appendix). The sim-
the mean values, θ~ i , and are assigned a COV of 1% to partially ac- plest such shape is the nφ -dimensional hypercube, where for each
count for the uncertainty in these parameters. The PDFs for the design variable, φi , an upper and lower border has to be defined.
excitation model parameters, pðθϵ Þ, are taken directly from Fig. 9 displays a sample estimate of the failure probability ob-
Taflanidis (2007), whereas for the modeling error, a zero mean jective function in a two-dimensional design problem for a six-unit-
normal PDF is chosen. A summary of the PDFs assigned to the cell SM, where Φ pertains to a two-dimensional hypercube with
parameter vector, θ, is given in Table 1. lower and upper borders φi ¼ ½0.3; 0.6. Apparently, the objective
function exhibits some symmetry characteristics. Assuming that
this attribute is also valid for an arbitrary number of unit cells,
Optimization N, the initial design space is selected to be an nφ -simplex, where
The optimization problem defined in Eq. (8) can be treated in for each design variable the following is true:
many ways, including gradient-based and gradient-free methods
(Ruszczyński and Shapiro 2003; Spall 2003). The approach φi < φiþ1 ; ∀ i ∈ ½1; nφ  ð22Þ

© ASCE 04017181-8 J. Eng. Mech.

J. Eng. Mech., 2018, 144(3): 04017181

 (a) (b) (c)

Fig. 9. Surface and contour plots of the failure probabilities in a six-unit-cell chain with nonidentical unit cells with RMS threshold ϑRMS ¼ 0.4 in the
limit state function, where φ ≈ ð0.44; 0.54Þ; the dotted line represents the identical unit-cell case; the failure probabilities were estimated using
Downloaded from ascelibrary.org by MICHIGAN TECHNOLOGICAL UNIVERSITY on 09/07/19. Copyright ASCE. For personal use only; all rights reserved.

Monte Carlo simulations with N ¼ 1,000 independent samples; at the far right the FRFs for both systems in the optimal 1D and 2D design cases are
presented

This places all admissible design variable vectors, φ, in a space 3. Find optimal subset, I  . This is the subset, I, with the mini-
where the individual design variables, φi , are arranged in ascend- mum sensitivity that contains a maximum of ρN samples:
ing order.
NI
I  ¼ arg minHðIÞ ≈ arg min ĤðIÞ ¼ arg min ð24Þ
SSO I∈Φ I∈Φ I∈Φ VI
The SSO algorithm (Taflanidis and Beck 2008a), with an initial
design space, Φ, in the shape of an nφ -dimensional simplex, iter- 4. Delete all samples that lie outside this identified subset, I.
atively identifies subsets I i in the shape of hyperelliptical segments, Compute Ĥ and COV Ĥ [Eq. (23)].
D, (see Appendix 7), which are likely to contain the optimal design 5. If Ĥ and COV Ĥ do not exceed the set threshold values, re-
variable, φ . With respect to Fig. 10, the SSO consists of the fol- start at Step 1, otherwise stop the algorithm.
lowing steps: The SSO identifies ever smaller subsets in the initial design
1. Initialize the SSO by selecting a desired number of samples, space, Φ. As stated in Taflanidis and Beck (2008a), it is desirable
N SSO , and a constraint value, ρ. Specify a threshold value for the to switch to a different algorithm to pinpoint the exact optimal sol-
average relative sensitivity of pðφjFÞ to φ: Ĥ and COV Ĥ as stop- ution, φ . The efficiency of the SSO algorithm depends strongly on
ping criteria. The sensitivity at each step is estimated as the chosen subset shape and sample ratio, ρ. In accordance with the
guidelines presented in Taflanidis (2007), the sample ratio is chosen
N Ii =V Ii to ρ ¼ 0.15.
HðIÞ ≈ ĤðIÞ ¼ ð23Þ
N Ii−1 =V Ii−1 SPSA with CRN
The SPSA optimization, as introduced in Spall (2003) and extended
where N I designates the number of samples in the subset and, V I , by Kleinman et al. (1999) to include a variance reduction technique
the subset volume. known as common random numbers (CRN), takes the typical re-
2. Use a sampling procedure to simulate N SSO samples from cursive form
pðφjFÞ (e.g., MCS, or MCMC). This corresponds to simulating
failed samples in the design variable space. φkþ1 ¼ φk þ ak ĝk ðφk Þ ð25Þ

(a) (b) (c)

(d) (e) (f)

Fig. 10. SSO algorithm for a reliability objective problem of a SM consisting of six unit cells with two independent properties [φ ¼ ðφ1 ; φ2 Þ]; the last
step shown is just before the stopping criteria were fulfilled; the optimal solution is marked with a cross and was pinpointed using the SPSA algorithm:
(a) algorithm Step 1; (b) algorithm Step 2; (c) algorithm Step 3; (d) algorithm Step 3—magnification; (e) algorithm Step 1; (f) algorithm Step 2

© ASCE 04017181-9 J. Eng. Mech.

J. Eng. Mech., 2018, 144(3): 04017181

 where the gradient is estimated as described. Different facets of the optimization outcome are consid-
ered in the following, including the influence of an increasing num-
P̂ðFjφk þ ck Δk;nφ ; Ωk;N Þ − P̂ðFjφk − ck Δk;nφ ; Ωk;N Þ ber of different unit cells on the optimal solution, the claimed
ĝk ðφk Þ ¼
2ck Δk;nφ symmetry of the objective function, and the influence of the exter-
nal stiffness.
ð26Þ

with the same set of random numbers for the positive and neg- Number of Independent Unit Cells nφ
ative perturbation. In Spall (2003), numerous admissible proba- The aim herein is to investigate whether the failure probability can
bility distributions for the perturbation direction are presented be reduced if an increasing number of differing unit cells is used.
(e.g., segmented uniform, u-shaped and symmetric Bernoulli 1). The actual number of unit cells, N, is kept fixed, whereas the num-
Downloaded from ascelibrary.org by MICHIGAN TECHNOLOGICAL UNIVERSITY on 09/07/19. Copyright ASCE. For personal use only; all rights reserved.

Herein, the symmetric Bernoulli 1 distribution is adopted, result- ber of design variables is increased. The unit-cell properties are then
ing in possible values for the perturbation vector Δnφ merely deduced from the design variables as already discussed.
The results are visualized in Table 2 and Figs. 12 and 13. The
Δnφ ¼ ½δ 1 ; : : : ; δ nφ ; with δ i ¼ f−1; 1g ∀ i ð27Þ highest failure probability is observed in the unprotected case
(kext ¼ ∞ s−2 ), followed by the disconnected case (kint ¼ 0 s−2 ),
The algorithm is initialized at the center of the last identified whereas the lowest failure probability, and therefore optimal de-
subset and normalizes the search space to be a unit radius nφ ball. sign, is achieved when all unit-cell properties are chosen independ-
This is beneficial if the subsets extend differently in the various ently. Fig. 12 reveals why a higher number of different cells does
dimensions. not necessarily imply a positive influence on the overall response.
Four illustrative runs of the algorithm starting at the center of the Each resonator effectively reduces one peak response value. In this
last identified subset φ ¼ ½0.4; 0.6 are shown in Fig. 11. It is a context, the SM does not create effective band gaps, i.e., locations
direct continuation of the last identified subset in the preceding in the system’s FRF that are smaller than 0 dB, but acts more like a
paragraph’s SSO example. The algorithm takes different paths form of damping.
but converges to similar optima, φ . It is further not forced to re- The limit state function in this case considers only the absolute
main within the last identified subset. A boundary condition is en- acceleration as a design criterion. Whereas the estimated optimum
forced only if the initial design space, Φ, is left. design variables, φ , are not guaranteed to be the optimal design
choice if relative displacement were a design criterion, it is worth
noting that the relative displacements are also reduced by the present
Parametric Study variable choice. That is, the displacement response is reduced with
respect to the initial unprotected system rather than amplified for the
optimal acceleration design variables, φ , as seen in Fig. 13.
Frequency Domain Analysis
The SM chain with an attached oscillator (G2 ) is now investigated Symmetry of the Optimal Design Variables φ

in the frequency domain, where the attached oscillator serves as It was already observed that the optimal design variables lie at
a simplified model for a protected structure. The optimization is points that lead to the flattest FRF. A pattern that emerges in higher
carried out by the two-stage optimization framework previously design dimensions is how these points are sequentially located at

(a) (b) (c) (d)

Fig. 11. Four runs of the SPSA algorithm starting at the center of the last identified subset φ ¼ ½0.4; 0.6 to pinpoint the optimal solution φ :
(a) P̂ðFjφ̂ Þ ¼ 3.42%; (b) P̂ðFjφ̂ Þ ¼ 3.46%; (c) P̂ðFjφ̂ Þ ¼ 3.52%; (d) P̂ðFjφ̂ Þ ¼ 3.49%

© ASCE 04017181-10 J. Eng. Mech.

J. Eng. Mech., 2018, 144(3): 04017181

 Table 2. Optimal Parameters, φ , Using Model G2 with kext ¼ 100 s−2 Determined through the Two-Stage Optimization Framework, for Different Numbers
of Independent Unit Cells nφ ¼ 1, 2, 3, and 6 and the Disconnected Case (kint ¼ 0 s−2 ) and the Unprotected Case (kext ¼ ∞ s−2 )
k~ ext ¼ 100 s−2
φ and P̂ðFjφÞ kext ¼ ∞ s−2 kint ¼ 0 s−2 nφ ¼ 1 nφ ¼ 2 nφ ¼ 3 nφ ¼ 6
φ1 jf 1 Hz — — 0.4794j0.48 0.4175j0.42 0.3485j0.36 0.2147j0.26
φ2 jf 2 Hz — — — 0.5494j0.56 0.4481j0.45 0.4022j0.40
φ3 jf 3 Hz — — — — 0.5722j0.59 0.4121j0.41
φ4 jf 4 Hz — — — — — 0.4142j0.41
φ5 jf 5 Hz — — — — — 0.5025j0.51
Downloaded from ascelibrary.org by MICHIGAN TECHNOLOGICAL UNIVERSITY on 09/07/19. Copyright ASCE. For personal use only; all rights reserved.

φ6 jf 6 Hz — — — — — 0.6081j0.65
P̂ðFjφÞ 59.2% 37.8% 24.8% 20.23% 18.03% 17.13%
Note: The failure probabilities, P̂ðFjφÞ, were estimated using Monte Carlo simulations with N ¼ 3,000 samples and a limit state function with a threshold of
ϑ ¼ 0.4 m s−1 as defined in Eq. (13).

(a) (b)

(c) (d)

(e) (f)

Fig. 12. Magnitude of the absolute acceleration FRF jHẍ ðfÞj for the optimal variables for the acceleration design criterion, found through the two-
stage optimization framework; the internal frequencies, f i , from the design variables are shown as black lines; the parameters, θ, stem from a sample
realization in the parameter space, Θ: (a) unprotected oscillator; (b) kint ¼ 0 s−2 ; (c) nφ ¼ 1; (d) nφ ¼ 2; (e) nφ ¼ 3; (f) nφ ¼ 6

peaks of the FRF obtained by the optimal design in nφ − 1 dimen- 1. The optimum in one design space, φ ∈ Φ, is not at the same
sions (Fig. 12). Assuming that the order in which the unit-cell prop- location in the mirrored design space, Φ ~ i.
erties are assigned is of secondary importance to the failure 2. The optimum is at the same location, but some variable ar-
probability, one can deduce that there are totally nφ local optima rangements are more favorable [with respect to the failure proba-
in nφ simplices that are symmetrical to our initial design space, Φ. bility, PðFjφÞ] than others.
The optimal solutions for this initial design space for an increasing Although this is no proof, the observations made in Fig. 9
number of independent unit cells are presented in Table 2. To fur- indicate that variables in the design space, Φ, do generally yield
ther explore the above assumption about the secondary importance slightly different failure probabilities from their mirrored counter-
of the property assignment order, a sensitivity analysis is performed parts. This would point to the adoption of the second explanation.
by changing the order of the design variables and observing the So, to identify the global optimum, it would be necessary to inves-
failure probabilities, as shown in Table 3. This effectively places tigate the failure probability in the mirrored design spaces as well.
~ The analysis is con- This would mean expanding the initial design space from a simplex
the solution in mirrored design spaces, Φ.
to a hypercube. The high number of local optima, in high dimen-
ducted for the nφ ¼ 2 and three-dimensional design spaces only,
sions, would, however, require the use of different types of optimi-
because the nφ ¼ 6 dimensional design space already has 6! − 1 ¼
~ i . The results show that the failure zation algorithms.
719 mirrored design spaces, Φ
probability is in fact not the same if the results lie in the alternate External Stiffness kext Influence on φ

design spaces, Φ ~ i , (even when considering the COV of 1%). Two The optimization of the previous section is repeated for different
possible explanations for this are external stiffness values (k~ ext ¼ 100 , 500, and 1,000 s−2 ), and

© ASCE 04017181-11 J. Eng. Mech.

J. Eng. Mech., 2018, 144(3): 04017181

 (a) (b)
Downloaded from ascelibrary.org by MICHIGAN TECHNOLOGICAL UNIVERSITY on 09/07/19. Copyright ASCE. For personal use only; all rights reserved.

(c) (d)

(e) (f)

Fig. 13. Magnitude of the relative displacement FRF jHu ðfÞj for the optimal variables for the acceleration design criterion, found through the two-
stage optimization framework; the internal frequencies, f i , from the design variables are shown as black lines; the parameters, θ, stem from a sample
realization in the parameter space, Θ: (a) unprotected oscillator; (b) kint ¼ 0 s−2 ; (c) nφ ¼ 1; (d) nφ ¼ 2; (e) nφ ¼ 3; (f) nφ ¼ 6

Table 3. Symmetry of the Optimal Solution, φ , by Calculating the Failure Whereas, for example, P̂ðFjφÞ reduces by 4.57% between nφ ¼ 1
Probabilities for the Optimal Design Variables from Table 2 in Different and nφ ¼ 2, the reduction between nφ ¼ 3 and nφ ¼ 6 is only
Mirrored Design Spaces, Φ or Φ ~i
0.9%. This could be explained by looking again at the FRF in
Design space nφ ¼ 2 nφ ¼ 3 Fig. 12. The SM structure only has one predominant natural fre-
φ
∈Φ P̂½Fjðφ1 ; φ2 Þ ¼ 20.23% P̂½Fjðφ1 ; φ2 ; φ3 Þ ¼ 18.03%
quency, f0 , in the seismically active range, therefore the response
φ ~1
∈Φ P̂½Fjðφ2 ; φ1 Þ ¼ 19.90% P̂½Fjðφ3 ; φ2 ; φ1 Þ ¼ 20.77% reduction occurs mostly around this frequency. After nφ ¼ 3, the
φ ~2
∈Φ — P̂½Fjðφ3 ; φ1 ; φ2 Þ ¼ 19.97% FRF is already flattened out considerably, with no major peaks re-
φ ~3
∈Φ — P̂½Fjðφ2 ; φ3 ; φ1 Þ ¼ 19.77% maining. Allowing additional independent unit cells does not help
φ ~4
∈Φ — P̂½Fjðφ2 ; φ1 ; φ3 Þ ¼ 18.53% with increasing the protection because there are no more predomi-
φ ~5
∈Φ — P̂½Fjðφ1 ; φ3 ; φ2 Þ ¼ 18.50% nant peaks to smooth out. In the case of multiple response peaks
Note: The failure probabilities P̂ðFjφÞ were estimated using Monte Carlo [e.g., multidegree-of-freedom (MDOF) protected oscillator system],
simulations with N ¼ 3,000 samples. The COV therefore is ≈1%. further failure probability reduction can be expected.
Another observation is the behavior of the optimal variable
P
mean, f̄  ¼ n1φ f i . As seen in Fig. 15, the mean tends to follow
the resulting optimal design variables, φ , are summarized in the fundamental frequency of the disconnected case, f 0 . This is not
Tables 2–5. It is observed that the failure probability decreases with surprising because the fundamental frequency is where the pre-
a higher number of independent unit cells (i.e., nφ ). Fig. 14 gives a dominant response is expected, and the response reduction is there-
visual representation of this performance. fore most effective in its proximity. The optimal frequencies tend to
A first observation is that the protection gets worse for higher lie below f0 as Fig. 4 indicates, where the band gap locations are
external stiffness, kext . More interesting, however, is the fact that shown. The band gaps, or in the finite case effective reduction
with a higher external stiffness, increasing the number of indepen- zones, start at the natural frequency of the unit cells f 0;int , but
dent unit cells does not significantly decrease the failure probability. extend to higher frequencies depending on the band gap width.

Table 4. Optimal Parameters, φ , Using Model G2 with kext ¼ 500 s−2 Determined through the Two-Stage Optimization Framework, for Different Number
of Independent Unit Cells nφ ¼ 1, 2, 3, and 6 and the Disconnected Case (kint ¼ 0 s−2 ) and the Unprotected Case (kext ¼ ∞ s−2 )
k~ ext ¼ 500 s−2
φ and P̂ðFjφÞ kext ¼ ∞ s−2 kint ¼ 0 s−2 nφ ¼ 1 nφ ¼ 2 nφ ¼ 3 nφ ¼ 6
φ1 jf 1 Hz — — 0.6614j0.73 0.6148j0.66 0.6018j0.64 0.2029j0.25
φ2 jf 2 Hz — — — 0.6883j0.78 0.6329j0.68 0.4976j0.50
φ3 jf 3 Hz — — — — 0.7113j0.82 0.4981j0.50
φ4 jf 4 Hz — — — — — 0.5153j0.52
φ5 jf 5 Hz — — — — — 0.6022j0.64
φ6 jf 6 Hz — — — — — 0.6948j0.79
P̂ðFjφÞ 59.2% 54.77% 45.63% 43.10% 42.80% 43.12%
Note: The failure probabilities P̂ðFjφÞ were estimated using Monte Carlo simulations with N ¼ 3,000 samples and a limit state function with a threshold of
ϑ ¼ 0.4 m s−1 as defined in Eq. (13).

© ASCE 04017181-12 J. Eng. Mech.

J. Eng. Mech., 2018, 144(3): 04017181

 Table 5. Optimal Parameters φ Using Model G2 with kext ¼ 1,000 s−2 Determined through the Two-Stage Optimization Framework, for Different Number
of Independent Unit Cells nφ ¼ 1, 2, 3, and 6 and the Disconnected Case (kint ¼ 0 s−2 ) and the Unprotected Case (kext ¼ ∞ s−2 )
k~ ext ¼ 1,000 s−2
φ and P̂ðFjφÞ kext ¼ ∞ s−2 kint ¼ 0 s−2 nφ ¼ 1 nφ ¼ 2 nφ ¼ 3 nφ ¼ 6
φ1 jf 1 Hz — — 0.6952j0.79 0.6253j0.67 0.6356j0.69 0.1451j0.22
φ2 jf 2 Hz — — — 0.7198j0.83 0.7073j0.81 0.4847j0.49
φ3 jf 3 Hz — — — 0.7077j0.82 0.5068j0.51
φ4 jf 4 Hz — — — — — 0.6440j0.70
φ5 jf 5 Hz — — — — — 0.6505j0.71
Downloaded from ascelibrary.org by MICHIGAN TECHNOLOGICAL UNIVERSITY on 09/07/19. Copyright ASCE. For personal use only; all rights reserved.

φ6 jf 6 Hz — — — — — 0.7289j0.85
P̂ðFjφÞ 59.2% 55.63% 50.27% 49.97% 49.93% 50.33%
Note: The failure probabilities, P̂ðFjφÞ, were estimated using Monte Carlo simulations with N ¼ 3,000 samples and a limit state function with a threshold of
ϑ ¼ 0.4 m s−1 as defined in Eq. (13).

Fig. 16. Results from an optimization run using a time domain analysis
Fig. 14. Failure probability estimators for the optimal design variables, up until the last converged solution; the optimal solution estimator, φ̂ ,
φ , for different numbers of independent unit cells, nφ ; the case of displayed is the one obtained from a frequency domain optimization;
nφ ¼ 0 corresponds to the nonconnection case (kint ¼ 0 s−2 ); the fail- the number of SSO samples used were N ¼ 1,000 with a sample ratio
ure probabilities were estimated using standard MCS with N ¼ 3,000 of ρ ¼ 0.15
samples and a limit state function with a threshold of ϑ ¼ 0.4 m s−1 as
defined in Eq. (13)

To effectively reduce a peak response, it is therefore necessary to

ensure the band gap covers this peak response value. Therefore,
unit cells with internal frequencies, f i , that are lower than f0 are
required.

Time Domain Analysis

Time domain optimization under uncertainty is by default computa-
Fig. 17. Comparison of the attached oscillator absolute accelera-
tionally expensive, attributable to the relatively long time it takes to
tion using model G2 with linear and nonlinear internal springs; the
evaluate the objective function. The presented SSO algorithm is
effectivity of the SM is slightly reduced, if the internal springs deform
very efficient in that case and allows a region to be identified that
nonlinearly
is likely to contain the optimal solution. Owing to computational

(a) (b) (c) (d)

Fig. 15. Location of the optimal frequencies, f , depending on the external stiffness, kext , for different number of design variables, nφ ; the plots also
include the fundamental frequency for the disconnected case (kint ¼ 0 s−2 ) as a function of the external stiffness; additionally, the mean of the optimal
frequencies, f̄ , is shown: (a) nφ ¼ 1; (b) nφ ¼ 2; (c) nφ ¼ 3; (d) nφ ¼ 6

© ASCE 04017181-13 J. Eng. Mech.

J. Eng. Mech., 2018, 144(3): 04017181

 limitations, the second step of the two-stage framework is not ac- domain responses and corresponds to a design quantity of great
tivated herein. interest for engineering purposes.
To investigate whether the time domain optimization converges The results of a sample SSO process are displayed in Fig. 16
to a similar optimum as the frequency domain optimization in together with the optimal solution from the frequency domain op-
the previous section, the optimization is repeated for the case of timization. The last identified subset does, in fact, contain the op-
nφ ¼ 2. The chosen setup is again the sixth unit-cell chain with timal solution from the frequency domain. This does not imply that
an attached linear oscillator. The limit state function is the maxi- the optimal solutions for both analysis are located at the same point,
mum response threshold of the absolute acceleration response with they are however likely to be in close proximity. This indicates that
ϑ ¼ 2.5 m s−2 , as defined in Eq. (13). This is well suited for time a frequency domain optimization could be used to obtain a low
Downloaded from ascelibrary.org by MICHIGAN TECHNOLOGICAL UNIVERSITY on 09/07/19. Copyright ASCE. For personal use only; all rights reserved.

(a) (b)

(c) (d)

(e) (f)

(g) (h)

(i) (j)

(k) (l)

Fig. 18. Comparison of the relative displacements between the internal and external mass of each unit cell in the six cell SM using model G2 and
linear and nonlinear internal springs; for brevity, the relative displacements are numbered as follows: ui ¼ ui;2 − ui;1 ; the design variables for the unit
cells were chosen according to φ from M2 for the case of nφ ¼ 6: (a) uðtÞ ¼ u1;2 − u1;1 ; (b) R1 ; (c) uðtÞ ¼ u2;2 − u2;1 ; (d) R2 ; (e) uðtÞ ¼ u3;2 − u3;1 ;
(f) R3 ; (g) uðtÞ ¼ u4;2 − u4;1 ; (h) R4 ; (i) uðtÞ ¼ u5;2 − u5;1 ; (j) R5 ; (k) uðtÞ ¼ u6;2 − u6;1 ; (l) R6

© ASCE 04017181-14 J. Eng. Mech.

J. Eng. Mech., 2018, 144(3): 04017181

 failure probability subset that is likely to contain the optimal sol- The authors found that a higher number of independently chosen
ution of the time domain analysis. From a computational perspec- internal stiffnesses will generally increase the efficiency of the SM
tive, this is very helpful because the time required to obtain the in mitigating the vibration response when compared with identically
presented time domain results was approximately 50 times higher chosen unit-cell properties. This study thereby confirmed the find-
(≈2 h for the frequency and ≈100 h for the time domain optimi- ings of Krödel et al. (2015) and were able to generalize them to
zation). This discrepancy in computational time is expected to in- uncertain seismic excitations. The findings also showed the mechan-
crease further if SPSA were to be employed as well, because of the ics the SM approach has in common with well-established protec-
higher number of required objective-function evaluations. tion methods (BI and TMD).
It was further shown that the external stiffness is a key factor in
the SM design process. It not only directly influences the failure
Downloaded from ascelibrary.org by MICHIGAN TECHNOLOGICAL UNIVERSITY on 09/07/19. Copyright ASCE. For personal use only; all rights reserved.

Nonlinear Internal Springs probability, but also indirectly determines the location of the optimal
As a final investigation, the introduction of nonlinear springs that resonator frequencies. When considering soil as the medium be-
can potentially be used to obtain a more realistic system response is tween the individual unit cells, additionally the linearity assumption
examined. It must be noted that performing optimization for non- of the external springs is not justified, and their nonlinearity has to
linear SM structures would imply the inclusion of the nonlinear be taken into account.
parameters in the uncertain parameter vector, θ, but at a prohibitive Initially, the goal was to carry out the optimization in the time
computational cost for the time domain analysis. To this end, domain, owing to computational limitations, however, the fre-
the system model, G2 , with the optimized model parameters, φ , quency domain was chosen as a viable alternative. This is not nec-
for the nφ ¼ 6 independent unit-cell case with k~ ext ¼ 100 s−2 is essarily a bad choice, because the frequency reduction zones are
reexamined with the integration of nonlinear internal springs. A only visible in the frequency domain, it does however prohibit the
hysteretic Bouc-Wen model (Charalampakis and Koumousis 2008) use of nonlinear models and does not take into account the transient
is utilized, with post yield stiffness kpost ¼ 10%kint , β ¼ 1, response of the structure. For linear systems, the frequency domain
γ ¼ 0.5, n ¼ 3, and fy ¼ 0.01 m s−2 . approach delivers satisfying results.
Fig. 17 indicates how the absolute acceleration of the protected From an algorithmic standpoint, the optimization becomes more
linear oscillator changes depending on whether linear or nonlinear challenging with an increase in dimensionality, as for every added
internal oscillator springs are used. A slight reduction in the effec- independent resonator property one expects a factorial increase in
tiveness of the SM in protecting the linear oscillator is detected. local optimal solutions. To circumvent this problem, the authors
When the oscillator reaches its yield deformation, dy , its effective chose to optimize only for resonators that have an increasing inter-
stiffness is greatly reduced to 10% of its original value. This re- nal stiffness. This places the design space in a simplex with only
duces the oscillator’s natural frequency and moves it away from one expected optimum. This is a limitation that should be addressed
the optimized frequencies. If the frequencies had been tuned to in future works.
higher frequencies, this might have a positive effect on the SM ef- On a similar note, the arrangement of the resonators has to be
fectiveness. The internal springs are ordered with an ascending considered in greater detail. In this paper, these were arranged as
internal stiffness (Fig. 18), which is seen in the hysteretic curves. homogeneously as possible, to obtain a more symmetric objective
As expected, the stiffer oscillators attract more force, and therefore function and thus justify the reduction of the design space to the
deform more than the soft oscillators. This becomes especially ap- mentioned simplex. This assumption will be relaxed and form a
parent in the relative displacement, u6 , where the peak deformation further design parameter in future work.
is approximately twice the deformation in the linear case.

Appendix. Geometric Shapes

Conclusion
The geometric shapes used throughout this paper are introduced
In this paper, the tools of robust-to-uncertainties optimal reliability here. They are used for the initial design space, Φ, and the subsets,
design were applied on the relatively new concept of metadamping. I i . The shape of I i should be chosen so that the challenging opti-
This was achieved by modeling a seismic metamaterial as a one- mization [Eq. (24)] can be solved efficiently. It is also beneficial,
dimensional unit-cell chain incorporating internal resonators that especially in the final iterations of the SSO algorithm, if the subset
show metadamping characteristics. Previous work demonstrated shape is similar to the objective function (Taflanidis 2007). Shapes
the effectivity of such metamaterials in reducing applied excitations that have been proposed in the past are the hyper-rectangle
in certain frequency ranges, but showed also excessive responses (Taflanidis and Beck 2008b) and the hyperellipse (Taflanidis and
outside of these reduction zones. Expanding on these reports, the Beck 2009). The shape chosen in this work is a hyperellipse seg-
authors used the tools of robust to uncertainties optimization to ment that moves towards a full hyperellipse in the course of the
optimize the SM model properties to produce the best possible re- algorithm. This adaptive shape of the subset was found to satisfy the
sponse reduction when excited by uncertain seismic events. problems posed by the problem formulation. Namely, to lie inside
It was found that monotonic relationships may be inferred be- the initial design space, Φ, that has the form of an (nφ − 1)-simplex
tween certain model parameters and the failure probability that was and adaption to the shape of the objective function in later algorithm
optimized. Higher external stiffness values will always lead to an stages.
inferior vibration behavior. Higher ratios between the external and
internal mass will have the same effect, and naturally so will lower Simplex
damping ratios. A higher number of unit cells, as previously re-
ported (Wagner et al. 2016b), will result in a better vibration mit- As described previously, the initial design variable space, Φ, has
igation inside the target frequency range. To this end, it was decided the form of an (nφ − 1)-simplex. Formally this simplex is a subset
that optimization is focused on exclusively varying the internal of the nφ hypercube given by
stiffness properties (i.e., internal natural frequency) of the unit-cell
resonators with a fixed number of six unit cells. S ¼ fφ ¼ ðφ1 ; : : : ; φnφ Þ ∈ Rnφ jφi < φiþ1 ∀ ig ð28Þ

© ASCE 04017181-15 J. Eng. Mech.

J. Eng. Mech., 2018, 144(3): 04017181

 This simplex contains all design points with monotonically in- kext = unit cell external mass;
creasing coordinates. In two dimensions, this simplex is given by a M= stochastic system model class;
triangle, in three dimensions by a tetrahedron. m̂ = equivalent chain mass;
mext = unit cell external mass;
Hyperellipse mint = unit cell internal mass;
A generalization of the ellipse to higher dimensions is the so called N= number of unit cells;
hyperellipse. It is defined in the design space, Φ, as nθ = dimensions parameter space;
nφ = dimensions design space;
E ¼ fφ ¼ ðφ1 ; : : : ; φnφ Þ ∈ Rnφ jðφ − cÞ⊺ Mðφ − cÞ ≤ 1g ð29Þ Δnφ = perturbation vector;
Downloaded from ascelibrary.org by MICHIGAN TECHNOLOGICAL UNIVERSITY on 09/07/19. Copyright ASCE. For personal use only; all rights reserved.

ϵ= seismic excitation;
where M ∈ Rnφ ×nφ is positive-definite symmetric matrix defining
η= model prediction error;
the size and orientation of the hyperellipse and c ∈ Rnφ is its center.
Note that the above can be equivalently expressed as ζ= damping ratio;
Θ= parameter space;
E‘ ¼ fφ ¼ ðφ1 ; : : : ; φnφ Þ ∈ Rnφ jjjL⊺ ðφ − cÞjj ≤ 1g ð30Þ θ= parameter vector;
ξ= system response;
where L = Cholesky decomposition of the matrix, M. ση = standard deviation modeling error;
Φ= design variable space;
Hyperellipse Segment Φ ~ = mirrored design space;
In the first step of the SSO algorithm, a subset inside the simplex is φ= design variable vector;
identified by performing the optimization in Eq. (24). This corre- φ = optimal design variable vector;
sponds to finding the subset with the largest volume, that contains a ω̂ = equivalent natural frequency;
high enough samples ratio, ρ. This challenging optimization gets ω0;int = internal fundamental frequency; and
increasingly harder in high dimensions. Already in low dimensions, ωint = internal natural frequency.
it is difficult for ellipse shapes to cover the whole simplex.
Problems arise especially in the corners, which cannot be reached
without reducing the samples ratio, ρ. For this reason, the chosen References
subset shape is a hyperellipse segment.
This segment is defined by Achaoui, Y., Ungureanu, B., Enoch, S., Brûlé, S., and Guenneau, S. (2016).
“Seismic waves damping with arrays of inertial resonators.” Extreme
D ¼ S ∩ E ≔ fφ ¼ ðφ1 ; : : : ; φnφ Þ Mech. Lett., 8, 30–37.
Au, S.-K., and Beck, J. L. (2003). “Subset simulation and its application
∈ Rnφ j½φi < φiþ1 ∀ i ∧ ½ðφ − cÞ⊺ Mðφ − cÞ ≤ 1g ð31Þ
to seismic risk based on dynamic analysis.” J. Eng. Mech., 10.1061
/(ASCE)0733-9399(2003)129:8(901), 901–917.
that is, it is defined by the intersection of points that lie within the
Beck, J. L., and Katafygiotis, L. S. (1998). “Updating models and
simplex, S, and the ellipse, E.
their uncertainties. I: Bayesian statistical framework.” J. Eng. Mech.,
10.1061/(ASCE)0733-9399(1998)124:4(455), 455–461.
Boore, D. M. (2003). “Simulation of ground motion using the stochastic
Acknowledgments method.” Pure Appl. Geophys., 160(3), 635–676.
Charalampakis, A., and Koumousis, V. (2008). “Identification of Bouc–
This paper is based on a master’s thesis carried out during a stay
Wen hysteretic systems by a hybrid evolutionary algorithm.” J. Sound
of one of the authors as a visiting student researcher at Caltech. Vibr., 314(35), 571–585.
The authors thank Prof. Dr. Alexandros Taflanidis for his support Cheng, Z., Shi, Z., Mo, Y. L., and Xiang, H. (2013). “Locally resonant
during the implementation of the subset optimization algorithm. periodic structures with low-frequency band gaps.” J. Appl. Phys.,
Additional thanks also to Prof. Dr. Apostolos Papageorgiou for 114(3), 033532.
his clarifying remarks on the authors’ questions about the synthetic Chey, M. H., Chase, J. G., Mander, J. B., and Carr, A. J. (2010). “Semi-
time histories algorithm. active tuned mass damper building systems: Application.” Earthquake
Eng. Struct. Dyn., 39(1), 69–89.
Chopra, A. K. (2007). Dynamics of structures: Theory and appli-
Notation cations to earthquake engineering, Prentice Hall, Upper Saddle River,
NJ.
The following symbols are used in this paper: Colombi, A., Colquitt, D., Roux, P., Guenneau, S., and Craster, R. (2016a).
BI = base isolation; “A seismic metamaterial: The resonant metawedge.” Sci. Rep., 6,
CRN = common random numbers; 27717.
Colombi, A., Roux, P., Guenneau, S., Gueguen, P., and Craster, R. (2016b).
E = stochastic excitation model;
“Forests as a natural seismic metamaterial: Rayleigh wave bandgaps
fex = excitation frequency; induced by local resonances.” Sci. Rep., 6, 19238.
G = stochastic system model; Dertimanis, V. K., Antoniadis, I. A., and Chatzi, E. N. (2016). “Feasibility
g = limit state function; analysis on the attenuation of strong ground motions using finite peri-
g~ = augmented limit state function; odic lattices of mass-in-mass barriers.” J. Eng. Mech., 10.1061/(ASCE)
ĝk = gradient estimator; EM.1943-7889.0001120, 1–10.
Finocchio, G., et al. (2014). “Seismic metamaterials based on isochronous
I F = failure indicator function;
mechanical oscillators.” Appl. Phys. Lett., 104(19), 191903.
j = stochastic performance evaluation model; Harvey, P. S., Jr., and Kelly, K. C. (2016). “A review of rolling-type seismic
k̂ = equivalent chain stiffness; isolation: Historical development and future directions.” Eng. Struct.,
kint = unit cell internal mass; 125, 521–531.

© ASCE 04017181-16 J. Eng. Mech.

J. Eng. Mech., 2018, 144(3): 04017181

 Hatzigeorgiou, G. D., and Kanapitsas, G. (2013). “Evaluation of fundamen- Papadimitriou, C., Beck, J. L., and Katafygiotis, L. S. (2001). “Updating
tal period of low-rise and mid-rise reinforced concrete buildings.” robust reliability using structural test data.” Probab. Eng. Mech., 16(2),
Earthquake Eng. Struct. Dyn., 42(11), 1599–1616. 103–113.
Huang, J., and Shi, Z. (2013). “Application of periodic theory to rows of Ruszczyński, A., and Shapiro, A. (2003). “Optimality and duality in sto-
piles for horizontal vibration attenuation.” Int. J. Geomech., 10.1061 chastic programming.” Handb. Oper. Res. Manage. Sci., 10, 65–139.
/(ASCE)GM.1943-5622.0000193, 132–142. Skinner, R. I., Beck, J. L., and Bycroft, G. N. (1975). “A practical system
Jaynes, E. T. (2003). Probability theory: The logic of science, Cambridge for isolating structures from earthquake attack.” Earthquake Eng.
University Press, Cambridge, U.K. Struct. Dyn., 3(3), 297–309.
Kaynia, A. M., Biggs, J. M., and Veneziano, D. (1981). “Seismic effective- Spall, J. C. (2003). Introduction to stochastic search and optimization,
ness of tuned mass dampers.” J. Struct. Eng., 107(8), 1465–1484. Wiley, Hoboken, NJ.
Kleinman, N. L., Spall, J. C., and Naiman, D. Q. (1999). “Simulation-based Taflanidis, A. (2007). “Stochastic system design and applications to sto-
Downloaded from ascelibrary.org by MICHIGAN TECHNOLOGICAL UNIVERSITY on 09/07/19. Copyright ASCE. For personal use only; all rights reserved.

optimization with stochastic approximation using common.” Manage. chastically robust structural control.” Ph.D. thesis, California Institute
Sci., 45(11), 1570–1578. of Technology, Pasadena, CA.
Taflanidis, A., and Beck, J. L. (2008a). “An efficient framework for optimal
Krödel, S., Thomé, N., and Daraio, C. (2015). “Wide band-gap seismic
robust stochastic system design using stochastic simulation.” Comput.
metastructures.” Extreme Mech. Lett., 4, 111–117.
Methods Appl. Mech. Eng., 198(1), 88–101.
Makris, N., and Gazetas, G. (1992). “Dynamic pile-soil-pile interaction. II:
Taflanidis, A., and Beck, J. L. (2008b). “Stochastic subset optimization for
Lateral and seismic response.” Earthquake Eng. Struct. Dyn., 21(2),
optimal reliability problems.” Probab. Eng. Mech., 23(2–3), 324–338.
145–162.
Taflanidis, A., and Beck, J. L. (2009). “Stochastic subset optimization
Maldovan, M. (2013). “Sound and heat revolutions in phononics.” Nature, for reliability optimization and sensitivity analysis in system design.”
503(7475), 209–217. Comput. Struct., 87(5–6), 318–331.
Mavroeidis, G. P., and Papageorgiou, A. S. (2003). “A mathematical Taflanidis, A., Scruggs, J. T., and Beck, J. L. (2008). “Probabilistically
representation of near-fault ground motions.” Bull. Seismol. Soc. Am., robust nonlinear design of control systems for base-isolated structures.”
93(3), 1099–1131. Struct. Control Health Monit., 15(5), 697–719.
Miniaci, M., Krushynska, A., Bosia, F., and Pugno, N. (2016). “Large scale Wagner, P.-R., Dertimanis, V. K., Chatzi, E. N., and Antoniadis, I. A.
mechanical metamaterials as seismic shields.” New J. Phys., 18(8), (2016a). “Design of metamaterials for seismic isolation.” Conf. Proc.
083041. Society for Experimental Mechanics Series, Vol. 2, Orlando, FL,
Palermo, A., Krödel, S., Marzani, A., and Daraio, C. (2016). “Engineered Springer, New York, 275–287.
metabarrier as shield from seismic surface waves.” Sci. Rep., 6, 39356. Wagner, P.-R., Dertimanis, V. K., Chatzi, E. N., and Antoniadis, I. A.
Papadimitriou, C., Beck, J. L., and Katafygiotis, L. S. (1997). “Asymp- (2016b). “On the feasibility of structural metamaterials for seismic-
totic expansions for reliabilities and moments of uncertain dynamic sys- induced vibration mitigation.” Int. J. Earthquake Impact Eng., 1(1–2),
tems.” J. Eng. Mech., 10.1061/(ASCE)0733-9399(1997)123:12(1219), 20–56.
1219–1229. Zhu, J., et al. (2013). “Acoustic rainbow trapping.” Sci. Rep., 3(1728), 1–6.

© ASCE 04017181-17 J. Eng. Mech.

J. Eng. Mech., 2018, 144(3): 04017181