Engineering Structures 29 (2007) 2069–2079

www.elsevier.com/locate/engstruct

Neural network for bending moment in continuous composite beams

considering cracking and time effects in concrete
Umesh Pendharkar a , Sandeep Chaudhary b , A.K. Nagpal a,∗
a Department of Civil Engineering, Indian Institute of Technology, Delhi 110016, India
b Department of Structural Engineering, Malaviya National Institute of Technology, Jaipur, India

Received 31 May 2006; received in revised form 28 October 2006; accepted 14 November 2006
Available online 13 December 2006

Abstract

A methodology using a neural network model has been developed for the continuous composite beams to predict the inelastic moments
(typically for 20 years, considering instantaneous cracking, and time effects, i.e. creep and shrinkage, in concrete) from the elastic moments
(neglecting instantaneous cracking and time effects). It is shown that the redistribution of elastic moment at a support due to instantaneous cracking
along with time effects depends primarily on the instantaneous cracking at the support and adjacent supports and also that the redistribution is
independent of absolute span lengths. The proposed neural network model predicts the inelastic moment ratio (ratio of elastic moment to inelastic
moment) using eight input parameters. The training and testing data for the neural network is generated using a hybrid analytical–numerical
method of analysis. The models have been validated for four example beams and the errors are shown to be small. The methodology enables
rapid estimation of inelastic moments and requires a computational effort that is a fraction of that required for the time dependent analysis. The
methodology can be extended for the composite building frames resulting in huge savings in computational time.
c 2006 Elsevier Ltd. All rights reserved.

Keywords: Neural network; Creep; Shrinkage; Composite beam; Concrete cracking

1. Introduction the structural analysis of masonry infilled steel frames [4],

Redistribution of elastic moments in continuous composite
beams (Fig. 1(a)) occurs due to instantaneous cracking in
concrete in hogging moment regions, as well as time effects
(creep and shrinkage) in concrete. Methods are available in the
literature for analysis of the beams, which take into account
this moment redistribution. These methods are based either
on an incremental or iterative approach. Both the approaches
require a computational effort, which is many times more than
that required for the elastic analysis (neglecting instantaneous
cracking and time effects). The use of neural networks may be
made to drastically reduce the computational effort.
Principles and applications of neural networks in civil
engineering have been summarized in the works by Flood and
Kartam [1,2] and Adeli [3]. Neural network based models have
been recently applied in the field of structural engineering for
∗ Corresponding author. Tel.: +91 11 26591234; fax: +91 11 26581117.
E-mail address: aknagpal@civil.iitd.ernet.in (A.K. Nagpal).
prediction of shear strength of RC beams [5], prediction of
ultimate shear strength of deep beams [6], structural reanalysis
using iterative method [7], generation of artificial earthquake
and response spectra [8], diagnosis of structural damage [9],
damage assessment of cable stayed bridges [10], damage
detection of bridges [11,12], evaluation of ultimate strength of
steel panels [13], evaluation of existing bridges [14], optimum
design of cold-formed space structures [15], prediction of time
effects in reinforced concrete frames [16], establishing size
effect on shear strength of beams without shear stirrups [17],
shear strength design of beams without stirrups and with
stirrups [18,19] and estimation of concrete strength [20–
22]. These studies reveal the strength of neural networks
in predicting the solutions of different structural engineering
problems.
In this paper, a methodology using a neural network has
been developed to estimate the inelastic moments in continuous
composite beams, M i (considering the instantaneous cracking
and time effects in concrete) from the elastic moments, M e

c 2006 Elsevier Ltd. All rights reserved.

0141-0296/$ - see front matter
doi:10.1016/j.engstruct.2006.11.009
2070 U. Pendharkar et al. / Engineering Structures 29 (2007) 2069–2079

Nomenclature
E modulus of elasticity of concrete
I cr transformed moment of inertia of steel section
and reinforcement about top fibre
I un transformed moment of inertia of composite
section about top fibre
Me elastic moment
Fig. 1(a). Composite cross-section.
Mi inelastic moment
M cr cracking moment
S stiffness
l span length
n number of spans
w uniformly distributed load Fig. 1(b). Cracked span length beam element.
w cr cracking load
η percentage variation in elastic moment

subscript
j, k, n support or span number

(neglecting the instantaneous cracking and time effects in

concrete). M e can be obtained from any of the readily available
software. The methodology enables rapid estimation of M i and
requires a computational effort that is a fraction of that required
for the methods available in the literature.
The neural network models predict the M i (typically at
20 years) due to instantaneous cracking and time effects.
The neural network models have been validated for four
example beams. The errors are shown to be small for practical
purposes. The methodology can easily be extended for large
composite building frames where a very significant saving in
computational effort would result.
2. Analysis of continuous composite beams
For generalized and efficient neural networks, a huge
number of training data sets are required for which a
highly efficient method is desirable. Recently, a hybrid
analytical–numerical procedure has been developed [23] to take
into account the nonlinear effects of concrete cracking and
time-dependent effects of creep and shrinkage in composite
beams and frames. The procedure is analytical at the element
level and numerical at the structural level. A cracked span
length beam element consisting of uncracked zone in the middle
and cracked zones at the ends (Fig. 1(b)) has been used in the
procedure.
The analysis in the hybrid procedure is carried out in two
parts. In the first part, instantaneous analysis is carried out
using an iterative method. In the second part, time-dependent
analysis is carried out by dividing the time into a number of
time intervals to take into account the progressive nature of
cracking of concrete. In time analysis crack length is assumed
to be constant and equal to that at the beginning, as shown in
Fig. 1(c).
The Age-Adjusted Effective Modulus method is used to
account for the variation of the modulus of elasticity with
Fig. 1(c). Progressive nature of cracking.
the age of loading and creep [25]. Further, in the procedure
the variation of tensile strength with time has been taken into
account in accordance with the provisions of the CEB-FIP
model code 1990 [26].
The procedure has been validated by comparison with the
experimental results, analytical results, and FEM model using
ABAQUS. This procedure is adopted in the present work.
3. Significant extent of propagation of the effect of cracking
Cracking in continuous composite beams occurs in the
end portions (where hogging moment occurs) of spans at
intermediate supports when subjected to loading that causes
instantaneous cracking in end portions. This instantaneous
cracking may further progress due to time effects. The
elastic bending moments M e at an instantaneous state gets
redistributed to M i at an instantaneous state owing to cracking
and at a final state (typically 20 years) gets further redistributed
owing to time effects that may also result in further concrete
cracking.
The redistribution of the moments along the length of
a beam due to instantaneous cracking at a support reduces
along the distance from the support. A preliminary numerical
study is carried out to estimate the significant extent of
propagation of the effect of instantaneous cracking, at a support
at instantaneous and final stages. For the study, a typical multi-
span (number of spans = n) continuous composite beam,
shown in Fig. 2(a), is considered. The cross-sectional properties
throughout the beam are kept constant unless otherwise stated.
The nature of the elastic moment diagram for the beam with
 U. Pendharkar et al. / Engineering Structures 29 (2007) 2069–2079
Fig. 2. A typical multi-span continuous composite beam: (a) geometry and
loading; (b) elastic bending moment diagram (equal spans and constant
loading); (c) span 1 loaded heavily; (d) elastic and inelastic bending moment
diagrams (span 1 loaded heavily); (e) span 5 loaded heavily; (f) elastic and
inelastic bending moment diagrams (span 5 loaded heavily).
equal spans (l1 = l2 = l3 · · · = ln ) and the same load intensities
(w1 = w2 = · · · = w j · · · = wn−1 ) is shown in Fig. 2(b) with
the maximum moment occurring at the penultimate supports.
Let the loading be such that instantaneous cracking occurs
at joint j. Let the ratio of the cracking moment M cr to the
elastic moment at joint j, (neglecting cracking) M ej be R j
(cracking moment ratio); thus a smaller value of R j indicates
greater cracking. The change (redistribution) {Mkie } R j in elastic
moment Mke at any support k at an instantaneous or final stage
depends upon R j , and is expressed as
{Mkie } R j = {Mki − Mke } R j .
Further the change in {Mkie } R j , 1{Mkie } R j from its value at
R j = 1, is the effect of instantaneous cracking occurring at
joint j. Its normalized value ηk, j may be taken as a measure
of the extent of propagation of the effect of cracking. The
elastic moment at support j, at R j = 1, may be taken as the
normalizing factor. The expression for ηk, j may therefore be
written as
ηk, j = 1{Mkie } R j /{M ej } R j =1
in which,
1{Mkie } R j = {Mki − Mke } R j − {Mki − Mke } R j =1 .
2071
Fig. 3. Variation of ηk, j with R2 at different supports for cracking at
penultimate support (support 2).
It may be noted that the second term in Eq. (3) is equal to
zero at the instantaneous state.
For the preliminary numerical study, a beam with n = 7,
l j ( j = 1 to n) = 4.0 m and M cr = 25.5 kN m is considered.
Two cases are considered to identify the significant extent
of propagation of the effect of instantaneous cracking at an
instantaneous state and final state. In the first case, the beam
is made to crack at an instantaneous state over a penultimate
support (support 2, j = 2), whereas in the second case cracking
at an instantaneous state is made to occur simultaneously over
intermediate supports (support 4 and 5 j = 4, 5) of 4th span.
For both the cases, initially the loading on the spans is kept
equal to the cracking load, w cr (the load at which the moment,
M e at any section of a beam just becomes equal to M cr ).
Consider the first case. The load on span 1 (Fig. 2(c)) is
increased, keeping the loads on other spans constant, such that
the instantaneous cracking takes place at support 2 (Fig. 2(d)).
The redistribution varies with the cracking moment ratio R2
(cracking moment ratio at support 2). Seven values of R2 ,
1.0, 0.90, 0.75, 0.60, 0.50, 0.40 and 0.25, in the practical
range resulting from the increase in w1 are considered. Values
of R2 equal to 1.0 and 0.25 indicate initiation of cracking
and maximum cracking respectively. The variations ηk,2 (k =
2, 3, 4, 5), at instantaneous state and at final state, are shown in
Fig. 3. It is seen that the variations at the two states are nearly
the same. Further, only the variations of η2,2 and η3,2 may be
considered to be significant.
Now consider the second case in which instantaneous
cracking is made to occur simultaneously at intermediate
(1)
supports (supports 4 and 5, j = 4, 5). For this purpose w4
is increased (Fig. 2(e)). The natures of M e and M i diagrams
for this case are shown in Fig. 2(f). Again, seven values of
R4 (R5 being equal to R4 for the case considered), as for R2 ,
are considered. The variations of η4,4 (=η5,5 ) (at the cracked
supports), of η3,4 , η6,5 (at first supports next to the cracked
supports), and of η2,4 , η7,5 (at second supports next to the
cracked supports), at the two stages are shown in Fig. 4. It is
again seen that the variations at the two stages are nearly the
(2) same. Also only, the variations ηk,4 (k = 4, 3) and ηk,5 (k =
5, 6) may be considered to be significant.
Therefore, it can be stated that, for both the cases the
(3) significant extent of propagation of the effect of instantaneous
2072 U. Pendharkar et al. / Engineering Structures 29 (2007) 2069–2079
Fig. 4. Variation of ηk, j with R4 (=R5 ) at different support for cracking at an
internal support (support 4).
cracking is up to the support next to the instantaneously cracked
support on either side of the cracked support.
Similar numerical studies are carried out for beams with
different l j , n and M cr . From these studies also (not reported),
it is again observed that the extent of propagation of the effect
of instantaneous cracking and time effects is significant up to
the first supports next to the instantaneously cracked supports.
It therefore follows that in order to establish redistribution
of moment at a support (support j) with sufficient accuracy,
cracking at the support and adjacent supports (support j + 1
and support j − 1) only needs to be considered.
4. Probable structural parameters and sensitivity analysis
The moment at any cross-section along the length of a
span can be obtained from the support moments and loading.
Inelastic moment Mki , at a support k, may be obtained from
the inelastic moment ratio, Mke /Mki , at the support and elastic
moment Mke . The ratio Mke /Mki would be the output parameter
for the neural network. Therefore, a sensitivity analysis is
carried out to determine the significant structural parameters
governing Mke /Mki . As stated subsequently, the sensitivity
analysis also yields the number of training data sets.
For the sensitivity analysis, two span (n = 2), three span
(n = 3) and seven span (n = 7) beams are considered. It may
be noted that the minimum number of spans in a continuous
composite beam in which cracking occurs is two. Further, the
seven span beams may represent the beams in which the number
of spans is greater than three. Therefore, these sets of beams
may be considered to represent continuous composite beams
with any number of spans.
As has been observed in the previous section, Mke /Mki at
a support, k is significantly affected by the cracking at the
support and adjacent supports (support k −1 and support k +1).
Accordingly, the probable structural parameters, which may
influence Mke /Mki at a support k of a beam, with the same cross-
section throughout the length, are listed below:
1. Age of loading, t0 .
2. Stiffness ratio of adjacent spans, Sk−1 /Sk (Sk = E I un /lk ,
where E = modulus of elasticity of concrete, and
I un = transformed moment of inertia of composite section
about top fibre, the reference axis).
Table 1
Practical range of probable structural parameters
Parameter t0 Sk−1 /Sk Rk wk−1 /wk I cr /I un Rk−1 Rk+1 Gr
7 days 0.25 0.25 0.25 0.38 0.25 0.25 20
Range – – – – – – – –
21 days 4.0 4.0 4.0 0.54 4.0 4.0 40
3. Cracking moment ratio at the support, Rk .
4. Load ratio of the adjacent spans, wk−1 /wk .
5. Composite inertia ratio, I cr /I un (I cr = transformed
moment of inertia of steel section and reinforcement about
top fibre, the reference axis).
6. Cracking moment ratio at left adjacent support, Rk−1 .
7. Cracking moment ratio at right adjacent support, Rk+1 .
8. Grade of concrete, Gr.
For a two span beam, the parameters Rk−1 and Rk+1 are
absent.
The practical range of probable structural parameters is
given in Table 1. Sensitive studies are reported for the left
penultimate support (support 2) of two span, three span and
seven span beams and for a typical internal support (support
4) of a seven span beam. In these studies, only one parameter
is varied at a time, keeping the other parameters constant,
either equal to the minimum or the maximum values of the
parameters. These sets are designated as Minimum value set
and Maximum value set respectively. It may be further noted
that since the cross section of beams is the same throughout the
length, the required stiffness ratios Sk−1 /Sk for the studies are
achieved by varying the length of spans.
4.1. Penultimate support
The variation of Mke /Mki with the probable structural
parameters for support 2 (k = 2) of two span, three span
and seven span beams is studied and structural parameters are
identified. The lengths l j ( j = 2 to n) are initially taken as
8.0 m. The span length l1 is varied to achieve the specific
stiffness ratios.
It may be noted that out of the eight probable structural
parameters [t0 , S1 /S2 , R2 , w1 /w2 , I cr /I un , R1 , R3 , Gr], the
cracking moment ratio R1 at the left adjacent support need not
be considered for two span, three span and seven span beams
as the moment is zero at support 1. Similarly, cracking moment
ratio R3 at the right adjacent support need not to be considered
for two span beams, as the moment at support 3 is zero.
The sensitivity analysis is carried out for each of the
structural parameters, in turn.
4.1.1. Age of loading (t0 )
The redistribution of M e at support 2 may vary with the age
of loading t0 ; hence t0 is considered as a probable structural
parameter. For two span beam the values of the other structural
parameters for the two sets considered are: Minimum value set:
S1 /S2 = 0.25, R2 = 0.25, w1 /w2 = 0.25, I cr /I un = 0.38,
Gr = M20; and Maximum value set: S1 /S2 = 4.0, R2 = 4.0,
w1 /w2 = 4.00, I cr /I un = 0.54, Gr = M40. The two sets
 U. Pendharkar et al. / Engineering Structures 29 (2007) 2069–2079
Fig. 5. Variation of M2e /M2i with t0 for two, three and seven span beams.
for three span and seven span beams are: Minimum value set:
S1 /S2 = 0.25, R2 = 0.25, w1 /w2 = 0.25, I cr /I un = 0.38,
Gr = M20 R3 = 0.25; and Maximum value set: S1 /S2 = 4.0,
R2 = 4.0, w1 /w2 = 4.00, I cr /I un = 0.54, Gr = M40
R3 = 4.0. It may be noted that R2 = 0.25 in the Minimum
value set represents maximum cracking; and R2 = 4.0 in
the Maximum value set represents absence of cracking. The
variation for the above two sets are shown in Fig. 5. It is
observed that the natures of variations for the three beams
for the two sets are similar. These can be represented fairly
accurately by considering two values of t0 (7 and 21 days),
designated as sampling points.
4.1.2. Effect of stiffness ratio (S1 /S2 )
The redistribution of elastic moment M2e at the support
(support 2) may vary with the relative stiffness of the spans;
hence stiffness ratio S1 /S2 has been considered a probable
structural parameter. For two span beams, the Minimum values
set and Maximum values set chosen are: t0 = 7 days R2 =
0.25, w1 /w2 = 0.25, I cr /I un = 0.38, Gr = M20, and
t0 = 21 days R2 = 4.0, w1 /w2 = 4.00, I cr /I un = 0.54,
Gr = M40, respectively. For three span and seven span beams
Minimum value sets and Maximum value sets are: t0 = 7 days,
R2 = 0.25, w1 /w2 = 0.25, I cr /I un = 0.38, R3 = 0.25,
Gr = M20, and t0 = 21 days R2 = 4.0, w1 /w2 = 4.00,
I cr /I un = 0.54, R3 = 4.0, Gr = M40, respectively. The
variations for all the beams are shown in Fig. 6. The variations
are seen to be mild for two span beams. Five values of S1 /S2
(0.25, 0.5, 1.0, 1.5 and 4.0) represents fairly accurately all the
variations.
4.1.3. Effect of cracking moment ratio R2
Since redistribution of M e at a support would vary with
R2 (R2 < 1), this is selected as a probable structural parameter.
For two span beam, Minimum value set and Maximum value
set chosen are: t0 = 7 days, S1 /S2 = 0.25, w1 /w2 = 0.25,
I cr /I un = 0.38, Gr = M20, and t0 = 21 days, S1 /S2 = 4.0,
w1 /w2 = 4.0, I cr /I un = 0.54, Gr = M40, respectively. For
three span beam and seven span beam the Minimum value set
and the Maximum value set are: t0 = 7 days, S1 /S2 = 0.25,
w1 /w2 = 0.25, I cr /I un = 0.38, R3 = 0.25, Gr = M20, and
t0 = 21 days, S1 /S2 = 4.0, w1 /w2 = 4.0, I cr /I un = 0.54,
2073
Fig. 6. Variation of M2e /M2i with S1 /S2 for two, three and seven span beams.
Fig. 7. Variation of M2e /M2i with R2 for two, three and seven span beams.
R3 = 4.0, Gr = M40, respectively. The variations are shown
in Fig. 7, and all are seen to be significant. The nature of
variations for all the beams is similar and can be represented
fairly accurately by five sampling points (R2 = 0.25, 0.45, 0.80,
1.0 and 4.0).
4.1.4. Effect of load ratio (w1 /w2 )
Ratio of loading on the adjacent spans of a support may
affect the redistribution of M e , therefore, w1 /w2 is also
considered as a probable structural parameter. For two span
beam, the Minimum value set and the Maximum value set are:
t0 = 7 days, S1 /S2 = 0.25, R2 = 0.25, I cr /I un = 0.38,
Gr = M20, and t0 = 21 days, S1 /S2 = 4.0, R2 = 4.00,
I cr /I un = 0.54, Gr = M40, respectively. For three span
beam and seven span beam the Minimum value set and the
Maximum value set are: t0 = 7 days, S1 /S2 = 0.25, R2 = 0.25,
I cr /I un = 0.38, R3 = 0.25, Gr = M20, and t0 = 21 days,
S1 /S2 = 4.0, R2 = 4.00, I cr /I un = 0.54, R3 = 4.0, Gr =
M40, respectively. The variations are shown in Fig. 8. Little
variation is observed for the Maximum value sets, whereas
mild variations are observed for the Minimum value sets. Two
sampling points (w1 /w2 = 0.25, 4.0) are sufficient to represent
the variations for all the beams.
4.1.5. Effect of composite inertia ratio (I cr /I un )
The stiffness of a composite section, I un , reduces to that of a
steel section, I cr , on cracking, which leads to redistribution of
2074 U. Pendharkar et al. / Engineering Structures 29 (2007) 2069–2079
Fig. 8. Variation of M2e /M2i with w1 /w2 for two, three and seven span beams.
Fig. 9. Variation of M2e /M2i with I cr /I un for two, three and seven span beams.
forces. Therefore, I cr /I un is considered as a probable structural
parameter. For two span beam, the Minimum value set and the
Maximum value set chosen are: t0 = 7 days, S1 /S2 = 0.25,
R2 = 0.25, w1 /w2 = 0.25, Gr = M20, and t0 = 21 days,
S1 /S2 = 4.0, R2 = 4.00, w1 /w2 = 4.00, Gr = M40,
respectively. For three span beam and seven span beam the
Minimum value set and the Maximum value set are: t0 = 7
days, S1 /S2 = 0.25, R2 = 0.25, w1 /w2 = 0.25, R3 = 0.25,
Gr = M20, and t0 = 21 days, S1 /S2 = 4.0, R2 = 4.00,
w1 /w2 = 4.00, R3 = 4.0, Gr = M40, respectively. The
variations are shown in Fig. 9. The variations for all the beams
can be represented fairly accurately by four sampling points
(I cr /I un = 0.380, 0.425, 0.446 and 0.540).
4.1.6. Effect of cracking moment ratio at right adjacent support
R3
As seen earlier (Figs. 3 and 4), the redistribution of elastic
moment M e at a support depends on cracking at the adjacent
supports also, therefore R3 is selected as a probable parameter
for variation of M2e /M2i for three and seven span beams. For
three span beam and seven span beam the Minimum value set
and Maximum value set are: t0 = 7 days, S1 /S2 = 0.25,
R2 = 0.25, w1 /w2 = 0.25, I cr /I un = 0.38, Gr = M20,
and t0 = 21 days, S1 /S2 = 4.0, R2 = 4.00, w1 /w2 = 4.00,
I cr /I un = 0.54, Gr = M40, respectively. The variations are
shown in Fig. 10. The nature of variations are the same for
three and seven span beams and further these tend to assume
Fig. 10. Variation of M2e /M2i with R3 for two, three and seven span beams.
Fig. 11. Variation of M2e /M2i with Gr for two, three and seven span beams.
a constant value for R3 greater than 1. Five sampling points
(R3 = 0.25, 0.4, 0.55, 1.0 and 4.0) are considered to be
sufficient to represent fairly accurately these variations.
4.1.7. Effect of grade of concrete (Gr)
The redistribution of M e due to cracking and time effects
depends upon the grade of concrete Gr, and, therefore Gr
is selected as a probable parameter. For two span beam, the
Minimum value set and the Maximum value set chosen are:
t0 = 7 days, S1 /S2 = 0.25, R2 = 0.25, w1 /w2 = 0.25,
I cr /I un = 0.38, and t0 = 21 days, S1 /S2 = 4.0, R2 = 4.00,
w1 /w2 = 4.00, I cr /I un = 0.54, respectively. For three span
beam and seven span beam the Minimum value set and the
Maximum value set are: t0 = 7 days, S1 /S2 = 0.25, R2 = 0.25,
w1 /w2 = 0.25, R3 = 0.25, I cr /I un = 0.38, and t0 = 21 days,
S1 /S2 = 4.0, R2 = 4.00, w1 /w2 = 4.00, R3 = 4.0, I cr /I un =
0.54, respectively. The variations are shown in Fig. 11, and are
seen to be mild. Two sampling points (Gr = M20 and M40) are
considered to be sufficient to represent fairly accurately these
variations.
Similar studies for each of the above eight parameters
(t0 , S j−1 /S j , M cr /M ej , w j−1 /w j , I cr /I un , M cr /M ej−1 ,
M cr /M ej+1 , Gr) are carried out for l j ( j = 2 to n) = 8.0
m and 12.0 m also. Almost the same variations as shown in
Figs. 5–11 are obtained. Therefore the absolute span length is
not a parameter governing the redistribution of moments at the
penultimate support.
 U. Pendharkar et al. / Engineering Structures 29 (2007) 2069–2079 2075

Fig. 12. Variation of M4e /M4i with t0 for two, three and seven span beams. Fig. 14. Variation of M4e /M4i with R4 for seven span beam.

Fig. 15. Variation of M4e /M4i with w3 /w4 for seven span beam.
Fig. 13. Variation of M4e /M4i with S3 /S4 for seven span beam.

Further, the nature of the variations for the penultimate

support on the other side of the beams, i.e. support n for n
span beams (n = 3, 7) would be similar to that observed for
support 2 of three span and seven span beams when loaded in a
corresponding manner.

4.2. Internal support

Similar studies, as performed for sensitivity analysis of

support 2, are carried out for the sensitivity analysis of a typical
internal support (support 4) of a seven span beam for the two
extreme value sets of the probable structural parameters, t0 ,
S3 /S4 , M cr /M4e , w3 /w4 , I cr /I un , M cr /M3e , M cr /M5e , Gr. The
span lengths l j ( j = 1 to 3 and 5 to 6) are initially taken equal
to 8.0 m. The span length l4 is varied to achieve the specified
stiffness ratios. Corresponding variations obtained are shown in
Figs. 12–18.
It is observed that the nature of variations for the internal
support, shown in Figs. 12–18, are similar to the corresponding
variations for the penultimate support, shown in Figs. 5–11,
though the numerical values are different. Therefore, the same
structural parameters and sampling points are applicable for
internal supports and penultimate supports.
Again, similar studies are carried out by taking span lengths
l j ( j = 1 to 3 and 5 to 6) = 4.0 m and 12.0 m in turn, and
similar results are obtained irrespective of the span lengths, as
has been observed earlier for the penultimate support.
Fig. 16. Variation of M4e /M4i with I cr /I un for seven span beam.
4.3. Sensitivity analysis summary
It is observed from the sensitivity studies carried out for a
penultimate support and an internal support (Figs. 5–18) that
the variations for two span beam are different from three span
and seven span beams. The variation for three span beam differs
marginally from seven span beam. From this observed marginal
difference it may be postulated that a seven span beam may
represent all the beams having span more than three. Therefore,
two span, three span and seven span beams may be considered
to represent continuous composite beams with any number of
spans. It has also been inferred above that Mke /Mki for a support
is independent of the absolute span lengths.
2076 U. Pendharkar et al. / Engineering Structures 29 (2007) 2069–2079

Fig. 17. Variation of M4e /M4i with R3 , (R5 ) for seven span beam.

Fig. 19. Neural network model.

I cr /I un , Gr, and the output is M2e /M2i . For penultimate support

of three span beam, seven span beam, and internal support
of a seven span beam, the input for a support k, consists of
eight parameters, t0 , Sk−1 /Sk , M cr /Mke , wk−1 /wk , I cr /I un ,
M cr /Mk−1 e , M cr /M e
k+1 and Gr, and the output is Mk /Mk . One
e i

hidden layer is chosen for each network in which the number of

neurons is decided in the learning process by trial and error.
Fig. 18. Variation of M4e /M4i with Gr for seven span beam.
6. Training of neural network
Table 2
Sampling points for generation of training data sets Since the training of the neural network is an essential
Parameter Sampling points step in its performance, a sufficiently large database should be
generated for training and testing. The performance in terms
t0 7 days, 21 days
Sk−1 /Sk 0.25, 0.5, 1.0, 1.5, 4.0
of generalization and prediction qualities of a neural network
Rk 0.25, 0.45, 0.80, 1.0, 4.0 depends significantly on the training and testing data and the
wk−1 /wk 0.25, 4.0 domain this data covers [16].
I cr /I un 0.380, 0.425, 0.446, 0.540 The data sets are generated only for two span, three span and
Rk−1 0.25, 0.4, 0.55, 1.0, 4.0
seven span beams since these have been postulated to represent
Rk+1 0.25, 0.4, 0.55, 1.0, 4.0
Gr 20, 40 all the beams. Table 3 gives the number of sampling points
for the structural parameters and the resulting number of data
sets for two span, three span and seven span beams. The total
The sampling points that are common to penultimate
number of data sets required to be generated for the beams
supports and internal supports have been identified above and
are 36 800. It comprises of 800 data sets for the penultimate
are summarized in Table 2.
supports of two span beam, 4000 data sets for each of the two
5. Configuration of neural networks penultimate supports of a three span and seven span beams, and
20 000 data sets for the internal support (support 4) of a seven
The neural network model chosen in the present study is span beam. Three neural networks, one each for penultimate
a multilayered feed-forward network with neurons in all the end support of two span beam, three span beam and seven span
layers fully connected in a feed-forward manner (Fig. 19). beam and designated as NETM21, NETM31 and NETM71
Sigmoid function is used as an activation function and the back respectively and one network for internal support of a seven
propagation-learning algorithm is used for training. The back span beam, designated as NETM72, are trained, with the total
propagation algorithm has been used successfully for many number of training data sets being 800, 8000, 8000 and 20 000
civil engineering applications and is considered as one of the respectively. In the network designation the first number (2, 3,
most efficient algorithms for engineering applications [8]. 7) indicates the number of spans in the beam, and the second
For the penultimate support of two span beam, the input number (1, 2) represents penultimate end support and internal
consists of six parameters, t0 , S1 /S2 , M cr /M2e , w1 /w2 , support respectively.
 U. Pendharkar et al. / Engineering Structures 29 (2007) 2069–2079 2077

Table 3
Total number of data sets considered for training and testing

Beam Support Number of sampling points for the structural parameters Number of data sets
No. Type t0 Sk−1 /Sk Rk wk−1 /wk I cr /I un Rk−1 Rk+1 Gr
Two span 1 Penultimate 2 5 5 2 4 – – 2 800 (=2 × 5 × 5 × 2 × 4 × 2)
Three span 2 Penultimate 2 5 5 2 4 – 5 2 4 000 (=2 × 5 × 5 × 2 × 4 × 5 × 2)
3 2 5 5 2 4 5 – 2 4 000 (=2 × 5 × 5 × 2 × 4 × 5 × 2)
Seven span 2 Penultimate 2 5 5 2 4 – 5 2 4 000 (=2 × 5 × 52 × 4 × 5 × 2)
4 Internal 2 5 5 2 4 5 5 2 20 000 (=2 × 5 × 5 × 2 × 4 × 5 × 2)
7 Penultimate 2 5 5 2 4 5 – 2 4 000 (=2 × 5 × 5 × 2 × 4 × 5 × 2)

Table 4 carried out. Two thirds of the data sets are used for training
Configuration of networks, mean square errors and number of epochs as training patterns whereas one third of the data sets are used
Network Number of neurons in MSE MSE Epochs for testing. For this partitioning, the ‘hold out method’ [27],
layers training testing in which partitioning is done randomly, has been adopted. The
Input Hidden Output configurations of four optimum networks with mean square
NETM21 6 14 1 0.00004 0.00047 60 000 error and number of epochs are given in Table 4.
NETM31 8 15 1 0.00075 0.00175 60 000
NETM71 8 15 1 0.00060 0.00126 45 000
NETM72 8 15 1 0.00076 0.00136 70 000
7. Validation of neural networks

A two span continuous composite validation beam (Fig. 20)

Fig. 20. Validation beam.
Further it may be noted that it may not be possible to
generate all the data sets with the practical loading in which the
values of the sampling points of Rk−1 , Rk , Rk+1 , in the lowest
and the highest portion of the range (0.25, 4.0) are achieved
simultaneously. For example with the practical loading it is not
possible to achieve the values of Rk−1 = 0.25, Rk = 4.0, and
Rk+1 = 0.25, simultaneously. For NETM21, however, all the
required data sets can be generated. For NETM31, NETM71
and NETM72, 6698, 7424 and 16 800 data sets have been
generated, respectively.
In order to bring all the input parameters and output
parameters in the range 0.0–1.0, the input as well output data
are multiplied by the normalization factors. For NETM31 and
NETM71, the penultimate supports have been modeled by
assigning a high value of 20 for the parameters R1 and Rn .
The training is carried out using the Stuttgart Neural
Network Simulator (SNNS) [24]. For training, several trials
with different numbers of neurons in the hidden layer are
the analytical results of which are available, Gilbert [28], is
considered. The network predicts inelastic moment at support
2 as 59.9 kN m, against 62.8 kN m reported by Gilbert. The
two values are quite close for practical purposes.
Trained neural networks are validated with a number of
beams with a wide variation of input parameters. Four example
beams with n = 2, 3, 5 and 15 are considered (Fig. 21(a)–(d))
with age of loading as 10, 14, 14, 10 days respectively; cracking
moment as 34.05 kN m, 33.22 kN m, 33.22 kN m, 33.33 kN m
respectively; and moment of inertia of steel beam about its own
axis as 40.04 × 10−6 m4 , 85.00 × 10−6 m4 , 40.04 × 10−6 m4 ,
40.04 × 10−6 m4 , respectively.
The input parameters are shown in Table 5. None of the
input parameters have been considered in the training or
testing. Inelastic moments (20 years) obtained from the hybrid
analytical–numerical procedure and the neural networks are
reported in Table 6, along with elastic moments. The root mean
square percentage errors in prediction of inelastic moments are
1.55%, 0.54%, 0.75%, and 0.99% for two span, three span, five
span and fifteen span example beams respectively. The error at
the supports having the maximum percentage redistribution of
moments is 0.24% for three span beam (at support 2), 2.39%
for five span beam (at support 3), and 3.7% for fifteen span
beams (at support 9). The percentage errors are generally small

Fig. 21. Example beams. (a) Two span; (b) three span; (c) five span; (d) fifteen span.
2078 U. Pendharkar et al. / Engineering Structures 29 (2007) 2069–2079

Table 5
Input parameters in example beams

Beam Network used Support no. t0 Sk−1 /Sk wk−1 /wk Rk−1 Rk Rk+1 I s /I co Gr
Two span NETM21 2 0.4545 0.3241 0.5722 – 0.1159 – 0.401 0.675
Three span NETM31 2 0.6364 0.1796 0.0922 0.9524a 0.1929 0.0459 0.4521 0.7805
3 0.6364 0.1049 0.1096 0.0372 0.2380 0.9524a 0.4521 0.7805
Five span NETM71 2 0.6364 0.2046 0.2988 0.9524a 0.1495 0.0761 0.4015 0.6585
NETM72 3 0.6364 0.2061 0.1755 0.1477 0.3898 0.2976 0.4015 0.6585
4 0.6364 0.2588 0.2873 0.3898 0.2976 0.2456 0.4015 0.6585
NETM71 5 0.6364 0.2374 0.2276 0.0581 0.2486 0.9524a 0.4015 0.6585
Fifteen span NETM71 2 0.4545 0.2778 0.1560 0.9524a 0.1356 0.0323 0.4015 0.6585
NETM72 3 0.4545 0.1355 0.1655 0.1339 0.1652 0.1619 0.4015 0.6585
4 0.4545 0.4390 0.3712 0.1652 0.1619 0.1561 0.4015 0.6585
5 0.4545 0.1518 0.1687 0.1619 0.1561 0.1541 0.4015 0.6585
6 0.4545 0.3920 0.3526 0.1561 0.1541 0.1679 0.4015 0.6585
7 0.4545 0.1355 0.1420 0.1541 0.1679 0.1172 0.4015 0.6585
8 0.4545 0.4634 0.3441 0.1679 0.1172 0.1190 0.4015 0.6585
9 0.4545 0.1386 0.1979 0.1172 0.1190 0.1706 0.4015 0.6585
10 0.4545 0.4246 0.4207 0.1190 0.1706 0.1770 0.4015 0.6585
11 0.4545 0.2076 0.3252 0.1706 0.1770 0.1667 0.4015 0.6585
12 0.4545 0.2744 0.1540 0.1770 0.1667 0.1521 0.4015 0.6585
13 0.4545 0.1355 0.1655 0.1667 0.1521 0.1697 0.4015 0.6585
14 0.4545 0.4390 0.3712 0.1521 0.1697 0.1430 0.4015 0.6585
NETM71 15 0.4545 0.1536 0.2442 0.0331 0.1448 0.9524a 0.4015 0.6585
a Boundary condition.

Table 6 is 20. For the use of neural networks only one analysis is
Elastic and inelastic moments in the example beams required; therefore, the computational time required in the
Beam Support Elastic moment Long-term inelastic neural network is, say, about one twentieth of that required in
no. (kN m) moment (kN m) the hybrid procedure.
Actual Network
Two span 2 71.66 66.94 67.98 8. Conclusions
Three span 2 42.52 58.31 58.45
3 34.46 47.45 46.95 1. A methodology has been presented for predicting inelastic
Five span 2 54.87 67.02 66.17 bending moments (considering concrete cracking and time
3 20.79 34.92 35.79 effects in concrete) of a continuous composite beam from
4 27.24 39.66 39.72 elastic bending moments (neglecting cracking) by using
5 32.99 50.39 50.96
neural networks. The methodology is based on the following
Fifteen span 2 60.52 70.79 70.77 findings:
3 49.07 50.09 49.63
a. Redistribution of elastic moments at a support of a
4 50.06 60.66 57.52
5 51.90 62.12 64.51 continuous composite beam due to cracking of concrete
6 52.59 43.38 65.84 and time effects can be obtained with sufficient accuracy
7 48.26 46.73 44.13 if cracking at the support and only adjacent supports is
8 69.17 75.04 71.34 considered.
9 68.13 95.89 99.46
10 47.50 58.87 59.07
b. The structural parameters that govern Mke /Mki at a support
11 45.79 57.28 53.53 k are identified as t0 , Sk−1 /Sk , Rk , wk−1 /wk , I cr /I un ,
12 48.63 56.50 54.93 Rk−1 , Rk+1 and Gr.
13 53.30 68.17 69.73 c. Absolute span length is not a parameter governing the
14 47.77 46.91 47.06
inelastic moment ratio Mke /Mki .
15 56.68 67.33 67.38
2. Using the methodology, four networks have been developed.
The networks for seven span beam, NETM72 and NETM71
for practical purposes particularly for the supports where there can be used to predict redistribution of elastic moment for
is higher redistribution of moments. beams having any number of spans greater than three.
In the hybrid procedure for beams with sufficient cracking, 3. The four networks can predict inelastic moments with
generally 10–12 iterations are needed for convergence in the reasonable accuracy from the elastic moments within a small
first part (instantaneous analysis). In the second part time fraction of the time required for the time dependent analysis.
analysis is carried out at, say, eight time instances. Therefore, The networks have been validated for different beams with
for both parts on average the total number of analyses required the maximum root mean square percentage error found to be
 U. Pendharkar et al. / Engineering Structures 29 (2007) 2069–2079
1.55%, and the maximum percentage error at a support with
highest redistribution of moments is 2.5%.
This application of the methodology to the composite
beams is a step towards rapid estimation of long-term inelastic
moments in large composite building frames where a very
large computational effort is required in the iterative methods
available in the literature.
References
[1] Flood I, Kartam N. Neural networks in civil engineering I: Principles and
understanding. Journal of Computing in Civil Engineering, ASCE 1994;
8(2):131–48.
[2] Flood I, Kartam N. Neural networks in civil engineering II: Systems and
application. Journal of Computing in Civil Engineering, ASCE 1994;8(2):
149–62.
[3] Adeli H. Neural networks in civil engineering, 1989–2000. Computer-
Aided Civil and Infrastructure Engineering 2001;16(2):126–47.
[4] Subramanian K, Mini KM, Josephine KF. Neural network-based
modeling of infilled steel frames. Structural Engineering and Mechanics
2005;21(5):495–506.
[5] Mansour MY, Dicleli M, Lee JY, Zhang J. Predicting the shear strength
of reinforced concrete beams using artificial neural networks. Engineering
Structures 2004;26:781–99.
[6] Sanad A, Saka MP. Prediction of ultimate shear strength of reinforced
concrete deep beams using neural networks. Journal of Structural
Engineering ASCE 2001;127(7):818–28.
[7] Jenkins WM. Structural Reanalysis using a neural network-based iterative
method. Journal of Structural Engineering ASCE 2002;128(7):946–50.
[8] Lee SC, Han SW. Neural-network-based models for generating artificial
earthquakes and response spectra. Computers and Structures 2002;80:
1627–38.
[9] Tsai CH, Hsu DS. Diagnosis of reinforced concrete structural damage
base on displacement time history using the back-propagation neural
network technique. Journal of Computing in Civil Engineering, ASCE
2002;6(1):49–58.
[10] Cho HN, Cho YM, Lee SC, Hur CK. Damage assessment of cable stayed
bridges using probabilistic neural network. Structural Engineering and
Mechanics 2004;17(3):483–92.
[11] Lee JJ, Lee JW, Yi JH, Yun CB, Jung HY. Neural network-based damage
detection for bridges considering errors in baseline finite element models.
Journal of Sound and Vibration 2005;280:555–78.
[12] Yeung WT, Smith JW. Damage detection in bridges using neural networks
for pattern recognition of vibration signatures. Engineering Structures
2005;27:685–98.
2079
[13] Pu Y, Mesbahi E. Application of artificial neural network to evaluation of
ultimate strength of steel panels. Engineering Structures 2006;28:1190–6.
[14] Augusto VM, Karen CC. Evaluation of existing bridge structures using
neural networks. Structural Engineering and Mechanics 2002;13(2):
187–209.
[15] Tashakori A, Adeli H. Optimum design of cold-formed steel space
structures using neural dynamic models. Journal of Constructional Steel
Research 2002;58:1545–66.
[16] Maru S, Nagpal AK. Neural network for creep and shrinkage deflections
in reinforced concrete frames. Journal of Computing in Civil Engineering,
ASCE 2004;18(4):350–9.
[17] Oreta AWC. Simulating size effect on shear strength design of RC beams
without stirrups using neural networks. Engineering Structures 2004;26:
681–91.
[18] Cladera A, Mari AR. Shear design procedure for reinforced normal and
high-strength concrete beams using artificial neural networks. Part I:
Beams without stirrups. Engineering Structures 2004;26:917–26.
[19] Cladera A, Mari AR. Shear design procedure for reinforced normal and
high-strength concrete beams using artificial neural networks. Part II:
Beams with stirrups. Engineering Structures 2004;26:927–36.
[20] Lee SC. Prediction of concrete strength using artificial neural networks.
Engineering Structures 2003;25:849–57.
[21] Kim DK, Lee JJ, Lee JH, Chang SK. Application of probabilistic neural
networks for prediction of concrete strength. Journal of Materials in Civil
Engineering, ASCE 2005;17(3):353–62.
[22] Kim JI, Kim DK, Feng MQ, Yazdani F. Application of neural networks for
estimation of concrete strength. Journal of Materials in Civil Engineering,
ASCE 2004;16(3):257–64.
[23] Chaudhary S, Pendharkar U, Nagpal AK. A hybrid procedure for cracking
and time-dependent effects in composite frames at service load. Journal of
Structural Engineering ASCE 2007;133(2).
[24] Sttutgart Neural Network Simulator (SNNS) user manual. Version 4.2.
University of Sttutgart: Institute For Parallel and Distributed High
Performance Systems (IPVR),
http://www-ra.informatik.uni-tuebingen.de/SNNS/, 1998.
[25] Gilbert RI, Bradford MA. Time-dependent behavior of composite beams
at service loads. Journal of Structural Engineering ASCE 1995;121(2):
319–27.
[26] Comite’ Euro International du Beton-Fe’de’ration International de la
Pre’contrainte, (CEB-FIP). Model code for concrete structures. London:
Thomas Telford; 1993.
[27] Reich Y, Barai SV. Evaluating machine learning models for engineering
problems. Artificial Intelligence in Engineering 1999;13:257–72.
[28] Gilbert RI, Bradford MA. Time-dependent behavior of composite beams
at service loads. Journal of Structural Engineering ASCE 1995;121(2):
319–27.