Journal of Science and Technology in Civil Engineering, NUCE 2021.

15 (1): 30–40

PROBABILISTIC PUSHOVER ANALYSIS

OF REINFORCED CONCRETE FRAME STRUCTURES
USING DROPOUT NEURAL NETWORK
Dang Viet Hunga,∗, Nguyen Truong Thanga , Pham Xuan Data
a
Faculty of Building and Industrial Construction, National University of Civil Engineering,
55 Giai Phong road, Hai Ba Trung district, Hanoi, Vietnam
Article history:
Received 07/07/2020, Revised 04/08/2020, Accepted 02/11/2020

Abstract
When taking into consideration nonlinear phenomena such as material plasticity, plastic hinge, and P-Delta
effect, the pushover analysis can provide more realistic structures’ nonlinear responses. However, this method
is not widely used in practice as it is more complex and requires more expertise than elastic approaches. On
the other hand, the data-driven method emerges as an increasingly appealing alternative since it requires only
input parameters, then directly yields results in conditions that enough training data are provided, as well as an
appropriate machine learning model is devised. Thus, this study develops a probabilistic data-driven approach
using the Multiple Layer Perceptron network coupled with the Dropout mechanism to perform the pushover
analysis of reinforced concrete (RC) frame structures, predicting base shear, lateral displacement, as well as
their relationship between the two formers. Moreover, corresponding confidence intervals of predicted values
are also available owing to the probabilistic nature of the method, thus helping engineers design conservative
solutions.
Keywords: pushover analysis; reinforced concrete; structure; probabilistic analysis; machine learning; dropout
mechanism; OpenSees.
https://doi.org/10.31814/stce.nuce2021-15(1)-03 © 2021 National University of Civil Engineering

1. Introduction

In practice, assessment of lateral displacement of structures is of great importance to ensure its

safety as well as the comfortability of users. Pushover analysis, a nonlinear static technique, is one
of the most well-known methods to perform this task. However, though it can provide more accurate
results, the nonlinear structural analysis is still less favorable compared to the linear counterpart due
to its complexity, high computational cost and expertise required. In recent years, data-driven method
has been considered as a promising alternative by which users only feed inputs to a properly trained
model to obtain reliable results. A number of data-driven models have been developed in the field
of civil engineering [1, 2]. However, when structures become more complex or the nonlinearity is
more pronounced, a deterministic approach seems to be inappropriate because the uncertainty or the
confidence of calculated results needs to be estimated. That is why probabilistic data-driven models
gain increasing attention among the engineering community.

∗
Corresponding author. E-mail address: [email protected] (Hung, D. V.)

30
 Hung, D. V., et al. / Journal of Science and Technology in Civil Engineering

For the probabilistic collapse analysis of structures due to vehicle impact, an artificial neural
network, which is a surrogate model efficient in computational time, was validated by comparison
with experience as presented in [3]. It is also noticed that the surrogate model facilitates various types
of sensitivity analysis, providing more insights into the effects of input parameters on the structure’s
behavior. To assess the damage severity of building after an earthquake, i.e., from minor damage to
collapsed structures, Mangalathu proposed a deep learning-based approach trained on a dataset of
3423 buildings recorded after the 2014 South Napa earthquake [4]. The method achieved a highly
accurate result of 86% and could be integrated into a mobile application, thus enabling a broad range
of users to estimate the vulnerability of the structure immediately following seismic events. Besides,
Feng et al. proposed a probabilistic framework dealing with the progressive collapse of reinforced
concrete (RC) structures due to column removal based on the probability density evolution method
[5]. Such a method allows for quantifying uncertainty in the quantities of interest, i.e., resisting forces,
displacement, etc. with statistic features such as mean variation, standard deviation and probability
density function. Brunesi et al. [6] carried out both deterministic analysis and probabilistic analysis of
progressive collapse in low-rise RC buildings, through which the authors emphasized the contribution
of secondary structural elements such as framing beam in progressive collapse resistance by providing
additional alternative load path. Guo et al. [7] conducted a series of Monte Carlo simulation with the
Latin hypercube sampling method on a 1-h rated steel beam and proved that the beam is able to
resist natural fire event with a probability of failure under 10%. Although Monte Carlo-based method
is able to perform probabilistic analysis but due to its computational expensiveness, another more
practical probabilistic framework for industry application was suggested to be explored. Worrell et
al. [8] investigated a machine learning-based (ML) model for probabilistic assessment of the safety
of nuclear power plants against fire hazard, specifically various ML models are implemented and
compared together, such as K-nearest neighbor, support vector machine, Decision Tree regressor, and
algebraic fire models. It is noticed that these ML models require a large number of training data to
reach a satisfying prediction accuracy. Recently, Fengfu investigated the progressive collapse of steel
frame under fire by leveraging both ML approaches and Monte Carlo simulations [9]. The obtained
results are promising, though it requires sufficient large dataset for the training process which hinder
its applicability in real-word applications.
In an effort to extend the probabilistic data-driven method to estimate the lateral displacement
of building structures, this study develops a probabilistic data-driven approach using the Multiple
Layer Perceptron network coupled with the Dropout mechanism to perform the pushover analysis of
RC frame structures. Once trained, the method can provide quantities of interest such as base shear,
lateral displacement, as well as their relationship between the two formers without requiring building
a numerical model such as Finite Element Method (FEM) with much reduced computational time.
Moreover, the corresponding confidence intervals of predicted values are also available owing to the
probabilistic nature of the method, thus helping engineers design conservative solutions.

2. Dropout Neural Networks

In this section, the architecture of Dropout Neural Networks inspired from the work of Gal and
Ghahramani [10] is investigated and adapted for the progressive collapse problem. The idea behind
the use of Dropout mechanism in this work is that such type of neural network (NN) allows effectively
handling the scarcity of collapsed data by providing probabilistic distribution and its statistic moments
rather than point estimates. Dropout Network consists of three key aspects: (i) Multiple layers of
perceptron includes input layers, one or more hidden layers and output layer as in the conventional
31
data bydata by providing
providing probabilistic
probabilistic distribution
distribution and itsand its statistic
statistic momentsmoments
ratherrather than point
than point
estimates.
estimates. Dropout
Dropout Network
Network consists consists of three
of three key aspects:
key aspects: (i) Multiple
(i) Multiple layerslayers
of of
perceptron
perceptron includes
includes input input
layers,layers,
one orone moreor more
hidden hidden
layerslayers and output
and output layer layer as in the
as in the
conventional
conventional Deep
DeepHung, D.Neural
Neural et al. /Networks
V.,Networks (DNN);
Journal (DNN);
of Science(ii) (ii) Dropout
andDropout
Technology in Civilmechanism
mechanism is applied
is applied
Engineering to theto the
Deephidden
hidden layerslayers
Neural and theandinput
Networks the input
(DNN); (ii)layer
layer with with
various
Dropout various
ratios,ratios,
mechanism iswhose whose
appliedoptimaloptimal
to the valuesvalues
hidden arethe
are obtained
layers and obtained
input
layerfrom
from fine-tuning
fine-tuning
with various ratios,processes;
processes;
whoseand andvalues
(iii)
optimal (iii) weight
The Theobtained
are weight
of wholeof whole
from neuralneural
fine-tuning networknetwork
processes;is and is
(iii)shared
shared The
weight
among of
among whole
thinned neural
thinned network
neuralneural is
networks
networks shared among
resulting
resulting thinned neural
from Dropout.
from Dropout. networks resulting
The architecture
The architecture of theoffrom
NN Dropout.
thewith
NN with
The architecture of the NN with Dropout mechanism is schematically illustrated in Fig. 1.
DropoutDropout mechanism
mechanism is schematically
is schematically illustrated
illustrated in Figure
in Figure 1. 1.

(a) Conventional neural network (b) Conventional neural network

(a) Conventional
(a) Conventional neuralneural network
network (b) Dropout
(b) Dropout neuralneural network
network
Figure 1. Architecture of Neural Network with Dropout mechanism
FigureFigure 1. Architecture
1. Architecture of Neural
of Neural NetworkNetwork with Dropout
with Dropout mechanism
mechanism
(a) The conventional plain neural network where every perceptron in a layer are connected to all perceptrons
of(a)
(a) The theThe
nextconventional
conventional plain
layer; (b) plain
The neural
NN neural
withnetworknetwork
Dropout where where
where aevery every perceptron
perceptron
fraction of number in a are
layer
inofaperceptron
layer are connected
connected
(circled tois to
cross)
all perceptrons
all perceptrons of the of thelayer.
next next layer.
(b) The
randomly (b)NN
Thewith
NNDropout
neglected with Dropout
during wherewhere
calculation a fraction
a fraction of number
of number of of
perceptron
perceptron (circled
(circled cross)cross) is randomly
is randomly neglected
neglected duringduring calculation.
calculation.
2.1. Deep Neural Networks (DNN)
2.1. Deep Neural
2.1. Deep Networks
Neural (DNN)
Networks (DNN)
Conceptually, DNN mimic the information analysis process of the brain, in which meaningful
sets of information are extracted from received data through a complex system of neurons connected
3
together. Hence, a typical DNN architecture consists 3of multiple layers, each layer comprised various
number of perceptrons, together they form a nonlinear mapping from input data to output results.
Fig. 1(a) depicts an example of the DNN architecture. Mathematically, one can formulate this nonlin-
ear mapping as follows [11]:

Y = F(X|W) = fL (...( f2 ( f1 (X|W1 )|W2 )...WL ) (1)

where L is the total number of layers in the network including the input and output layers, W is the
matrix of the network’s parameters; X and Y denote the input and output vectors, respectively.
The connection between two consecutive layers is performed by combining a linear matrix oper-
ation and a nonlinear activation function as follows:

fl (Xl |Wl ) = h(Wl × Xl + bl ) with l = 1, 2, . . . , L (2)

where Xl is the input of the l layer; bl is bias vector; and h denotes the nonlinear activation function.
In this study, one adopts the Rectified Linear Unit (ReLU) function for hidden layers and the softmax
function for the output layers.
Herein, the network parameters are computed in a supervised fashion, which means the structural
state is labeled in advance, then the database is divided into three groups, namely training, validation,
32
 Hung, D. V., et al. / Journal of Science and Technology in Civil Engineering

and testing datasets. The values of network parameters are randomly initialized, then they are itera-
tively updated via the backpropagation algorithm to minimize the deviation denoted by a loss function
L between prediction values and annotated structural states. At a step t of the training process, the pa-
rameter values are updated by:
ij ij j
wl (t) = wl (t − 1) + ηδil xl−1 (t) (3)

where l denotes the layer number; δil is the error at node i of layer l, which is calculated backward from
the derivation at output layer δL ; and η stands for the learning rate. The training process terminates
when a tolerable error level is met or the number of iterations reaches to a limit.

2.2. Dropout mechanism

In general, deep neural networks contain a large number of parameters, i.e. weight matrix W,
thus training such models requires a significant amount of available and well-curated data. If data is
limited, these are many possible configurations which can provide accurate results on training data,
but most of these formers eventually perform poorly on unseen testing data, namely the overfitting
phenomenon. To tackle these difficulties, one can adopt a combination approach averaging prediction
results of different configurations. Because a combined model is always more generalized than a sin-
gle one, it will improve the prediction performance. However, it also required significantly increasing
resources both in time and in efforts. More recently, Srivastava et al. [12] proposed the dropout mech-
anism which efficiently deals with limited data by providing a probability distribution rather than a
deterministic estimate. Then, the statistical properties such as the mean value and the standard devi-
ation of quantities of interest can be derived. In this way, one is able not only to obtain a predicted
result but also to assess how much the result confidence is. The mechanism of the dropout technique
is described in detail as follows.
Considering a node j in layer l, its state is independently sampled from a Bernoulli distribution
with a predefined probability p, as follows:

r(l)
j ∼ Bernoulli(p) (4)

where r(l)
j is a binary variable indicating the state of the considered node with 0 corresponding to be
dropped out, while 1 to be active. Then the input vector for layer l of the network is updated by:

el = Xl ⊗ rl
X (5)

where ⊗ denotes the element-wise product, Xl are input values for layer l of the original plain network,
el for updated dropout network. Then, Eq. (2) is rewritten as below,
X

fl (Xl |Wl ) = h(Wl × X

el + bl ) (6)

The above process is repeated for every layer in the network architecture except the output layer.
After that, the network is trained by performing a number of iteration of forward-pass for calculating
output results and backward-pass to update the weight values W. Note that, at each iteration, differ-
ent nodes are eliminated with a probability p, forming a thinner architecture than the original plain
architecture.
Intuitively, the Dropout Network architecture can potentially improve the prediction ability thanks
to two-folds: i) it works as a regularization technique reducing the overfitting problem by not being
33
 Hung, D. V., et al. / Journal of Science and Technology in Civil Engineering

over-relied on any node in the network structure, as it can be eliminated. ii) It is well-known that
for a specific database, there is a particular best-fitted architecture providing the highest prediction
accuracy, but it is not generalizable for other data. In contrast, the Dropout mechanism provides a
more generalized architecture but not necessarily larger as the effect of ISSN
Journal of Science and Technology in Civil Engineering 2020
eliminating
1859-2996
nodes, which can
provide non-optimal but satisfied prediction accuracy with different data.
3. Pushover analysis database
3. Pushover analysis database
Pushover analysis (PA) is a widely used static procedure to approximate non-linear
Pushover analysislateral
building (PA)deformations
is a widelybeyond
used the elastic
static range, up to
procedure to approximate
the failure occurrence.
non-linear building lat-
Compared to elastic analysis, PA can encompass more realistic behaviors including
eral deformations beyond the elastic range, up to the failure occurrence. Compared to elastic analysis,
plastic hinges, non-linear constitutive law of material, P-Delta effect, and staged
PA can encompass more realistic behaviors including plastic hinges, non-linear constitutive law of
construction depending user-defined tasks. In this study, to perform PA of structures,
material, P-Delta effect,
one adopts theand staged program
open-source construction
OpenSees depending
[13] crediteduser-defined tasks. In this study, to per-
by the Pacific Earthquake
form PA of structures, one adopts the open-source program OpenSees [13] credited by the Pacific
Engineering Research Center thanks to its accurateness, openness, and high computation
Earthquake Engineering Research
speed, appreciated Center thanks
by the earthquake researchtocommunity.
its accurateness, openness, and high computation
speed, appreciated by the earthquake research community.

Figure 2. TypicalFigure 2. Typical

RC frame underRCaframe under a concentrated
concentrated horizontalhorizontal forcetop
force at the at the topfor
floor floor
pushover analysis
for pushover analysis

In this section,Inanthis section,using

example an example
OpenSeesusingtoOpenSees
carry outto acarry
PA isout a PAdemonstrated.
briefly is briefly Let’s take an
example of a 2-Ddemonstrated. Let’s take
asymmetric an example
3-story of a 2-D
and 2-bay RCasymmetric
frame with 3-story and 2-baystory
a constant RC frame
height H, and its bay
with a constant story height 1, and its bay widths 2 as shown in Figure 2. The frame is
widths B as shown in Fig. 2. The frame is subjected to uniformly distributed vertical load acting on
subjected to uniformly distributed vertical load acting on all its beams in addition to the
all its beams in addition to the self-weight. A concentrated horizontal force is applied at the leftmost
self-weight. A concentrated horizontal force is applied at the leftmost top floor.
top floor.
The nonlinear constitutive law of steel is constructed using the model of Filippou et al. [14]. At
first, the behavior of the steel is approximated by a bilinear curve composed of a branch for an initial
linear elastic state with slope E0 , and another branch for hardening state characterized by a ratio b
resulting in a slope E1 = bE0 . Next, the transition between the two states is controlled by a curvature
parameter R0 , as shown in Fig. 3(a). The values of b and R0 are initially adopted as 0.02 and 20 as
recommendations in general cases by [13].
The constitutive law of concrete is simulated using the Kent-Scott-Park model [15] composed of
three parts, as shown in Fig. 3(b) where negative and positive values denote compression and tension,
respectively. The elastic state of compression starts 6 with an initial slope E0 , reaching the maximum
strength of f pc , i.e., 28-day compressive strength, and the corresponding strain ε0 . Mathematically, E0
can be calculated by the following equation:
2 f pc
E0 = (7)
ε0
34
 Journal
Journal of Science
of Science and Technology
and Technology in Civil
in Civil Engineering
Engineering 2020 2020 ISSNISSN 1859-2996
1859-2996
Hung, D. V., et al. / Journal of Science and Technology in Civil Engineering

(a)(a)(a) Bilinear
Bilinear
Bilinear constitutive
constitutive
constitutive law of law
law ofsteelof steel
steel (b)(b)
Kent-Scott-Park
Kent-Scott-Park model
(b) Kent-Scott-Park model model ofofconcrete
concrete
of concrete
Figure
Figure
Figure 3. 3. Material’s
3. Material’s
Material’s elasto-plastic
elasto-plastic
elasto-plastic constitutive
constitutive
constitutive law law
law

AfterThe The nonlinear

nonlinear
reaching constitutive
constitutive
f pc , the strength law of lawsteel
decrease, of steel
while theiscompressive
constructed
is constructed using using the model
the model
strain ofincreases
Filippou up
of Filippou
continuously
to et et [14].
εu , al. al. [14].
so-called At At first,
first,
crushing the
strain,theandbehavior
behavior of the
respective of crushing
the steel
steel is approximated
is approximated
strength f pcU =by0.2aby .aOn
bilinear
bilinear
f pc thecurve curve
other hand,
composed
thecomposed of
of a branch
tensile behavior a branch for an
for an initial
of concrete initial linear
linearby
is modeled elasticelastic state
state with
a bilinear with
curveslope slope
of which 3
34 , and , and
4 another
the another branch
branchhas the
initial branch
same for
slope hardening
E , after state
reaching characterized
the maximum by a ratio
tensile / resulting
strength
for hardening state characterized by a ratio / resulting in a slope 35 = 5/34 . Next,
0 f t , in
the a slope
tensile 3
strength = /3 4 . Next,
decreases the a
the with
softening slope
transition E .
tsbetween the two states is controlled by a curvature parameter 6 , as shown in
transition between the two states is controlled by a curvature parameter 64 , as4shown in
In terms of section modelling, the section of RC element is simulated using the fiber approach
Figure Figure3(a). 3(a).
The The values values
of /of and / and64 are 64 are initially
initially adopted adopted asand as 0.02
0.02 and 20 as
which is able to account for moment-curvature, axial force-deformation, theirand 20 as
interaction at the
same recommendations
recommendations
time [16] thus being in
in general general
superiorcases cases
than bythe by
[13]. [13]. section approach calculating independently the
uniaxial
bending and normalThe constitutive
The constitutive law of law
stress. Specifically,of concrete
concrete is simulated
each concrete
is simulated usingusing
section
the the Kent-Scott-Park
is decomposed
Kent-Scott-Park model
[15][15]
into a predefined
model num-
ber of fibers,
composedhorizontally
of three
Journal of Science and vertically
parts,
and Technologyas shownas shown in
in Figure
in Civil EngineeringFig. 4.
3(b)For the
where rebar, the
negative section
and is defined
positive valuesei-
with
composed
Journal of
of three
Science and
parts, as shown
Technology in Civil
in Figure
Engineering
3(b)2020
2020
where negative andISSN
ISSN
1859-2996
positive
1859-2996
values
ther steel evenly distributed around the perimeter as in columns or per-layer (top/intermediate/bottom)
denote compression and tension, respectively. The The elasticstatestate of compression startsstarts
as denote
in beams. compression and tension, respectively. elastic of compression
with with an initial
an initial slopeslope 34 , reaching
34 , reaching the maximum
the maximum strength
strength of '78of, i.e.,'78 , 28-day
i.e., 28-day compressive
compressive
strength,
strength, and the andcorresponding
the corresponding strainstrain 94 . Mathematically,
94 . Mathematically, 34 can 34 becancalculated
be calculated by the
by the
following
following equation:
equation:
2 f pc2 f pc
E0 =E0 = e (7) (7)
e 0 0

AfterAfter
reaching '78 , '78
reaching the, the strength
strength decrease,
decrease, whilewhile
the the compressive
compressive strain
strain
continuously
continuously increases
increases 9: , so-called
up toup9:to, so-called crushing
crushing strain,
strain, and and respective
respective crushing
crushing
strength ' '=
strength = 0.2'
0.2'
78;
78; .(a)
On78 .the
Onother
78
the other
hand,hand, the tensile
the tensile behavior
behavior of concrete
(b) of concrete is modeled
is modeled
(a)
(a) of which the initial branch has the (b) (b)
by a by a bilinear
bilinear curvecurve
of which the Figure
initial 4.branch samesame
slopeslope
has themodelling 34 , after
34 , after reaching
reaching the the
Fiber section
maximum
maximum tensile
tensile strength
strength '
the? ,tensile
' ,Figure the
Figure tensile
4. 4.
Fiber strength
section
strength
Fiber decreases
modelling
decreases
section withwith
modelling a softening
a softening 3 . 3?@ .
slopeslope
?
In terms of section modelling, the section of RC element is simulated using the fiber?@
In terms of section modelling, the section of RC element is axial
approach which is able to account for moment-curvature, simulated using the fiber and
force-deformation,
The output
approach of interest
which is able of the simulation is the pair (D , F ), which are the drift ratio of the top
their interaction at to
theaccount
same timefor moment-curvature, top axial force-deformation,
[16] thus being superior base
than the uniaxial and
section
floor their
and the corresponding
interaction base
at the same shear. Fig.
time [16] the 5 depicts
thusbending a typical
being superior relationship
thanstress. between
the uniaxial Dtop and Fbase
sectioneach
approach calculating independently and normal Specifically,
duringapproach
PA. It can be seen independently
calculating that with a small bending
drift ratio (<normal the base
0.01), stress. shear linearly
each increases with
concrete section is decomposed the and
into a predefined number of fibers, Specifically,
horizontally and
increasing drift,
concrete after
section passing an
is decomposed yield point, a likely plateau region occurs before a softening trend
vertically as shown in Figureinto a predefined
4. For the rebar, number
the section of fibers, horizontally
is defined andsteel
with either
drivesvertically
the structure’s behavior.
evenlyas shown in Figure
distributed 4. For the
around the rebar, the section
perimeter as isindefined with either
columns steel
or per-layer
evenly distributed around the 7 35 7 as in columns or per-layer
perimeter
(top/intermediate/bottom) as in beams.
(top/intermediate/bottom) as in beams.
 approach calculating independently the bending and normal stress. Specifically, each
concrete section is decomposed into a predefined number of fibers, horizontally and
vertically as shown in Figure 4. For the rebar, the section is defined with either steel
evenly distributed around the perimeter as in columns or per-layer
(top/intermediate/bottom)
Hung, D. V., et al. / as in beams.
Journal of Science and Technology in Civil Engineering

Figure
Figure 5. Typical
5. Typical force-displacementcurve
force-displacement curveobtained
obtained from
from PA
PA using
usingFEM.
FEM
The output of interest of the simulation is the pair (A?B7 , CDE@F ), which are the drift
ratio of the top floor and the corresponding base shear. Figure 5 depicts a typical
In summary, the input for each simulation includes 15 variables, and the output is the correspond-
relationship between A?B7 and CDE@F during PA. It can be seen that with a small drift
ing base shear. By varying randomly the values of 15 variables: the number of stories N s in [1, 10], the
ratio (<0.01), the base shear linearly increases with increasing drift, after passing an
number of bays Nbay in [1, 6], the story height H s in [2, 6] m, the column size Hcol in [300, 600] mm,
the beam height Hbeam in [300, 600] mm, the beam width Bbeam in [300, 600] mm, Young’s modulus
of steel E s in [180, 220] MPa, yield strength of steel fy [300, 400] MPa, tensile strength of concrete
ft in [2-5] MPa, compressive strength of concrete8 fc in [20-40] MPa, compressive strain of concrete
ε0 in [0.002, 0.004], crushing strain of concrete εu in [0.08, 0.12] and tensile strain of concrete εt
in [0.0002, 0.0003]. These ranges of values are taken from the literature and engineering practices
[17, 18]. In total, 1000 simulations with OpenSees are carried out, then the base shear is recorded for
10 values ratio drift from 0 to 5% with an increment of 0.5%, resulting in a database of 10000 lines
as shown in Table 1. This database is further divided into training/validation/test data with a widely
used ratio 8/1/1 for training the Dropout network.

Table 1. Pushover database for training the Dropout network obtained from FEM

Steel Concrete
Hs Lbay Hcol Hbeam Bbeam Fbase
No Ns Nbay E fy fc ε0 ft εt εu ∆%
(m) (m) (m) (m) (m) kN
GPa MPa MPa % MPa % %
1 3 3 2 3 0.3 0.3 0.4 180 300 20 0.2 2 0.02 8 0.5 300
2 3 3 2 3 0.3 0.3 0.4 180 300 20 0.2 2 0.02 8 1 340
3 3 3 2 3 0.3 0.3 0.4 180 300 20 0.2 2 0.02 8 1.5 380
... ... ... ... ... ... ... ... ... ... ... ... ... ...
Min 1 1 2 2 0.3 0.3 0.3 180 300 20 0.2 2 0.02 8 0.5 4
Max 10 6 6 6 0.6 0.6 0.6 220 400 40 0.4 5 0.03 12 5 717

4. Computational results

In this section, the training process and test results of the Dropout network are presented. For the
architecture of the network, a preliminary study conducted by the authors pointed out that an architec-
ture of 15/256/128/64/32/1, meaning 5 layers: input layer with 15 perceptrons for 15 variables and 4
hidden layers with the numbers of perceptrons are 256, 128, 64 and 32, respectively and output layer
is 1 corresponding to the base shear value. More detail of the grid search for selecting an appropriate
architecture of MLP can be found in [1]. For the value of Dropout probability, one adopts the value
36
 Hung, D. V., et al. / Journal of Science and Technology in Civil Engineering

0.5 recommended by [12]. In terms of training process hyper-parameters, one used the Adam opti-
mizer with an initial learning rate l = 0.001, which was reduced by a factor of 0.01 when there was
no validation loss improvement for five consecutive iterations, and the total number of iterations is
300. The performance criteria, a.k.a, the loss function is defined using the mean square error (MSE)
measuring the average of the squared difference between the predicted base shears and actual values
as follows:
N
1 X  actual predict 2

MS E = Fbase,i − Fbase,i (8)
N i=1
Journal of Science and Technology in Civil Engineering 2020 ISSN 1859-2996
The evolution of the mean square error (MSE) between base shears predicted by the Dropout
network and those from the FEM model during the training process is depicted in Fig. 6. As expected,
the average of the squared difference between the predicted base shears and actual
the MSE decreases with increasing numbers of iterations; in addition, the histogram of base shears
values as follows:
predicted are also plotted at the bottom, visually show the improvement of the model’s performance.
5 7UFVR8? W
At the end of the training process, anMN3 = ∑O
excellent (C E8?:E" −between
agreement CDE@F,R ) results is apparently
(8) observed.
O RT5 DE@F,R

Figure 6. EvolutionFigure 6. Evolution

of mean of mean
square error square
(MSE) error (MSE)
between betweennetwork’s
the Dropout the Dropout network’s
results and those of FEM
results and those
modelofduring
FEM model during
training training process.
process
The evolution of the mean square error (MSE) between base shears predicted by
After the the Dropout
training networkthe
process, andDropout
those from the FEMismodel
network duringusing
validated the training process isPA results re-
experimental
ported in [19].depicted in Figure 6. consists
The experiment As expected,
of athe MSE decreases
2-story and 2-bay with
RCincreasing
frame, as numbers
shownofin Fig. 7 with
iterations; in addition, the histogram of base shears predicted 2 also plotted at the
are
detailed dimensions, whose columns’ cross-sections is 400 × 400 mm with 10-D19 rebar and beams’
bottom,
cross-section is 300 ×visually
400 mm show2 the improvement of the model’s performance. At the end of the
with 8-D19 rebar, respectively.
training process,
The comparison results an areexcellent agreement
illustrated between
in Fig. results is apparently
8, showing a very goodobserved.
agreement between the
experimental curveAfterand the
the training
predictionprocess,
curvethefrom
Dropoutthe network
presentismethod.
validatedInusing experimental
addition, corresponding 90%
PA results
confidence interval reported
(CI), in [19]. The
represented experiment
by shaded consists
area is alsoof aderived
2-story and
by 2-bay RC frame,
repeating the inference 100
as shown in Figure 7 with detailed dimensions, whose columns’
times, each time the architecture of the network is stochastically altered by the Dropout cross-sections is mechanism.
2 2
400´400 mm with 10-D19 rebar and beams’ cross-section is 300´400
As is known, 90% CI means there is a 90% probability that the actual value of concerned quantity mm with 8-D19
lies within therebar,
CI. respectively.
The comparison
Note that the elastic period has results are illustrated
relatively smallinCIFigure
than8,theshowing a very
plastic zone,good agreement
i.e., less uncertainty, and
between
the CI is widened the structure
as the experimental curve and
is more the prediction
damaged. curve
In short, from
this the present result
comparison method.reliably
In confirms
addition, corresponding
the accuracy of the present Dropout network. 90% confidence interval (CI), represented by shaded area is also
derived by repeating the inference 100 times, each time the architecture of the network

37
10
 is stochastically altered by the Dropout mechanism. As is known, 90% CI means there
is a 90% probability that the actual value of concerned quantity lies within the CI.
Note that the elastic period has relatively small CI than the plastic zone, i.e., less
uncertainty, and the CI is widened as the structure is more damaged. In short, this
Hung, D. V.,
comparison al. / Journal
etreliably
result confirmsofthe
Science
accuracyand Technology
of the in Civil
present Dropout Engineering
network.

Figure 7. Experimental RC frame [19]

Figure 7. Experimental RC frame [19]

Next, the Dropout network is used to perform parametric studies, providing more insights about
the influences of main input parameters on structures’ non-linear behavior. Starting from the above
two-bay RC frame, one varies the story height, the number of stories, the column size, and the bay
width, then calculates the base shear corresponding to 5% drift ratio, and plots evolution curves.
For clarity, Journal
only one parameter is modified each time, while the other inputsISSN
of Science and Technology in Civil Engineering 2020
remain the same as the
1859-2996
experimental model.
Figure 8. Validation results of Dropout network using experimental data.
Next, the Dropout network is used to perform parametric studies, providing more
insights about the influences of main input parameters on structures’ non-linear
behavior. Starting from the above two-bay RC frame, one varies the story height, the
number of stories, the column size, and the bay width, then calculates the base shear
corresponding to 5% drift ratio, and plots evolution curves. For clarity, only one
parameter is modified each time, while the other inputs remain the same as the
experimental model.

11

Figure
Figure 8. Base
8. Base shear
shear at 5%
at 5% drift
drift ratio
ratio in in
thethe functions
functions ofof inputparameters
input parameters
Figure 8 demonstrates the evolution of base shear at a 5% drift ratio in the functions
Fig. 8 demonstrates the evolution
of these investigated of base
parameters. shear
It can be atseena 5%thatdrift
the ratio in thedecreases
base shear functionswith
of these inves-
tigated parameters. It can be seen that the base shear decreases with
increasing story height, though a greater story height meaning a larger lateral increasing story height, though
a greater story height meaning
displacement. a largerthat
This implied lateral
the displacement.
columns lost their Thisstrength
implied with
that the
suchcolumns
large lost their
strength with
displacement. A decreasing trend of base shear is also observed as the number of storiesas the num-
such large displacement. A decreasing trend of base shear is also observed
ber of stories increases,
increases, but withbuta significantly
with a significantly slow to
slow rate owing rate
the owing to the
story beams story beams
contributing to thecontributing
to the frame’s lateral
frame’s rigidity
lateral rigidityand
andreduce columns’slenderness.
reduce columns’ slenderness. In contrast,
In contrast, increasing
increasing columncolumn size
strengthen size
the frame rigidity;
strengthen thus,rigidity;
the frame the basethus,
shear
themanifests
base shearan upwardan
manifests tendency. Meanwhile, the bay
upward tendency.
width interestingly
Meanwhile, shows two
the bay regimes:
width an upward
interestingly showstrend from 2antoupward
two regimes: 4 m, then
trendafrom
gradual
2 to downward
trend. Similar
4 m,parametric
then a gradualstudies for other
downward trend.parameters or even studies
Similar parametric multipleforparameters can be
other parameters or carried out
with the help of multiple
even the Dropout network
parameters candepending
be carried on outusers’
with theinterests.
help of the Dropout network
depending on users’ interests.
6. Conclusions
In this study, a data-driven approach38 for pushover analysis of RC frames is
investigated by leveraging the probabilistic nature of the Dropout mechanism and the
well-known nonlinear mapping capacity of the neural network, providing a novelty
alternative to the conventional model-based methods. Knowing that training data is of
paramount importance for any data-driven approach, a rigorous numerical simulation
 Hung, D. V., et al. / Journal of Science and Technology in Civil Engineering

5. Conclusions

In this study, a data-driven approach for pushover analysis of RC frames is investigated by leverag-
ing the probabilistic nature of the Dropout mechanism and the well-known nonlinear mapping capac-
ity of the neural network, providing a novelty alternative to the conventional model-based methods.
Knowing that training data is of paramount importance for any data-driven approach, a rigorous nu-
merical simulation suite has been carried out using the reliable OpenSees software accredited by the
engineering community accounting for the nonlinearity of both concrete and steel materials, as well
as the forming of plastic hinges. The simulation suite results in a 10000-sample database sufficiently
large for training a data-driven model.
After being trained, the Dropout network is validated through a comparison with experimental
measurement published in the literature, which confirms its accuracy and also displays the respective
confidence interval, which offers a margin of safety for design solutions. Next, parametric studies are
conducted with the help of the present method whose results reveal more insight into the structures’
behavior, such as the relationship between the number of stories and the base shear, the effect of
column size, and material properties in reducing the base shear. For the type of structure considered
in this study, one posits that there are a positive correlation between the lateral rigidity of the structure
with column sizes, a negative correlation with the story height, and a bimodal positive-then-negative
correlation with the bay width.
In the future, it is interesting to extend the method to structures’ behavior under cyclic loads
or even under extreme loading as ground motions, shock loading, but a more sophisticated neural
network architecture will be resorted. Another promising direction is to account for multiple criteria
rather than a single pair (Dtop , Fbase ) as done in the real design process of structures for which an
ensemble of dropout models will be developed; each model specifically solves a single criterion.

References
[1] Hung, D. V., Hung, H. M., Anh, P. H., Thang, N. T. (2020). Structural damage detection using hybrid
deep learning algorithm. Journal of Science and Technology in Civil Engineering (STCE)-NUCE, 14(2):
53–64.
[2] Hung, T. V., Viet, V. Q., Thuat, D. V. (2019). A deep learning-based procedure for estimation of ulti-
mate load carrying of steel trusses using advanced analysis. Journal of Science and Technology in Civil
Engineering (STCE)-NUCE, 13(3):113–123.
[3] Javidan, M. M., Kang, H., Isobe, D., Kim, J. (2018). Computationally efficient framework for probabilistic
collapse analysis of structures under extreme actions. Engineering Structures, 172:440–452.
[4] Mangalathu, S., Burton, H. V. (2019). Deep learning-based classification of earthquake-impacted build-
ings using textual damage descriptions. International Journal of Disaster Risk Reduction, 36:101111.
[5] Feng, D.-C., Xie, S.-C., Deng, W.-N., Ding, Z.-D. (2019). Probabilistic failure analysis of reinforced
concrete beam-column sub-assemblage under column removal scenario. Engineering Failure Analysis,
100:381–392.
[6] Brunesi, E., Nascimbene, R., Parisi, F., Augenti, N. (2015). Progressive collapse fragility of reinforced
concrete framed structures through incremental dynamic analysis. Engineering Structures, 104:65–79.
[7] Guo, Z., Gao, R., Zhang, X., Jia, X. (2018). Fire Resistances of Restrained Steel Beams Subjected to Fire
Loads. KSCE Journal of Civil Engineering, 22(8):3028–3038.
[8] Worrell, C., Luangkesorn, L., Haight, J., Congedo, T. (2019). Machine learning of fire hazard model
simulations for use in probabilistic safety assessments at nuclear power plants. Reliability Engineering &
System Safety, 183:128–142.
[9] Fu, F. (2020). Fire induced progressive collapse potential assessment of steel framed buildings using
machine learning. Journal of Constructional Steel Research, 166:105918.
39
 Hung, D. V., et al. / Journal of Science and Technology in Civil Engineering

[10] Gal, Y., Ghahramani, Z. (2016). Dropout as a bayesian approximation: Representing model uncertainty
in deep learning. In International Conference on Machine Learning, 1050–1059.
[11] Burkov, A. (2019). The hundred-page machine learning. Quebec City, Canada.
[12] Srivastava, N., Hinton, G., Krizhevsky, A., Sutskever, I., Salakhutdinov, R. (2014). Dropout: a simple
way to prevent neural networks from overfitting. The Journal of Machine Learning Research, 15(1):
1929–1958.
[13] McKenna, F., Fenves, G. L. OpenSees manual. PEER Center.
[14] Filippou, F. C., Popov, E. P., Bertero, V. V. (1983). Effects of bond deterioration on hysteretic behavior of
reinforced concrete joints. Report EERC 83-19, Earthquake Engineering Research Center, University of
California, Berkeley.
[15] Hisham, M., Yassin, M. (1994). Nonlinear analysis of prestressed concrete structures under monotonic
and cycling loads. PhD thesis, University of California, Berkeley.
[16] Taucer, F., Spacone, E., Filippou, F. C. (1991). A fiber beam-column element for seismic response anal-
ysis of reinforced concrete structures, volume 91. Earthquake Engineering Research Center, College of
Engineering, University of California.
[17] Engineering ToolBox. Concrete - Properties. Accessed 02 July 2020.
[18] Wang, C., Xiao, J., Sun, Z. (2016). Seismic analysis on recycled aggregate concrete frame considering
strain rate effect. International Journal of Concrete Structures and Materials, 10(3):307–323.
[19] Sung, Y., Lin, T.-K., Hsiao, C., Lai, M. (2013). Pushover analysis of reinforced concrete frames consid-
ering shear failure at beam-column joints. Earthquake Engineering and Engineering Vibration, 12(3):
373–383.

Appendix A.
As a second validation case, one compares the Dropout network with a FEM unseen on training
data in performing pushover analysis. The structure of interest has 9 stories with input parameters are
listed in Table A.1. The figure show that the predicted results encompasses those from FEM, thus
confirming the credibility of the proposed approach.
Table A.1. Input parameters for validation case with FEM

Steel Concrete
Hs Lbay Hcol Hbeam Bbeam
No Ns Nbay
(m) (m) (m) (m) (m) E fy fc ε0 ft εt εu
GPa MPa MPa % MPa % %
Journal of Science and Technology in Civil Engineering 2020 ISSN 1859-2996
1 9 3 2 3.2 0.35 0.3 0.25 200 460 27.5 0.2 2 0.02 8

Figure
Figure 9. Validation
A.1. Validation results
resultsofofDropout
Dropoutnetwork using
network FEM
using model.
FEM model
References 40
[1] Hung, D. V., Hung, H. M., Anh, P. H., & Thang, N. T. (2020). Structural damage detection
using hybrid deep learning algorithm. Journal of Science and Technology in Civil
Engineering (STCE)-NUCE, 14(2), 53-64.
[2] Hung, T. V., Viet, V. Q., & Thuat, D. V. (2019). A deep learning-based procedure for