Analysis of RC Buildings by Metamodel

Approaches

Deepak Sahu, Pradip Sarkar, and Robin Davis

Abstract Response of RC structures is very complex and dynamic in nature because

of the vulnerabilities that exist in geometry, material properties and loading. To repre-
sent the random dynamic responses accurately, stochastic analysis is chosen here.
The stochastic analysis can be done by two types of methods, like statistical and
the non-statistical approaches. In non-statistical approach a relationship is devel-
oped between the input and random output or responses. Computationally efficient,
simplified methods are required as an alternative to Monte–Carlo simulation which
is considered as an accurate method for stochastic analysis. The present study is
an evaluation of the different non-statistical metamodel-based approaches such as
high-dimensional model representation and using design of experiments approaches
like central composite design, Box–Behnken design and full factorial design for
the representation of the response surface. The effectiveness of high-dimensional
model representation over conventional response surface metamodel approaches is
discussed in the present study with regard to two contexts, namely free vibration
response and nonlinear time history responses of RC frames. The seismic fragilities
obtained using high-dimensional model representation are compared with estab-
lished metamodel approaches for effectiveness and computational efficiency. It is
found that the use of high-dimensional model representation yields fairly accurate
results with even less computational effort.

Keywords Stochastic analysis · HDMR · RSM

D. Sahu (B) · P. Sarkar · R. Davis

Department of Civil Engineering, National Institute of Technology Rourkela, Rourkela 769008,
India
e-mail: [email protected]
P. Sarkar
e-mail: [email protected]
R. Davis
e-mail: [email protected]

© Springer Nature Singapore Pte Ltd. 2021 817

R. M. Singh et al. (eds.), Advances in Civil Engineering, Lecture Notes
in Civil Engineering 83, https://doi.org/10.1007/978-981-15-5644-9_65
818 D. Sahu et al.

1 Introduction

Prediction of structural behaviour in the case of unexpected loadings, such as earth-

quakes, cyclones, tsunamis and hurricanes, is very complex and difficult, since incon-
sistency existed in both the nature of loading and structural resistance [1]. To address
this problem stochastic approach can be used to predict the response of the struc-
ture for these catastrophic events in a probabilistic manner. Monte-Carlo simulation
(MCS) is a leading technique [2], which consists of large number of sampling process
to obtain the responses, but becomes computationally affluent for problems having
small probabilities among the commonly used statistical approaches [3].
Apart from statistical approach, the response of the structure is evaluated at some
particular set of random parameters in the case of non-statistical method. A func-
tional relationship is generated between the input and output variables known as a
metamodel [4]. Response surface method (RSM) [5] is the most broadly used non-
statistical approach, which forms the metamodel. The selection of sampling points
for the functional evaluation of the responses in RSM can be done by different tech-
niques for design of experiments (DOE) such as central composite design (CCD)
[6] and Box–Behnken design (BBD) [7]. A high-dimensional model representation
(HDMR) [8] applied in different fields as method for stochastic analysis was intro-
duced recently, to represent the response surface in terms of a metamodel for the
study of structural behaviour [9] HDMR is used recently.
The focus of this work is to scrutinize the effectiveness of the recently introduced
HDMR method, to problems involving free vibration analysis of concrete buildings,
with reference to established response surface methods.

2 Review of Metamodel-Based Approaches

The present study considers the problem to obtain the random natural frequency
for stochastic analysis of a RC symmetric bare framed building. Stochastic analysis
of the selected problems requires the evaluation of metamodel using each method,
HDMR, and RSM using different sampling techniques such as CCD and BBD. This
section provides a review of the above methods.

2.1 RSM Approaches

The RSM metamodel can be represented by its general form as shown in Eq. (1)

y = f (xi ) + ε (1)
Analysis of RC Buildings by Metamodel Approaches 819

Here, y represents the response (output), x i represents the input variables, and ε
represents the error in estimation and can be neglected in the case of computer analysis
[10]. The response surface input variables are the parameters whose uncertainty
or randomness can cause uncertainty in the output or response. A second-order
(quadratic) function is selected in the present study to evaluate the response. The
form of such a function is shown in Eq. (2).


k 
k 
k−1 
k
y = β0 + βi xi + βii xi2 k + βi j xi x j (2)
i=1 i=1 i=1 j=1

where β is unknown coefficients, and k is the total number of input variables. To

determine the unknown constants (β), the design of experiments should be carried out
at the selected sampling points to obtain the response. Then, a functional relationship
is established between the random input parameters and output responses which is
called a metamodel. The different design methods implemented in this work are
CCD, and BBD gives a different combination of the input variables. Details on these
sampling methods can be found in the literature [11, 12].

2.2 HDMR Metamodel Approach

HDMR method is used to incarnate the probability analysis of a system which needs
a huge computational cost and effort. In other words, tedious processes like the MCS
can be carried out on this compact model effectively and efficiently [13].
The input variables and the output function in HDMR can be conveniently repre-
sented as the N-dimensional vector x = {x 1 , x 2 , …, x N } and f (x), respectively. The
value of N can vary up to the order of 100–1000 or more. Similar to other response
surface methods, the effect of input variables on the output function can be indepen-
dent and/or correlated. HDMR expresses the output f (x) as a hierarchical correlated
function expansion in terms of the input variables as in Eq. (3).


N   
f (x) = f 0 + f i (xi ) + f i1 i2 xi1 , xi2
i=1 1<i 1 <i 2 <N
    
+ f i1 i2 i3 xi1 , xi2 , xi3 + . . . + f i1 i2 ......i N xi1 , xi2 , ...xi N (3)
1<i 1 <i 2 <i 3 <N

where f 0 is the constant obtained by the response f (x) at a selected reference point
(mean point), f i (x i ) is the first-order
 term
 representing the individual contribution of
the variable on the output, f i1 i2 xi1 , xi2 is the second-order  term for the  cooperative
effects of the variables xi1 and xi2 on the output, f i1 i2 ...i N xi1 , xi2 , ..., xi N is the residual
dependence of all the input variables that influence the output f (x). In the present
820 D. Sahu et al.

Fig. 1 Two-dimensional
schematic representation of
sampling in HDMR

study, finite-element analysis (FEA) of the structure is carried out at the reference
points c = {c1 , c2 , c3 , . . . c N } consisting of the mean values of the selected variables.
The higher order terms are evaluated as per standard methods in the input variable
space around the reference point. The lower order expansion functions are subtracted
off to remove the dependency.
For the development of HDMR metamodel in this study, the three-point sampling
method is chosen and in Fig. 1, it is explained for two random parameters. The
horizontal and vertical axes represent the random parameters x 1 and x 2 , respectively.
There are five grid points considered to be sampling points for the two random param-
eters, x 1 and x 2 . The center point is represented by mean values of their parameter.
The output value at the five sampling points is required to solve the five constants
of the metamodel function involving two random variables. When there is a higher
number of random variables, the grid will have that many dimensions in space.

3 Description of the Structure

An RC frame with four stories and two bays in both of the two horizontal directions
is selected for the present study as shown in Fig. 2. The building frame is designed
for the highest seismic zone of India (Zone V with PGA of 0.36 g) as per relevant
Indian standards (IS 456-2000, IS 1893-2016) considering medium soil conditions.
The mean values of the characteristic strength of concrete and reinforcement
steel are taken as 25 and 415 MPa, respectively, for design. The selected frame is
modelled for elastic free vibration analysis. Open System for Earthquake Engineering
Simulation (OpenSees) Laboratory tool [17] is used for all the above analyses.
Analysis of RC Buildings by Metamodel Approaches 821

Fig. 2 Details of the selected building with mean values of dimensions

4 Free Vibration Response of RC Frame

The fundamental natural frequency is inexact due to the possible uncertainties in

the geometry and material properties of the structure. Stochastic free vibration anal-
ysis of selected RC building frame using non-statistical (metamodel based) methods
(HDMR, CCD and BBD) is presented in this study to verify the accurateness with
reference to statistical MCS method.

4.1 Modelling of Uncertainties

The uncertainties in the random structural properties are modelled by considering

the most significant parameters as random variables that can affect the output natural
frequency of the building. A total of seven variables, modulus of elasticity (E c ),
storey weight (W ), storey height (h), beam depth (Bd ), beam width (Bw ), column
depth (C d ), column width (C w ), are shown in Table 1, as well as the probability
distributions with statistical parameters of these variables shown in the table are
taken from the previous studies. It is to be noted here that some of the data associated
822 D. Sahu et al.

Table 1 Details of input random variables

S. No. Property Mean COV (%) Probability Source
distribution
1 Modulus of 29 kN/mm2 15 Lognormal Dimitri et al.
elasticity (Ec) (1997)
2 Storey weight 700 kN 10 Normal Ay BÖ (2008)
(W )
3 Storey height (h) 3000 mm 8 Lognormal Bal et al. (2008)
4 Beam depth (Bd ) 350 mm 1.5 Normal Dimitri et al.
(1997)
5 Beam width (Bw ) 300 mm 3 Normal Dimitri et al.
(1997)
6 Column depth 400 mm 1.5 Normal Dimitri et al.
(C d ) (1997)
7 Column width 400 mm 3 Normal Dimitri et al.
(C w ) (1997)

with the statistical details presented in Table 1 relies on local data, which may or
may not be proper for general consideration.

4.2 Development of Metamodel Using HDMR

The random output variable, fundamental natural frequency (ω) of the selected frame
is represented as a function of all the input random variables (given in Table 1). In
order to construct the metamodel, the computational model of the frame is generated
at selected sampling points of the input parameters to conduct free vibration analysis.
The values of the seven random variables obtained at the sampling points (refer to
Fig. 1) are presented in Table 2.

Table 2 Three-point sampling for HDMR

Property Unit μ–δ μ μ+δ
Ec kN/mm2 24.65 29.00 33.35
W kN 630 700 770
h mm 2760 3000 3240
Bd mm 440 450 460
Bw mm 291 300 309
Cd mm 394 400 406
Cw mm 388 400 412
μ and σ are mean and standard deviation of the random variables, respectively
Analysis of RC Buildings by Metamodel Approaches 823

The metamodel representing random natural frequency (ω) using HDMR can be
obtained from Eq. 3 as:

ω = f 0 + f (E c ) + f (W ) + f (h)
+ f (Bd ) + f (Bw ) + f (Cd ) + f (Cw ) (4)

The functions f (E c ), f (W ), f (h), f (Bd ), f (Bw ), f (Cd ), and f (Cw ) in Eq. (4)
are assumed as second-order functions of the respective random variables as,

f (z) = a0 + a1 z + a2 z 2 (5)

The coefficients a0 , a1 and a2 are established by taking the input–output combi-

nations of the three sampling points where all the random variables, except the
considered random variable (z), are at their mean values. The number of random
variables chosen in the problem derives the number of sampling points required in
the HDMR.
Table 3 compares the fundamental frequencies obtained from FEA with those
predicted using HDMR.
The coefficients computed for each random variable for the formation of HDMR
metamodel are presented in Table 4.
To evaluate the accuracy of the metamodel expressions, Fig. 3 depicts the plots
between the actual output frequencies from FEA and the frequencies predicted by
metamodels of various methods at all sampling points. In the figure it can be observed
that the actual analysis results and the results obtained by simulating the metamodel
are almost identical.

4.3 Comparison of the Performance of HDMR with Other

Methods

It is to be noted that once the metamodel is formulated, it can be used for simulations
of output frequencies without the need of further FEA. However, to have a reasonably
accurate estimation of the output frequency, a minimum number of simulations shall
be conducted on the metamodel. A convergence study is conducted considering
the mean frequency response by varying the number of simulations of metamodel
from 10 to 100,000. The variation in the mean frequency versus the number of
simulations obtained for all the metamodel methods is plotted in Fig. 4. The MCS is
also conducted for various number of samples of the same range as a reference. It can
be seen that the mean frequency is converged to a stable value corresponding to the
number of simulations of 100,000 for all the methods. Hence, 100,000 simulations
are conducted for further studies.
The metamodels obtained in each method are simulated for randomly generated
values of all the random input variables. The output natural frequency responses
 824

Table 3 Set of input and output combinations in HDMR

S. No. E c (kN/mm2 ) W (kN) h(mm) Bd (mm) Bw (mm) C d (mm) C w (mm) ω (Hz)
Calculated using FEA Predicted from metamodel
1 29 700 3000 450 300 400 400 2.292 2.292
2 24.65 700 3000 450 300 400 400 2.113 2.113
3 33.35 700 3000 450 300 400 400 2.458 2.458
4 29 630 3000 450 300 400 400 2.361 2.361
5 29 770 3000 450 300 400 400 2.228 2.228
6 29 700 2760 450 300 400 400 2.558 2.558
7 29 700 3240 450 300 400 400 2.069 2.069
8 29 700 3000 443.25 300 400 400 2.271 2.271
9 29 700 3000 456.75 300 400 400 2.313 2.313
10 29 700 3000 450 291 400 400 2.284 2.284
11 29 700 3000 450 309 400 400 2.299 2.299
12 29 700 3000 450 300 394 400 2.286 2.286
13 29 700 3000 450 300 406 400 2.314 2.314
14 29 700 3000 450 300 400 388 2.279 2.279
15 29 700 3000 450 300 400 412 2.336 2.336
D. Sahu et al.
Table 4 Values of input variables obtained by three-point sampling for HDMR
Coefficients E c (kN/mm2 ) W (kN) h (mm) Bd (mm) Bw (mm) C d (mm) C w (mm)
a0 −1.438 0.953 6.423 −2.673 −0.577 35.064 16.685
Analysis of RC Buildings by Metamodel Approaches

a1 5.95E−02 −1.77E+05 −3.26E−03 8.79E−03 3.01E−03 −1.78E−01 −8.58E−02

a2 −3.43E−04 5.89E+09 3.74E−07 −6.33E−06 −3.62E−06 2.25E−04 1.10E−04
825
826 D. Sahu et al.

From Metamodel (Hz) 2.6 2.6

R² = 0.9966

From Metamodel (Hz)

R² = 1
2.4 2.4

2.2 2.2

2 2
2 2.2 2.4 2.6 2 2.2 2.4 2.6
Using FEA (Hz) Using FEA (Hz)

(a) HDMR (b) CCD

2.9
From Metamodel (Hz)

R² = 0.9932
2.6

2.3

1.7
1.7 2 2.3 2.6 2.9
Using FEA (Hz)

(c) BBD

Fig. 3 Comparison of natural frequency responses at sampling points obtained from FEA and
metamodel methods

1.6
Mean Fundamental Frequency

1.5
(Hz)

1.4 HDMR
BBD
CCD
MCS
1.3
10 100 1000 10000 100000

Number of Simulations

Fig. 4 Convergence of mean fundamental frequency from different metamodel methods

Analysis of RC Buildings by Metamodel Approaches 827

obtained from this simulation are compared with the results obtained from MCS
as shown in Fig. 5. It can be seen that the metamodel-based approaches are able
to represent the random frequency responses fairly accurately. The mean, standard
deviation and number of FEA required in each method are summarized in Table 5.
It is observed that HDMR method could predict the random responses with least
number of computations.

(a) HDMR (b) CCD

(c) BBD

Fig. 5 Comparison of frequencies (Hz) predicted by metamodel methods with MCS

Table 5 Statistical parameters of natural frequency obtained from different methods

Methods Mean (Hz) Std. dev. (Hz) No. of samples required to create metamodel
HDMR 2.333 0.305 15
CCD 2.299 0.308 100
BBD 2.306 0.309 62
MCS (100,000) 2.306 0.311 −
828 D. Sahu et al.

Fig. 6 Comparison of frequencies (Hz) predicted by metamodel methods with MCS

It can be seen that the HDMR method performs better compared to other meta-
model methods as evident from the R2 values. Figure 6 shows the comparison of
probability distributions of output natural fundamental frequencies simulated by
various metamodel approaches along with the MCS method. The statistical difference
between the probability distributions of natural frequencies obtained by the various
methods is found to be negligible, and the frequencies are found to be normally
distributed.

5 Conclusions

Unpredictability exists in the responses of reinforced concrete structures due to the

uncertainties in geometry, material properties and loading. A stochastic analysis
integrating the uncertainties in the input properties is vital to study the random
dynamic responses. Non-statistical approaches based on metamodel concepts have
been reported as an alternative to computationally intensive statistical approaches in
previous studies. The present study explores the application of non-statistical HDMR
metamodel for the analysis of buildings. The effectiveness of non-statistical HDMR
is evaluated in the current study by comparing its results with those of the traditional
approaches such as response surface method using the design of experiments (DOE).
Noticeable conclusions noted from the case study of selected building are as follows:
• HDMR method is found to predict the random natural frequency responses (incor-
porating seven random variables) of four-storied RC building with least number of
computations (15 sampling points) compared to other metamodel DOE methods.
The error in the prediction is found to be less than 1%.
Analysis of RC Buildings by Metamodel Approaches 829

• Although the results are deduced based on the selected case studies, the
methodology is applicable and relevant to other structures also.

References

1. Xu J, Chen J, Li J (2012) Probability density evolution analysis of engineering structures via

cubature points. Comput Mech 50(1):135–156
2. Rubinstein RY, Kroese DP (1981) Simulation and the Monte Carlo method. Wiley, New York
3. Chakraborty S, Mandal B, Chowdhury R, Chakrabarti A (2016) Stochastic free vibration anal-
ysis of laminated composite plates using polynomial correlated function expansion. Compos
Struct 135:236–249
4. Mangalathu S (2017) Performance based grouping and fragility analysis of box-girder bridges
in California. Ph.D. Thesis, Georgia Tech
5. Bucher C, Bourgund U (1990) A fast and efficient response surface approach for structural
reliability problems. Struct Saf 7(1):57–66
6. Box G, Wilson KB (1951) On the experimental attainment of optimum conditions. J Royal
Stat Soc. Series B (Methodological) 13(1):1–45
7. Box GE, Behnken DW (1960) Some new three level designs for the study of quantitative
variables. Technometrics 2(4):455–475
8. Rabitz H, Mer FA, Shorter J, Shim K (1999) Efficient input output model representations.
Comput Phys Commun 117(12):11–20
9. Chowdhury R, Rao BN, Prasad AM (2009) High-dimensional model representation for
structural reliability analysis. Commun Numer Methods Eng 25(4):301–337
10. Towashiraporn P (2004) Building seismic fragilities using response surface metamodels. Ph.D.
thesis, Georgia Institute of Technology
11. Montgomery DC (2004) Design and analysis of experiments. Wiley, New York
12. Wu CF, Hamada M (2000) Experiments planning, analysis, and parameter design optimization.
Wiley, New York
13. Unnikrishnan VU, Prasad AM, Rao BN (2013) Development of fragility curves using high-
dimensional model representation. Earthq Eng Struct Dynam 42(3):419–430
14. IS 456 (2000) Indian standard plain and reinforced concrete–code of practice. Bureau of Indian
Standards, New Delhi
15. IS 1893 (2016) Indian standard criteria for earthquake resistant design of structures. Bureau of
Indian Standards, New Delhi, India
16. IS 13920 (2016) Indian standard ductile detailing of reinforced concrete structures subjected
to seismic forces. Bureau of Indian Standards, New Delhi, India
17. Mazzoni S, Mckenna F, Scott MH, Fenves GL (2006) The open system for earthquake
engineering simulation (OpenSEES) User Command-Language Manual.
18. Val D, Bljuger F, Yankelevsky D (1997) Reliability evaluation in nonlinear analysis of
reinforced concrete structures. Struct Saf 19(2):203–217
19. Mitropoulou CC, Lagaros ND, Papadrakakis M (2011) Life-cycle cost assessment of optimally
designed reinforced concrete buildings under seismic actions. Reliab Eng Syst Saf 96:1311–
1331
20. Ay BÖ (2008) Fragility based assessment of low-rise and mid-rise reinforced concrete frame
buildings in Turkey. M.S. Thesis, Middle East Technical University, Turkey
21. Bal IE, Crowley H, Pinho R (2008) Detail assessment of structural characteristics of Turkish
RC buildings stock for loss assessment models. Soil Dyn Earthq Eng 28:914–932