Soil Dynamics and Earthquake Engineering 85 (2016) 202–216

Contents lists available at ScienceDirect

Soil Dynamics and Earthquake Engineering

journal homepage: www.elsevier.com/locate/soildyn

Seismic response of underground reservoir structures

in sand: Evaluation of Class-C and C1 numerical simulations using
centrifuge experiments
Y.H. Deng a, S. Dashti b,n, A. Hushmand b, C. Davis c, B. Hushmand d
a
Chang’an University, Department of Geological Engineering, Xi’an, China
b
University of Colorado Boulder, Civil, Environmental and Architectural Engineering, Boulder, CO, United States
c
Los Angeles Department of Water and Power, Los Angeles, CA, United States
d
Hushmand Associates, Inc., Irvine, CA, United States

art ic l e i nf o a b s t r a c t

Article history: Centrifuge experiments were conducted to investigate the seismic response of stiff-unyielding buried
Received 10 December 2015 reservoir structures with varying stiffness in medium-dense, dry sand. The results of these tests were
Received in revised form used to evaluate the predictive capabilities of Class-C and C1, nonlinear, finite element analyses of the
29 March 2016
seismic response of these relatively stiff buried structures. All simulations were performed in two di-
Accepted 6 April 2016
Available online 18 April 2016
mensions using the pressure-dependent, multi-yield-surface, plasticity-based soil constitutive model
(PDMY02) implemented in OpenSees. For Class-C simulations, model parameters were calibrated based
Keywords: on the available cyclic simple shear tests on the test soil. For Class-C1 simulations, the same soil model
Centrifuge was used along with user-defined modulus reduction curves that were corrected for soil's implied shear
Seismic soil-structure interaction
strength. The use of shear modulus reduction curves, which modeled a softer soil response compared to
Nonlinear finite element analyses
PDMY02, generally improved the prediction of site response in the far-field as well as seismic racking
Buried reservoir structures
deformations, earth pressures, and bending strains on the structures. Experimentally, the dynamic
thrust, racking, and bending strains on or of the model structures were shown to primarily peak near the
strain-dependent fundamental frequency of the site, regardless of the fundamental frequency of the
structure itself. This influence in addition to other important response parameters were captured rea-
sonably well by Class-C1 simulations, with residuals ranging from
0.25 to 0.2.
& 2016 Elsevier Ltd. All rights reserved.

1. Introduction They exhibit some deformation depending on their stiffness, but

they deform less than a vertical structure buried in the ground.
Underground box structures have generally performed well This is because the walls of a reservoir structure are restrained
during previous earthquakes except for a few cases of damage against large deformations at their roof and base [12].
(e.g., Daikai Subway Station during the 1995 Hyogo Ken Nanbu Soil-structure interaction (SSI) near relatively stiff buried
Earthquake; Balboa water treatment plant during the 1971 San structures depends on the stiffness and geometric constraints of
Fernando Earthquake), which show the importance of designing the structure, properties of the backfill soil, soil-structure interface
them safely for seismic loading. The current state of practice for properties, and characteristics of the earthquake motion. Ad-
analyzing the seismic performance of underground reservoir vanced numerical tools can model the stiffness and geometric
structures relies heavily on simplified analytical methods that do constraints of structures, soil nonlinearities during strong dynamic
loading, and the interface conditions between the soil and struc-
not fully capture the kinematic constraints imposed on these
ture, in order to predict the seismic performance of structures.
structures. These procedures are often either based on the as-
Previous numerical simulations of the seismic response of yielding
sumption of a yielding wall, which displaces sufficiently to develop
or unyielding walls typically used linear, equivalent-linear vis-
an active condition (e.g., [19,21,25]), or a rigid-unyielding wall,
coelastic, or elastic perfectly-plastic backfill soil properties (e.g.,
which undergoes no deformation (e.g., [32]). Underground re-
[23,24,30,31,36]) to evaluate the influence of wall flexural rigidity,
servoir structures fall in neither of these two extreme categories. base and translational flexibility, and soil properties on the am-
plitude and distribution of earth pressures and deformations.
n
Corresponding author. Seismic earth pressures on stiff-unyielding structures have not
E-mail address: [email protected] (S. Dashti). been studied adequately using advanced numerical analyses with

http://dx.doi.org/10.1016/j.soildyn.2016.04.003
0267-7261/& 2016 Elsevier Ltd. All rights reserved.
 Y.H. Deng et al. / Soil Dynamics and Earthquake Engineering 85 (2016) 202–216 203

Table 1
Dimensions and properties of three equivalent model structures used in centrifuge in prototype scale [12].

Structure Height and width (m) (outer edge to outer edge) Thickness (cm) Lateral stiffness (kN/m2) Fundamental frequency (Hz)

Base Roof Walls

Baseline 10.4 and 12.2 69 37 56 9.0 4.0

Flexible 50 28 28 1.8 2.0
Stiff 146 112 113 131.4 9.9

Note: Model structures were 17.46 m long (approximately equal to the inside width of the centrifuge container).

nonlinear, pressure-dependent soil properties and realistic interface soil. Class-C1 simulations were performed with the same soil
conditions. Capturing soil's nonlinear response is particularly im- model as the Class-C but with manually adjusted shear modulus
portant in the analysis of reservoir structures, because they are of- reduction curves, in order to better capture centrifuge experi-
ten designed for strong levels of shaking inducing large strains in mental results. In this paper, we discuss and compare the
the backfill soil. Further, many of the previous numerical methods predictive ability of Class-C and C1, nonlinear finite element
with realistic conditions were not sufficiently validated against case simulations in capturing seismic SSI, accelerations, deformations,
histories or experimental studies for stiff-unyielding structures. and lateral earth pressures on underground reservoir structures.
Most of the previous experimental studies have focused on the
seismic response of yielding retaining walls [1,5,18,20,22,27,28].
Recent dynamic centrifuge tests have also been performed on re- 2. Overview of centrifugre experiments
latively flexible, rectangular tunnels in cohesionless soils (e.g.,
[6,,30]). These tunnels, however, were more flexible and buried A series of three centrifuge experiments were conducted at the
deeper than those representative of reservoir structures near the University of Colorado Boulder, all with similar soil conditions,
surface. instrumentation, and structure outer dimensions but different
A series of three centrifuge experiments were recently con- thickness and stiffness of the model underground structures.
ducted by Hushmand et al. [12] at the University of Colorado These three tests, referred to as T-Flexible, T-Baseline (or T-BL),
Boulder to evaluate the seismic performance of relatively stiff and T-Stiff, were named according to the relative stiffness of the
underground structures buried in an 18.6 m-thick layer of dry, underground structure, as detailed by Hushmand et al. [12]. All the
Nevada Sand at a relative density (Dr) of approximately 60%. Three three models were spun to 60 g of centrifugal acceleration, and a
different model box structures were designed to represent sim- sequence of five earthquake motions were applied to the model
plified prototype reinforced concrete buried reservoir structures of specimen in flight in the same order. All the dimensions presented
varying stiffness characterizing those evaluated by the Los Angeles in this paper are in prototype scale, unless stated otherwise.
Department of Water and Power (LADWP). The structures had
10.4 m high walls that were restrained against rotational move- 2.1. Properties of underground structures
ment at the top and bottom by their roof and floor. Additionally,
the reservoir's foundation could rock or slide on the soil. Hush- Three simplified, equivalent, scaled box structures were de-
mand et al. [12] showed experimentally that none of the com- signed in centrifuge to simulate the mass, lateral stiffness, and
monly used procedures (e.g., [2,19,21,25,32]) could adequately natural frequency of complex prototype, buried, reinforced con-
capture the loading and deformations experienced by reservoir crete, water reservoir structures. The structures were designed
structures for the range of stiffness and ground motions con- with uniform 1018 Carbon Steel (density ¼7870 kg/m3; Young's
sidered in their design. These experiments enabled a compre- Modulus ¼200 GPa; Poisson's ratio¼0.29). As summarized in
hensive and fundamental evaluation of the influence of structure Table 1, the outer dimensions of the three structures were kept the
stiffness and ground motion characteristics on seismic SSI as well same, while their thicknesses were varied to change their stiffness.
as lateral earth pressures, racking deformations, and bending
strains experienced by stiff-unyielding buried structures. 2.2. Soil properties, model preparation, and instrumentation
The centrifuge experiments performed by Hushmand et al. [12]
are used in this paper to evaluate the ability of Class-C and C1, Fig. 1 shows the model configuration and instrumentation in
nonlinear finite element analyses of the soil-structure system in the three experiments. Nevada sand (Gs ¼ 2.65; emin ¼0.56;
simulating the response of stiff-unyielding structures buried in emax ¼0.84; D50 ¼0.13 mm; Cu ¼1.67) was dry pluviated into a
medium-dense, dry sand. Even though centrifuge experiments do flexible-shear-beam (FSB) centrifuge container to achieve a uni-
not represent the complexities of real field conditions, they enable form soil layer with a relative density of Dr E60% or a dry unit
validation of numerical tools that may later be used in modeling weight of 15.6 kN/m3. The transfer functions of surface to base
more complex conditions. The prediction classification used in this accelerations in the far-field under centrifuge ambient vibrations
paper is based on that described by Lambe [16]. A Class-C pre- indicated a small-strain, soil fundamental frequency (fso) ranging
diction is one made after the experiment without knowing the from about 2.1 to 2.4 Hz (corresponding to an average, small-
results at the time of the prediction. A Class-C1 prediction is one strain, shear wave velocity, V̄s, ranging from about 156 to 179 m/s),
made after the event, while knowing the results of the experiment while the transfer functions obtained from earthquake motions
at the time of prediction. indicated strain-dependent, effective site fundamental frequencies
All simulations of the centrifuge tests presented in this paper (fso′) ranging from about 1.0 to 1.7 Hz (average, effective V̄s’ ran-
were performed using the pressure-dependent, multi-yield-sur- ging from 74 to 126 m/s). The estimated small strain Vs ¯ of the far-
face, plasticity-based soil constitutive model (PDMY02) im- field soil prior to shaking agreed with but was slightly smaller than
plemented in OpenSees by Elgamal et al. [9] and Yang et al. those estimated using the empirical procedures recommended
[34,35]. The PDMY02 soil constitutive model parameters for Class- by Seed and Idriss [26] and Bardet et al. [4] for sand.
C simulations were calibrated based on the available cyclic simple Teflon sheets were used on the container sides and ends of the
shear tests (e.g., detailed by Karimi and Dashti [14,15]) on the test structures to reduce friction at the structure-container interface.
204 Y.H. Deng et al. / Soil Dynamics and Earthquake Engineering 85 (2016) 202–216

Fig. 1. Schematic and layout of centrifuge experiments in prototype scale [12].

Nevada sand was glued to the base of the structures to provide a terms of intensity, frequency content, and duration) on the per-
more realistic interface friction between the structure and the soil. formance of the buried structures and their interaction with the
As shown in Fig. 1, data was collected with accelerometers, tactile surrounding soil. Fig. 2 shows the acceleration response spectra
pressure sensors, strain gauges, and LVDTs. The accelerometer ar- (5% damped) and Arias Intensity time histories of the base motions
ray A1-4 representing far-field conditions (approximating free- achieved (or measured) in T-BL, and Hushmand et al. [12] provided
field) was placed 11.1 m from the wall of the structure and 3.7 m more details on their properties. The achieved base motions varied
from the inner boundary of the flexible container. Even though use slightly during different experiments because of the variation in
of an FSB container is expected to reduce boundary effects, the far- model weights and natural frequencies. Therefore, it was im-
field location in these experiments was likely still influenced to portant to use the actual recording of the base motion during the
some degree by boundary effects as well as soil-structure inter- numerical simulation of the corresponding test.
action. Therefore, it was important to mimic similar conditions
numerically for a proper comparison, as opposed to simulating a
true free-field condition. The tactile sensors were first thoroughly 3. Numerical simulations
de-aired, conditioned, equilibrated, and calibrated statically ac-
cording to the procedure recommended by Tessari et al. [29]. Then, The centrifuge tests described in the previous section were
they were dynamically calibrated using the procedure described numerically simulated using the pressure-dependent, multi-yield-
by Gillis et al. [10]. surface, plasticity-based soil model (PDMY02) implemented in
OpenSees by Elgamal et al. [9] and Yang et al. [34]. Nonlinear soil
2.3. Ground motions constitutive models have many parameters to calibrate, often
making them impractical in engineering design. However, in the
Scaled, horizontal components of the following recordings seismic design of underground reservoir structures, strong ground
were applied to the base of the model in each experiment: Sylmar motions are regularly selected due to the sensitivity of these
Converter Station during the 1994 Northridge Earthquake (NSC52), structures. Under such strong levels of shaking and shear strains,
the LGPC Station during the 1989 Loma Prieta Earthquake equivalent-linear viscoelastic or elastic perfectly-plastic soil con-
(LGP000), and the Istanbul Station during the 1999 Izmit Earth- stitutive models become increasingly limiting. Hence, the PDMY02
quake in Turkey (IST180). These motions were chosen to evaluate model, despite its many parameters, was used in this study to
the influence of different ground motion characteristics (i.e., in model the nonlinear behavior of the retained soil. Hashash et al.

Fig. 2. Container base motions as recorded in T-BL and input into the numerical simulations: (a) 5%-damped spectral accelerations; and (b) Arias Intensity time histories.
 Y.H. Deng et al. / Soil Dynamics and Earthquake Engineering 85 (2016) 202–216 205

Fig. 3. (a) The small-strain Vs profile used in the numerical simulations of Nevada sand compared with two empirical procedures; (b) the maximum allowed and selected
element size with depth.

[11] showed the promise of this constitutive model in analyzing

site response in medium-dense, dry sand as compared to cen-
trifuge measurements, when combined with user-defined mod-
ulus reduction curves expected for the test soil. The numerical
analyses presented in this paper were conducted in prototype
scale and in two dimensions (2-D) assuming plane strain
conditions.
Soil and underground structures were all modeled using
SSPquad elements [17]. A linear-elastic material was adopted to
simulate the response of the underground structures, with ap-
propriate properties (density, Young's modulus, and Poisson's ra-
tio). Node-to-segment, frictional contact, zero-length interface Fig. 4. The mesh configuration used in the numerical simulation of centrifuge
experiments.
elements of type zeroLengthContactNTS2D [33] were used be-
tween soil and structure elements to allow for relative movements
representative model (T-BL). The displacement degrees of freedom
between the soil and a buried structure. This interface element
were tied for each pair of nodes on side boundaries at the same
follows the classical Coulomb friction law, and the interface fric-
elevation, to represent the physical conditions imposed by the
tion was defined as tan(ϕinterface) ¼0.7 tan(ϕsoil).
centrifuge container.
The element size was selected to allow shear wave propagation
The PDMY02 soil constitutive model parameters for Class-C
in the frequency range of interest. Fig. 3a shows the small-strain Vs
simulations were calibrated based on the available cyclic simple
profile in soil predicted from two empirical equations compared to shear (CSS) tests on Nevada sand with different values of confining
the default estimates by the PDMY02 model. A maximum fre- stress, relative density, and cyclic stress ratio performed by Ar-
quency (fmax) of 15 Hz was conservatively assumed during dy- ulmoli et al. [3] and Kammerer et al. [13]. Karimi and Dashti
namic loading, which was beyond the capacity of the shaking table [14,15] provided a comparison of the numerically simulated (using
under increased gravity (typically around 400 Hz in model scale, the PDMY02 model) and experimentally measured soil response at
which translates to 6.7 Hz in prototype scale). The maximum al- an element level in CSS tests, which are not repeated here for
lowable element size at each depth was then estimated as: brevity. Table 2 summarizes the calibrated soil model parameters
hallowable ¼minimum wavelength/8¼ λmin/8 ¼(Vs/fmax)/8. The used in this study.
adopted mesh size was smaller than the maximum allowable size, The same motions recorded in each centrifuge test were ap-
as shown in Fig. 3b. Fig. 4 shows the mesh configuration of a plied to the base of the corresponding numerical model in the
206 Y.H. Deng et al. / Soil Dynamics and Earthquake Engineering 85 (2016) 202–216

Table 2
Summary of PDMY model parameters for Nevada sand [14,15].

Parameter Value Unit Description

Dr 60 61 64 69 74 % Relative density
e 0.67 0.67 0.66 0.65 0.64 – Void ratio
ρ 1.59 1.59 1.60 1.61 1.62 ton/m3 Saturated unit weight
pr′ 101 101 101 101 101 kPa Reference effective confining pressure
Gmax,1,oct 68.95 70.25 73.42 78.23 83.86 MPa Octahedral low-strain shear modulus
γmax,r 0.1 0.1 0.1 0.1 0.1 – Maximum octahedral shear strain
Br 184.09 187.57 196.06 208.91 223.95 MPa Bulk modulus
d 0.5 0.5 0.5 0.5 0.5 – Pressure dependency coefficient
φ TXC 34.3 34.3 34.8 36.2 37.1 deg. Triaxial friction angle used by model
φPT 26.3 26.3 26.4 26.0 26.1 deg. Phase transformation angle
c1 0.040 0.042 0.036 0.020 0.019 – Control the shear-induced volumetric change, contraction tendency based on the dilation history, and
c2 2.85 2.72 2.30 1.50 1.49 – overburden stress effect, respectively
c3 0.21 0.21 0.19 0.15 0.15 –
d1 0.07 0.07 0.09 0.15 0.18 – Reflect dilation tendency, stress history, and overburden stress, respectively
d2 3.0 3.0 3.0 3.0 3.0 –
d3 0.0 0.0 0.0 0.0 0.0 –
NYS 20 20 20 20 20 – Number of yield surfaces generated by model
liq1 1.0 1.0 1.0 1.0 1.0 – Account for permanent shear strain (slip strain or cyclic mobility) in sloping ground
liq2 0.0 0.0 0.0 0.0 0.0 –

same sequence. After the application of each motion, the soil

densified, as confirmed by LVDT recordings at different locations.
The far-field settlement measurements in each test were used to
estimate, in a simplistic manner, the change in soil relative density
(Dr) under 1-D conditions. This change in soil Dr was taken into
account in estimating soil model parameters for the subsequent
motions in the numerical simulations, as shown in Table 2.
No modulus reduction curve was used other than that pre-
dicted by default by the PDMY02 model in OpenSees for Class-C
predictions. A small-strain, Rayleigh damping value of 2% was
implemented at the fundamental frequency of the site. Based on
the comparison of Class-C numerical results with experimental
measurements, particularly the response in the far-field, Class-C1
simulations were performed with the same soil constitutive model
parameters as Class-C but this time with user-defined, shear
modulus reduction curves (G/Gmax versus γ) that were corrected
for soil's implied shear strength (similar to the approach taken by Fig. 5. Normalized modulus reduction curves used in the Class-C and C1 numerical
[11]), in order to better capture the response measured in cen- simulations. Class-C simulations used the default PDMY02 curve in OpenSees.
trifuge. Table 3 summarizes the parameters used to obtain the Class-C1 simulations used the mean and mean þ1 standard deviation (upper)
curves proposed by Darendeli [7].
mean and one standard deviation above mean (upper) G/Gmax
curves proposed by Darendeli [7], and Fig. 5 compares these
4. Comparison of experimental and numerical results
curves with the default response of the PDMY02 model. The rea-
sons behind using these Darendeli's curves in Class-C1 simulations
4.1. Accelerations in the far-field and on the structure
and the comparison of results with centrifuge recordings will be
discussed in the following sections. Evaluation of numerical predictions started with site perfor-
mance (in terms of acceleration and settlement) in the far-field, a
location away from the buried structures to minimize the com-
Table 3
plexities introduced by soil-structure interaction. The goal was to
Parameters used to ob- evaluate the ability of nonlinear, time-domain, site-response
tain [7] modulus reduc- analyses in capturing the 1-D propagation of horizontal shear
tion curves for Nevada
sand.
waves through uniform dry sand, the strain-compatible effective
fundamental frequency (fso′) of the site, and shaking-induced
Parameters Values permanent settlements in the far-field. The transfer function (TF)
of accelerations at the soil surface in the far-field to container base
K0 0.44
PI (%) 0 were first compared among the experiments and simulations, as
N (cycles) 10 shown in Fig. 6 for a representative experiment (T-BL) and three
f (Hz) 1
representative motions. Fig. 7 compares numerically predicted
OCR 1
accelerations with those measured in centrifuge along the far-field
 Y.H. Deng et al. / Soil Dynamics and Earthquake Engineering 85 (2016) 202–216 207

Fig. 6. Experimental and numerical transfer functions (TFs) of surface to base acceleration in the far-field in T-BL during three representative motions.

Fig. 7. Experimental and numerical 5%-damped spectral accelerations and Arias Intensity-time histories in the far-field during the Northridge_L and Northridge_H motions
in T-BL.
208 Y.H. Deng et al. / Soil Dynamics and Earthquake Engineering 85 (2016) 202–216

Fig. 8. Experimental and numerical settlements in the far-field during three representative motions in T-BL.

soil column in terms of 5%-damped acceleration response spectra The fso′ of the far-field soil corresponding to the peak value of
and Arias Intensity time histories. Fig. 8 compares numerically surface to base TF's was experimentally observed to range from
predicted soil surface settlements in the far-field with those approximately 1–1.7 Hz during the motions employed in this
measured during different motions. study. Although far-field surface settlements were predicted well

Fig. 9. Experimental and numerical spectral ratios of structure to far-field acceleration (5%-damped) for different structures and three representative ground motions.
 Y.H. Deng et al. / Soil Dynamics and Earthquake Engineering 85 (2016) 202–216 209

by Class-C simulations (Fig. 8), the amplification of accelerations at Experimentally, increasing the confinement and structure stiffness
higher frequencies and hence, the fso′ of the site were over- were observed to reduce the amplification of structure to far-field
estimated during all motions (Fig. 6). This also led to the over- spectral ratios. This pattern was captured numerically. However,
estimation of acceleration response spectra at higher frequencies near the roof, all numerical simulations slightly underestimated
as well as Arias Intensities, particularly near the soil surface in the the amplification of accelerations on the structure. The connection
far-field (Fig. 7). In Class-C predictions, the shear modulus reduc- between soil and structure elements were therefore slightly too
tion of soil was estimated automatically by the PDMY02 model stiff, restraining the independent horizontal movement of the
(Fig. 5). From the far-field comparisons, the soil behavior predicted structure near its roof compared to what was observed experi-
by the PDMY02 model was judged to be too stiff. mentally, as shown in Fig. 9. Further, Class-C predictions with the
The effective fundamental frequency of the site (fso′) was default, stiffer, PDMY02 modulus reduction response often over-
known to have a significant impact on the response of relatively estimated the frequency at which the peak spectral ratios
stiff underground reservoir structures, as demonstrated experi- occurred.
mentally by Hushmand et al. [12]. Therefore, in Class-C1 predic-
tions, slightly reduced G/Gmax curves (the mean and mean þ 1 4.2. Racking deformations in the far-field and on the structure and
standard deviation curves proposed by Darendeli [7]) were numerical settlements in the far-field
manually defined in OpenSees to better match the acceleration
response in the far-field. These user-defined curves were corrected Seismically-induced racking displacement, defined as the re-
for the implied shear strength of soil (e.g., [11]). The predicted lative lateral displacement of the structure's roof with respect to
accelerations and fso′ generally improved near the surface, parti- its base, is an important measure of seismic performance for un-
cularly with the use of the upper (i.e. mean þ1 standard deviation) derground box structures. Experimentally, racking deformations
Darendeli G/Gmax curves, as shown in Figs. 6 and 7. Using a softer were obtained both in the far-field and on the structure from
soil response in OpenSees, however, increased the error in the double integrating and subtracting accelerometer recordings at
predicted far-field settlements, as shown in Fig. 8. the elevations corresponding to the structure's roof and base [12].
The change in accelerations due to the presence of the under- A band-pass, 5th order, acausal, Butterworth filter with corner
ground structure was monitored experimentally and numerically, frequencies of 0.2 and 15 Hz was applied to acceleration records,
to evaluate the influence of soil-structure interaction (SSI) on ac- followed by double integration and baseline correction to obtain
celerations near buried reservoir structures. Fig. 9 compares the displacement from accelerometers A12 and A14 on the structure
numerically predicted and experimentally measured spectral ra- and A2 and A4 in the far-field. Since the structures were designed
tios of acceleration at the roof of the structure to that in the far- to remain elastic, no permanent racking was expected, and
field soil surface during three representative motions. therefore using accelerometers was judged to be adequate.

Fig. 10. Experimental recordings and numerical predictions of racking in the far-field during three representative motions in T-BL: (a) time histories zoomed to 5–20 s for
clarity; (b) Fourier Amplitude Spectra.
210 Y.H. Deng et al. / Soil Dynamics and Earthquake Engineering 85 (2016) 202–216

Fig. 11. Experimental recordings and numerical predictions of racking on three different structures during three representative motions in T-BL, T-Flexible, and T-Stiff:
(a) time histories zoomed to 5–20 s for clarity; (b) representative Fourier Amplitude Spectra in T-Stiff.

Numerically, racking displacements were obtained directly from Most simulations, particularly Class-C1 predictions, could cap-
the predicted lateral displacements. ture this effect. Class-C simulations with the default PDMY02
Figs. 10 and 11 compare numerically predicted and experi- model G/Gmax curve often slightly overestimated racking at
mentally recorded seismic racking displacements in the far-field higher frequencies and underestimated racking at lower fre-
and on three structures, in terms of time histories and Fourier quencies both in the far-field and on the structures. As expected
Amplitude Spectra (FAS). Experimentally, racking deformations from the results in terms of far-field accelerations, Class-C1
both in the far-field and on the structures appeared to have a predictions with the upper Darendeli G/Gmax curve generally
significant content near 1 Hz, which coincided with the effective predicted racking better than other simulations, particularly on
fundamental frequency of the site (fso′) during these motions. the structure. Use of the mean Darendeli G/Gmax curve often led
 Y.H. Deng et al. / Soil Dynamics and Earthquake Engineering 85 (2016) 202–216 211

to an overestimation of structural racking at lower frequencies approximately triangular (increasing with depth) to a higher order
because of a softer retained soil. polynomial as the flexural rigidity of the wall increased.
Dynamic thrust was obtained experimentally on the structure
4.3. Lateral earth pressures on the structure wall by numerically integrating the dynamic earth pressures along
the height of the wall at each instance of time. The dynamic thrust
Evaluating the predictive capabilities of nonlinear numerical si- time histories obtained from tactile sensors were subject to a
mulations in terms of seismic lateral earth pressures was one of the band-pass, 5th order, acausal, Butterworth filter with corner fre-
main objectives of this study, due to their critical importance in de- quencies of 0.1 and 15 Hz, to remove low and high frequency noise
sign. To obtain lateral earth pressures experimentally, tactile pressure that was at times present in the recordings. As a result, any per-
sensors were employed. These sensors were conditioned, equili- manent change in thrust could not be captured after filtering.
brated, and calibrated both statically and dynamically, as detailed by From these time histories, however, the time corresponding to
Gillis et al. [10]. It was shown experimentally that dynamic earth peak thrust could be determined during each ground motion [12].
pressures acting on stiff-unyielding structures can be of engineering Dynamic thrust was similarly obtained numerically by integrating
significance and need to be considered in design [12]. Further, the dynamic earth pressures along the height of the wall. For a con-
distribution of dynamic earth pressures appeared to change from sistent comparison with experimental measurements, a similar

Fig. 12. Static (before and after shaking), total (static þ dynamic), and dynamic lateral earth pressure profiles at the time of maximum dynamic thrust on different structures
during the Northridge-L motion.
212 Y.H. Deng et al. / Soil Dynamics and Earthquake Engineering 85 (2016) 202–216

filter was applied to numerically predicted dynamic thrust time (Ko) and active (Ka) conditions. All simulations predicted static,
histories to remove any permanent change in thrust. pre-shake earth pressures well, which often fell between at-rest
Fig. 12 compares numerically predicted static (before and after and active conditions. The experimentally observed permanent
the motion prior to filtering), maximum total (static þdynamic), increase in static earth pressures due to earthquake loading was,
and maximum dynamic lateral earth pressure profiles with those however, largely underestimated numerically. This is because the
measured in different experiments during the Northridge_L mo- numerical model does not update soil properties and changes in
tion. The maximum total and dynamic earth pressures are shown soil density and contact of grains with the wall during a time
at the time corresponding to peak dynamic thrust in each ex- domain analysis.
periment or simulation. The results are also compared with the Total and dynamic earth pressures at the time of peak dynamic
theoretically expected static lateral earth pressures under at-rest thrust were generally overestimated by the stiffer soil models

Fig. 13. Experimentally measured and numerically predicted dynamic thrust on different structures during three representative motions: (a) time histories zoomed to 5–20 s
for clarify; (b) representative Fourier Amplitude Spectra in T-Stiff.
 Y.H. Deng et al. / Soil Dynamics and Earthquake Engineering 85 (2016) 202–216 213

(default PDMY02 and upper Darendeli) and in some cases slightly Friction was minimized on the face of the tactile sensors by
underestimated by the softer soil model (mean Darendeli), parti- using Teflon sheets (as detailed by Gillis et al. [10]) to reduce shear
cularly at shallow depths. The distribution of dynamic earth stresses on the sensors and improve their reliability in measuring
pressures was better captured by Class-C1 simulations with the normal pressure. However, this interface condition is different
mean Darendeli G/Gmax curve implemented. All simulations, from what is typically expected on the walls of buried, reinforced
however, failed to predict a parabolic distribution of dynamic concrete, reservoir structures and what was assumed in the nu-
earth pressures on the stiff structure and instead predicted a merical simulations. This difference may have been partly re-
roughly triangular distribution increasing with depth. This was sponsible for the observed differences in the predicted and mea-
contrary to previous experimental observations. sured lateral earth pressures in all Class-C and C1 simulations. In

Fig. 14. Experimentally measured and numerically predicted bending strains at the corner of different structures during three representative motions: (a) time histories
zoomed to 5–20 s; (b) representative Fourier Amplitude Spectra in T-Stiff.
214 Y.H. Deng et al. / Soil Dynamics and Earthquake Engineering 85 (2016) 202–216

addition, nonlinear interface properties and complex densification pressures, the peak values of dynamic strain were often measured
patterns of granular soils around the walls during shaking are near the effective fundamental frequency of the site (fso′), re-
difficult to replicate numerically with a continuum model. gardless of the structure”s stiffness. Class-C predictions (default
Numerically predicted dynamic thrust was compared with PDMY02) often overestimated bending strains at higher fre-
those measured in centrifuge for different structures and motions quencies and underestimated strains at lower frequencies. But the
in Fig. 13 in terms of time histories and Fourier Amplitude Spectra, Class-C1 simulations (both mean and upper Darendeli) better
respectively. Experimentally, dynamic thrust on all structures captured dynamic strains both in terms of amplitude and fre-
(regardless of its fundamental frequency) always showed sig- quency content. Similar trends were observed at other locations.
nificant content near the effective fundamental frequency of the
site (fso′), which was approximately 1–1.7 Hz during the motions 4.5. Summary of numerical and experimental comparisons
considered, as shown in Fig. 13b. The stiffer soil model (default
PDMY02) often overestimated the peak dynamic thrust and failed The accuracy of Class-C and C1 numerical predictions was
to capture the large content near site fso′. Generally, the predic-
evaluated in terms of residuals for different response parameters
tions improved in Class-C1 simulations, particularly with the up-
of interest:
per Darendeli curve, although all models tended to underestimate
dynamic thrust near the dominant frequencies of 1–1.7 Hz. ⎛ X experimental ⎞
Residual X = log ⎜ ⎟
⎝ Xnumerical ⎠ (1)
4.4. Dynamic bending strains on the structure
where X refers to a given quantity obtained numerically or ex-
The amplitude and distribution of dynamic bending strains perimentally. The variance in the predicted response was also
were another important measure of the demand imposed on each evaluated among different numerical simulations for different
structure and its performance during earthquake loading. Experi- motions considered. Fig. 15 summarizes the range of residuals and
mentally, dynamic bending strains were obtained from eight strain variances in different response parameters of interest: PGA profile
gauges installed on either side of the box structure walls. The in the far-field and along the structure, acceleration response
largest strains, as expected, were measured on the flexible struc- spectra and Arias Intensity time histories in the far-field surface,
ture and the smallest on the stiff structure. The corner strain racking of the far-field soil and structure, magnitude and location
gauges (SG8) measured the largest dynamic strains, which are of dynamic thrust time histories, and dynamic bending strain time
compared with numerical predictions for different structures and histories along the structure walls for each type of simulation. In
motions in Fig. 14 in terms of time histories and Fourier Amplitude general, Class-C1 predictions with the upper G/Gmax Darendeli
Spectra, respectively. Similar to seismic racking and earth curve provided improved predictions of most response parameters

Fig. 15. Range of residuals and variances in the prediction of different response parameters using different simulations.
 Y.H. Deng et al. / Soil Dynamics and Earthquake Engineering 85 (2016) 202–216 215

of interest with a reasonable range of variance. Hence, even model with many parameters may not always be a practical op-
though this model was not ideal for all response parameters, tion. Hence, a comparison of its performance with simpler soil
overall it captured the seismic demand on and the performance of constitutive models in the future will be of value to practitioners.
stiff-unyielding underground structures with residuals ranging
from about
0.25 to 0.2 and variances of less than approximately
0.3. Acknowledgments

The authors would like to acknowledge the Los Angeles De-

5. Concluding remarks partment of Water and Power (LADWP) for financial support of the
experimental component of this research and Mr. Zana Karimi for
The seismic response of relatively stiff underground reservoir assistance in developing the OpenSees models of the centrifuge
structures that are restrained against rotation at their base and experiments.
roof is currently not adequately understood. Seismic soil-structure
interaction (SSI), structure stiffness, connection to the backfill soil,
and soil nonlinearity need to be considered in their design under References
strong levels of shaking.
A series of Class-C and C1, nonlinear numerical simulations [1] Al Atik L, Sitar N. Seismic earth pressures on cantilever retaining structures. J
were performed, and their results were compared against cen- Geotech Geoenviron Eng 2010;136(10):1324–33.
trifuge measurements in terms of site response in the far-field, [2] Anderson DG, Martin GR, Lam IP, Wang JN. Seismic design and analysis of
retaining walls, buried structures, slopes and embankments.Washington, D.C.:
racking deformation of the soil and structure, and seismic lateral Transportation Research Board, National Cooperative Highway Research Pro-
earth pressures and bending strains along the walls of box struc- gram; 2008.
tures with varying stiffness. Two dimensional (2-D), Class-C, nu- [3] Arulmoli K, Muraleetharan KK, Hossain MM, Fruth LS. VELACS: Verification of
liquefaction analyses by centrifuge studies, laboratory testing program.Irvine,
merical simulations were performed in the finite element program CA: The Earth Technology; 1992.
OpenSees with the PDMY02 model. The model parameters were [4] Bardet JP, Huang Q, Chi SW. Numerical prediction for model no. 1. In: Pro-
calibrated to capture the response of Nevada sand (test soil) in ceedings of the international conference on the verification of numerical
procedures for the analysis of soil liquefaction problems. Balkema, Nether-
cyclic simple shear tests. The reduction of soil shear modulus at lands, vol. 1; 1993. pp. 67–86.
larger strains (G/Gmax versus γ) in these simulations was predicted [5] Bolton MD, Steedman RS. Centrifugal testing of micro-concrete retaining walls
automatically by the nonlinear constitutive model. Class-C1 si- subject to base shaking. In: Proceedings of conference on soil dynamics and
earthquake engineering, southampton. Balkema, Rotterdam. pp. 311–329.
mulations were subsequently performed with the same con- [6] Cilingir U, Madabhushi SG. Effect of depth on the seismic response of square
stitutive model parameters but with the addition of user-defined tunnels. Soils Found 2011;51(3):449–57.
G/Gmax curves proposed by Darendeli [7], which were corrected for [7] Darendeli MB. Development of a new family of normalized modulus reduction
and material damping curves.Austin, TX: University of Texas at Austin; 2001.
soil's implied shear strength. A softer soil response was desired in
[8] Dewoolkar MM, Ko H, Pak RYS. Seismic behavior of cantilever retaining walls
Class-C1 simulations to improve the predicted accelerations and with liquefiable backfills. J Geotech Geoenviron Eng ASCE 2001;127(5):424–
the overall performance of buried structures. Mean and one 35.
standard deviation above mean (upper) G/Gmax curves proposed [9] Elgamal A, Yang Z, Parra E. Computational modeling of cyclic mobility and
post-liquefaction site response. Soil Dyn Earthq Eng 2002;22(4):259–71.
by Darendeli [7] for sand were considered in Class-C1 predictions. [10] Gillis K, Dashti S, Hashash Y. Dynamic calibration of tactile sensors for mea-
The addition of user-defined G/Gmax curves that were softer surement of soil pressures in centrifuge. ASTM Geotech Test J 2015;38:261–74.
than the default PDMY02 model improved the prediction of site [11] Hashash Y, Dashti S, Romero Arduz MI, Ghayoomi M, Musgrove M. Evaluation
of 1-D seismic site response modeling of sand using centrifuge experiments.
response in terms of accelerations and the effective fundamental Soil Dyn Earthq Eng J 2015;78:19–31.
frequency of the site, but increased the error in far-field settlement [12] Hushmand A, Dashti S, Davis C, Hushmand B, Zhang M, McCartney J, Lee Y, Hu
predictions. Class-C1 simulations with the upper Darendeli G/Gmax J. Seismic performance of underground reservoir structures: insight from
centrifuge modeling on the influence of structure stiffness. J Geotech Geoen-
curve generally led to improved predictions of racking displace- viron Eng 2016. http://dx.doi.org/10.1061/GT.1943-5606.0001477 ASCE.
ments both in the far-field and on the structure. Both sets of Class- [13] Kammerer A, Wu J, Pestana J, Riemer M, Seed R. Cyclic simple shear testing of
C1 simulations also improved the amplitude and frequency con- Nevada sand for PEER Center, project 2051999.Berkeley, CA: University of
California; 2000.
tent of dynamic earth pressures and thrust. But, all simulations
[14] Karimi Z, Dashti S. Seismic performance of structures on liquefiable soils:
failed to predict the parabolic distribution of dynamic earth insight from numerical simulations and centrifuge experiments. J Geotech
pressures on the stiff structure. Class-C1 simulations with the Geoenviron Eng ASCE 2016. http://dx.doi.org/10.1061/GT.1943-5606.0001479.
[15] Karimi Z, Dashti S. Numerical and centrifuge modeling of seismic soil-foun-
upper Darendeli G/Gmax curve generally led to better predictions of
dation-structure-interaction on liquefiable ground. J Geotech Geoenviron Eng
dynamic bending strains on the structure, both in terms of am- ASCE 2015;04015061(14):1–14.
plitude and frequency content. The peak response of the structure [16] Lambe TW. Predictions in soil engineering. Geotechnique 1973;23(2):149–202.
in terms of racking, bending strains, and seismic earth pressures [17] McGann CR, Arduino P, Mackenzie-Helnwein P. Stabilized single-point 4-node
quadrilateral element for dynamic analysis of fluid saturated porous media.
was shown experimentally to occur near the effective fundamental Acta Geotech 2012;7(4):297–311.
frequency of the site (fso′) regardless of the stiffness of the struc- [18] Mikola R. Seismic earth pressures on retaining structures and basement walls
ture. Class-C1 predictions with the upper Darendeli G/Gmax curve in cohesionless soils.Berkeley, CA: University of California; 2012.
[19] Mononobe N, Matsuo M. On the determination of earth pressures during
often predicted this important influence from the dynamic prop- earthquakes, Proceedings of the World Engineering Congress, vol. 9; 1929. pp.
erties of the backfill soil. 179–187.
In summary, the PDMY02 soil constitutive model with its de- [20] Nakamura S. Reexamination of Mononobe-Okabe theory of gravity retaining
walls using centrifuge model tests. Soils Found 2006;46(2):135–46.
fault shear modulus reduction led to a soil response that was too [21] Okabe S. General theory of earth pressure. J Jpn Soc Civil Eng 1926;12.
stiff. This led to the overestimation of fso′ as well as the forces and [22] Ortiz LA, Scott RF, Lee J. Dynamic centrifuge testing of a cantilever retaining
deformations experienced by the buried structure at higher fre- wall. Earthq Eng Struct Dyn 1983;11(2):251–68.
[23] Psarropoulos PN, Klonaris G, Gazetas G. Seismic earth pressures on rigid and
quencies. The implementation of user-defined, upper Darendeli
flexible retaining walls. Int J Soil Dyn Earthq Eng 2005;25:795–809.
G/Gmax curve in OpenSees with the PDMY02 soil constitutive [24] Roth WH, Mehrain M. The meaning of seismic earth pressure. Annual SEAOC
model could better capture the key response parameters of in- Convention; 2010.
terest both in the far-field and near buried reservoir structures, [25] Seed HB, Whitman RV. Design of earth retaining structures for dynamic loads.
ASCE specialty conference, lateral stresses in the ground and design of earth
with residuals ranging from about
0.25 to 0.2 and a variance less retaining structures. Cornell University, Ithaca, New York; 1970. pp. 103–147.
than about 0.3. It is acknowledged, however, that the PDMY02 [26] Seed HB, Idriss IM. Soil moduli and damping factors for dynamic response
216 Y.H. Deng et al. / Soil Dynamics and Earthquake Engineering 85 (2016) 202–216

analyses. Berkeley, CA: Earthquake Engineering Research Center, University of [32] Wood JH. Earthquake induced soil pressures on structures.Pasadena, CA: Ca-
California; 1970. p. 40. lifornia Institute of Technology; 1973.
[27] Stadler AT. Dynamic centrifuge testing of cantilever retaining walls.Boulder, [33] Wriggers P, Vu VT, Stein E. Finite-element formulation of large deformation
Colorado: University of Colorado at Boulder; 1996. impact–contact problems with friction. Comput Struct 1990;37:319–31.
[28] Steedman RS, Zeng X. Centrifuge modeling of the effects of earthquakes on [34] Yang Z, Elgamal A, Parra E. Computational model for cyclic mobility and as-
free cantilever walls. Centrifuge’91; 1991. sociated shear deformation. J Geotech Geoenviron Eng 2003;129(12):1119–27.
[29] Tessari A, Abdoun T, Sasanakul I, Wroe E. Boundary corrected calibration of [35] Yang Z, Lu J, Elgamal A. OpenSees soil models and solid-fluid fully coupled
tactile pressure sensors. Phys Model Geotech 2014:331–6. elements: user”s manual.San Diego: Department of Structural Engineering,
[30] Tsinidis G, Pitilakis K, Madabhushi G, Heron C. Dynamic response of flexible University of California; 2008.
square tunnels: centrifuge testing and validation of existing design meth- [36] Zhai E, Davis CA, Yan L, Hu J. Numerical simulations of geotechnical centrifuge
odologies. Geo-technique 2015;65(5):401–17. Modeling of seismic earth pressures on an underground restrained structure.
[31] Veletsos AS, Younan AH. Dynamic modeling and response of soil-wall systems. International Efforts in Lifeline Earthquake Engineering. ASCE; December
J Geotech Eng ASCE 1994;120(12):2155–79. 2013. pp. 369–376.