Journal of Civil Engineering (IEB), 35 (2) (2007) 119-128

Prediction of short and long-term deflections of

two-way edge-supported reinforced concrete slabs
using artificial neural network
Tahsin Reza Hossain, Md. Taslim Uddin1 and Md. Saiful Alam Siddiquee
Department of Civil Engineering
Bangladesh University of Engineering and Technology, Dhaka 1000, Bangladesh
Received 10 July 2007

_____________________________________________________________________
Abstract
For designing a structure, members are so proportioned that will have adequate strength against
failure and at the same time must possess sufficient stiffness to ensure serviceability. ACI Code
provides minimum thickness for two-way edge-supported slabs so that the deflections are not
excessive. It also allows thinner slabs if calculated deflections are found tolerable. The method
given in ACI Code for deflection estimation is relatively straightforward in comparison to other
codes. It uses Branson's equation to take cracking into account for short-term deflection
calculation. As for long-term deflection, a simplified multiplier approach is proposed in the
Code. These calculation approaches, suitably incorporated in a finite element (FE) package, had
been used to estimate incremental and total long-term deflections of two-way edge-supported
slabs. However, use of this finite element package is neither easy to use in the design office nor it
is available to everybody. An attempt has been made in this paper to train a customized Artificial
Neural Network (ANN) program using the results of the FE package and use the trained network
to easily predict the deflection of two-way edge supported slab. ANN is particularly suitable for
predicting output parameters which depend upon a large number of input parameter like span,
aspect ratio, DL, LL, fc, fy etc. An example demonstrates some simple steps for calculating
short- and long-term deflections.
2007 Institution of Engineers, Bangladesh. All rights reserved.
Keywords: Concrete slab, deflection, slab thickness, serviceability

1.

Introduction

In strength design, the members are so proportioned that will have a proper safety
margin against failure under an overload state. It is also important that the member
1

Former graduate student

120

T.R. Hossain et al. / Journal of Civil Engineering (IEB), 35 (2) (2007) 119-128

performance in normal service be satisfactory. This performance, termed as

serviceability, is not guaranteed simply by providing adequate strength. Service load
deflections under full load may be excessively large, or long-term deflections due to
sustained load may cause damage to partition walls. There are other serviceability
related problems like visually disturbing wide tension cracks and discomforts due to
vibrations.
In the past, questions of serviceability were dealt with indirectly by limiting the stresses
in concrete and steel at service loads to some rather conservative values that resulted in
satisfactory performance. Now, with strength design methods in general use that permit
more slender members through more accurate assessment of capacity, such indirect
methods will no longer do. Use of high strength materials further contributes toward this
trend of smaller member sizes. ACI Code (2002) proposes minimum slab thickness to
ensure serviceability and at the same time allows thinner slabs if deflection calculation
permits so. ACI Code also provides a simple deflection calculation procedure. A
general-purpose finite element package, suitably adopted to take into consideration the
effect of cracking according to ACI Code, has been proved reasonable to predict
deflection in several previous studies (Hossain 1999, Hossain et al 2005, Hossain and
Alam 2004, Hossain and Alam 2003). However, use of this nonlinear finite element
package is neither easy to use in the design office nor it is available to everybody.
Ahmed and Chowdhury (1999) proposed design charts to calculate elastic deflections of
edge supported slabs. As in most cases slabs are cracked even in service load, thickness
selection on the basis of elastic deflection may not be always justified. Hossain and
Alam (2007) presented a number of design charts to take into account the effect of
cracking. Although, deflection depends on a number of geometric, material and loading
parameters which in turn control level of cracking, these charts simplified the nonlinear
relation and calculates reasonable deflection multiplier on the basis of level of cracking.
As deflection depends on a number of parameters, it is possible to employ an Artificial
Neural Network program which can establish a relation between a large number of inputs
and outputs by learning from known values. In the current work, a customized ANN
computer program (Siddiquee 2007) has been trained using the nonlinear FE (Hossain
1999) results so that it can be used to predict deflections as an alternative to performing
time-consuming explicit FE analysis.
2.

Description of the FE model

A program module based on global plate stiffness approach was developed by Hossain
(1999) to incorporate the different short- and long-term models for predicting deflection
of reinforced concrete slabs. The module acts as an integral part of the FE package FE77
(1999) and calculates modified elastic properties to represent cracking, creep and
shrinkage for each element, on the basis of stresses of FE solution, which are then fed
back into the assembly module of the FE package. Hossain and Vollum (2002) found
good correlation in the analysis of the real full scale 7 storied building at Cardington
using this FE module employing EC2 (1992) and CEB-FIP Model Code 1990 MC-90
(1990), where creep and shrinkage deflections are dealt with more rigorously along with
the effect of construction load. Deflection estimation procedure in ACI Code is simpler
than these codes where long-term deflections are calculated from instantaneous
deflection using multiplier. Bransons crack model (1977) which is also adopted in the
ACI Code (2002) to calculate instantaneous deflection has been used in the current work
for developing the database to train the neural network. Description of the model and its
incorporation was presented in previous works (Hossain 1999, Hossain and Alam 2007).

T.R. Hossain et al. / Journal of Civil Engineering (IEB), 35 (2) (2007) 119-128

121

This Branson/ACI model was found to give good prediction of experimental deflections.
A comparison is shown in the next section.
3.

Validation of the model with experiments (Shukla slabs)

Shukla and Mittal (1976) carried out a series of tests on two-way edge-supported slabs.
All the slabs were 214 cm square and 8 cm thick. The slabs were supported on reinforced
concrete walls with centre to centre span of 183 cm each way. Their corners were held
down by means of 40 mm diameter steel rods anchored to the floor. Loads were applied
to the test slab in increments of 2 tonnes each through an inverted waffle-tree system
which transferred load at 16 equidistant point of the slab. Two slabs (S-8 and S-11)
from this series were analysed there. S-8 and S-11 were isotropically reinforced with 10
mm bars to provide 5.24 and 4.36 cm2/m steel in each direction respectively. The two
slabs differ in concrete strengths which were 15.9 (S-8) and 22.0 (S-11) N/mm2. Moduli
of rupture and elasticity were not reported and hence have been estimated using the ACI
equations. Details of the slab dimensions and FE mesh can be found in Alam (2003).
Short-term load-deflection results from FE analysis and experiment were compared for
the central nodes of two slabs in Figs. 1 and 2. The FE analysis modelled experimental
deflection accurately to a possible extent.

Total load on slab, Ton

35
30
25
20
15
ACI model
Experiment
Neural network load case-1
Neural network load case-2

10
5
0
0

10

12

14

16

18

Central deflection, mm

Fig. 1. Comparison of deflection for edge-supported Shukla & Mittal slab S-8 and FE analysis

Total load on slab, Ton

30
25
20
15

ACI model
Experiment
Neural network load case-1
Neural network load case-2

10
5
0
0

10

12

Central deflection, mm

Fig. 2. Comparison of deflection for edge-supported Shukla & Mittal slab S-11 and FE analysis

122

T.R. Hossain et al. / Journal of Civil Engineering (IEB), 35 (2) (2007) 119-128

4.

Artificial neural network

4.1

ANN: from brain to mathematical prediction tool

Artificial Intelligence (AI) is a very versatile and potential technology in the field of
computer technology, which enables computer users in various fields to solve problems
for which algorithmic approach cannot be formulated and which normally requires
human intelligence and expertise. Expert Systems (ESs) and Artificial Neural Networks
(ANNs), the best known manifestations of AI, have today gained immense credibility
and acceptance in many professional fields. ANN approaches have successfully been
used in wide range Civil Engineering problems where conventional approaches based on
engineering mechanics were proved to be difficult to establish an explicit relationship
between causes and effects. These include a wide range of problems in the fields of
structural engineering, construction engineering and management, environmental and
water resources engineering, traffic engineering, highway engineering, and geotechnical
engineering (Jeng et al. 2003, Flood 2001, Adeli 2001, Kartam 1997, Flood and Kartam
1994).
Artificial neural networks are biologically inspired in the sense that neural network
configurations and algorithms are usually constructed with the natural counterpart in
mind. The tremendous processing power of human brain is basically the result of the
massively parallel processing units called neurons. A human brain functions with
hundreds of thousands of such biological neurons, which are interconnected by a highly
complex network. Every neuron consists of a cell body, axon and dendrites. Dendrites
extend from the cell body to the other neurons where they receive signals at a connection
point called the synapse. These inputs are communicated to cell body where all such
inputs are essentially summed up. If the resulting sum exceeds a specified threshold
value, the cell fires and a signal is sent down the axon.
An artificial neural network, also called a simulated neural network (SNN) or commonly
just neural network (NN) is an interconnected group of artificial neurons that uses a
mathematical or computational model for information processing based on a
connectionistic approach to computation. In most cases, an ANN is an adaptive system
that changes its structure based on external or internal information that flows through the
network. In more practical terms, neural networks are non-linear statistical data
modeling or decision making tools. They can be used to model complex relationships
between inputs and outputs or to find out patterns in data.
4.2

Back-propagation and training the network

Here in this paper, a general purpose ANN program (Siddiquee, 2007) is used, which is
basically a back propagation type of neural network. In this network, a set of input
parameters are connected to a set of output parameters through a set of weights and
hidden or middle layers as shown in the Fig. 3. The network is trained to recognise the
correct input-output pattern by adjusting the weight values of the interconnecting weight
matrix. The adjustment follows an error-correcting method called error backpropagation, from where the name of the method is developed. After sufficient number
of training, when the error becomes gradually diminished, the network becomes capable
of predicting any new data within the trained range of input data or any data outside the
range. Number of any hidden layers actually represent the number of the data inputoutput relationship. The number of hidden layers is determined by gradually increasing
its number and checking the error-norm of the trained data set. The best number of

T.R. Hossain et al. / Journal of Civil Engineering (IEB), 35 (2) (2007) 119-128

123

hidden layers is the number for which the error-norm is the lowest. In this paper the best
number of hidden layers was found out to be 16.
Input
Layer
Middle
Layer
Output
Layer

Fig. 3. Basic structure of a Neural Network

A large number of nonlinear FE analyses using Hossains (1999) module employing

Bransons crack model was carried out by varying the relevant geometric and material
parameters. Using the results of these analyses, the ANN was trained as shown in Fig. 4.
Once the amount of error, i.e. the residual becomes very small, the network is ready for
prediction.

Fig. 4. The Neural network being trained for corner panel of edge-supported slab

124

4.3

T.R. Hossain et al. / Journal of Civil Engineering (IEB), 35 (2) (2007) 119-128

Use of the trained network as prediction tool

The use of this ANN is shown in Fig. 5, where input parameter like span, loading and
material properties were given as inputs to predict maximum stress, elastic and cracked
deflections. It was found in Uddin (2005) that these predictions are reasonably accurate
in comparison with the FE analysis. Prediction of the Shukla slab deflections were also
tried using ANN with two different load paths and compared with the previously
presented FE results in Figs. 1 and 2. ANN predictions and FE results are almost
identical.

Fig. 5. The Neural network being used for predicting stresses and deflections

5.

ACI serviceability provisions

There are presently two approaches to control excessive deflections. The first one is
indirect and consists in setting suitable upper limits on the span-depth ratio. This is
simple, and is satisfactory in many cases where spans, loads and load distributions, and
member sizes and proportions fall in the usual range. Otherwise, it is essential to
calculate deflections and to compare those predicted values with specific limitations that
may be imposed by codes or by special requirements.
The 1963 ACI Code (1963), in which the coefficient method of analysis was introduced,
provided that the slab thickness should not be less than 3.5 inch and not less than the
total perimeter divided by 180. These limitations gave satisfactory results for edgesupported slabs. The more general unified method of two-way system analysis found in
recent edition of ACI Code (2002) contains three equations governing minimum slab
thickness. These equations account for the relative stiffness of slab and edge beams, the
ratio of long to short panel side dimensions, and conditions of restraint along the edges.

T.R. Hossain et al. / Journal of Civil Engineering (IEB), 35 (2) (2007) 119-128

125

The equation for beam-supported slab normally yields a higher slab thickness than
obtained from perimeter divided by 180.
5.1

ACI Code (2002) Thickness: Slabs with Beams on All Sides

The parameter used to define the relative stiffness of the beam and slab spanning in
either direction is . Then m is defined as the average value of for all beams on the
edges of given panel. According to ACI Code (2002), for m equal to or less than 0.2, the
minimum thickness of slabs without interior beam shall apply.
For m greater than 0.2 but not greater than 2.0, the slab thickness must not be less than

h=

ln (0.8 + f y / 200,000 )
36 + 5 ( m 0.2 )

(1)

and not less than 5.0 inch (125mm).

For m greater than 2.0, the thickness must not be less than

h=

ln (0.8 + f y / 200,000)
36 + 9

(2)

and not less than 3.5 inch (89mm).

where, ln= clear span in long direction, m = average value of for all beams on edges of
a panel, = ratio of clear span in long direction to clear span in short direction.
6.

Deflection calculation: ACI provisions

ACI Code (2002) also permits slab thickness less than that obtained from above
equations if it can be shown by computation that deflections will not exceed the limit
values given in the Code. Calculation of slab deflection is complicated due to the
presence of cracking even in service loads and also there are time-dependent
deformations due to concrete creep and shrinkage. ACI Code incorporates Bransons
equation for calculating short-term deflection considering cracking. As for long-term
deflection calculation, unlike EC2 (1992) and MC90 (1990), ACI Code (2002) adopted a
simplified multiplier approach. According to Nilson (1997), the philosophy behind it is
that uncertainties regarding material properties, effect of cracking and load history never
justifies precise deflection calculations. The calculated deflections must satisfy the
maximum permissible deflection tabulated in the Code to ensure serviceability.
6.1

Long-term deflection multiplier

On the basis of empirical studies, ACI Code (2002) specifies that additional long-term
deflection due to combined effects of creep and shrinkage shall be calculated by
multiplying the immediate deflection by the factor:

1 + 50

(3)

where, = a time-dependant coefficient, ACI Code (2002) suggested a five-year value of

=2.0, and =As/bd, is usually zero for slabs as compression steels are seldom used.

T.R. Hossain et al. / Journal of Civil Engineering (IEB), 35 (2) (2007) 119-128

126

7.

Example showing short- and long-term deflections estimation using the

trained ANN

To demonstrate the method of deflection calculation using the trained ANN program
following ACI Code, an example is worked out here. Unlike the approach shown in
Nilson (1997), cracking in slab is considered in the analysis. In the current example,
short- and long-term deflections of a 3.66 m x 4.27 m corner panel slab were estimated
with following parameters.
Slab thickness was calculated using recent formula shown in Equation 2 and found to be
94.7 mm (rounding to 100 mm was not done in this study). The following parameters are
assumed: fc= 20.7 MPa, f y= 414 MPa, Ec= 20.7 GPa, Poissons ratio =0.18 and Modular
ratio, n=10. A reduced value of 0.33fc (MPa) was used for rupture strength of concrete
instead of 0.60fc (MPa). Tam and Scanlon (1986) produced good correlation between
calculated deflection with 0.33fc value and mean field-measured deflection. This
approach of using reduced modulus of rupture to take into the effect of cracking due to
restraint shrinkage was reported in a series of papers (ACI Committee 435 1991,
Thompson and Scanlon 1988, Ghali 1990). The total dead load with 1.20 kN/m2 of floor
finish and 2.25 kN/m2 self-wt was 3.45 kN/m2 and live load was 3.83 kN/m2. From the
trained ANN, the immediate deflection was found for total dead and live load. The
calculation of long-term deflection was performed using sustained load of 30% live load
and =2.0 as proposed by the ACI Code. An earlier study (Hossain et al. 2005)
demonstrated a deflection calculation procedure with =3.0 as proposed by Branson for
slabs.
7.1

Long-term deflection calculation

Immediate deflection for dead load and live load predicted using the ANN as shown in
Fig. 4.

d +l = 4.60 mm (With nonlinear FE analysis for the same slab, the deflection found was
4.57 mm)
Assuming 50% of creep due to self-weight occurs before the finishing of the building
starts. The time-dependent portion of dead load deflection was

d = 4.60

1.20
2.25
1 + 4.60
2 = 1.42 + 1.52 = 2.94 mm
7.28
7.28

The long-term deflection due to sustained portion of the live load was

0.3 L = 4.60

3.83
0.3 3 = 2.17 mm
7.28

The instantaneous deflection due to application of short-term portion of the live load was

0.7 L = 4.60

3.83
0.7 = 1.69 mm
7.28

The total incremental deflection is = 2.94 + 2.17 + 1.69 = 6.8 mm

The ACI Code limitation of incremental deflection is

span
= 7.63 mm
480

T.R. Hossain et al. / Journal of Civil Engineering (IEB), 35 (2) (2007) 119-128

127

It is observed that slab thickness is adequate from incremental deflection consideration.

The total deflection was

total = 4.6

2.25
1.20
3 + 4 .6
3 + 2.17 + 1.69 = 10.40 mm
7.28
7.28

The ACI Code limitation of total deflection is

span
= 15.24 mm
240

From calculation, slab thickness was found to be adequate from total deflection
consideration.
8.

Conclusions

Hossain & Alam (2004) showed that in most cases the thickness provided by the ACI
Code are proved to be adequate where spans, live loads and concrete strength are within
normal range and when the slabs are mostly uncracked or slightly cracked. For shorter
spans with lighter loads, a smaller thickness may suffice from serviceability point of
view. The ACI Code allows slab thickness less than the specified value if calculated
values are within code-specified limits. So it would be economical to use thinner slabs in
such situations where deflection analysis permits so. On the contrary, for excessive live
load and larger panels, which generate high level of cracking in slab, providing ACI
Code minimum thickness may not be adequate. In such conditions, deflection
calculations should be mandatory to decide a higher thickness. In either case, a quick
estimation of slab deflection as demonstrated in the current paper using the trained ANN
software which is capable of producing results as good as the nonlinear FE analysis,
should indicate if ACI minimum thickness is appropriate to use.
References
ACI Committee 435 (1991), State-of-the-Art Report on Control of Two-way Slab Deflections,
Detroit: American Concrete Institute.
ACI Publication 318-02 (2002), Building Code Requirements for Reinforced Concrete, Detroit:
American Concrete Institute.
ACI Publication 318-63 (1963). Building Code Requirements for Reinforced Concrete, Detroit:
American Concrete Institute.
Adeli, H. (2001), Neural Networks in Civil Engineering: 1989-2000, Computer-Aided Civil
and Infrastructure Engineering , 16 (2), 126142.
Ahmed, B. and Chowdhury, S.R. (1999), Simplified calculation method for serviceability
deflection of edge supported slabs; Part 1: Modelling slab response, Journal of the Civil
Engineering Division, IEB, 27, 29-40.
Ahmed, B. and Chowdhury, S.R. (1999), Simplified calculation method for serviceability
deflection of edge supported slabs; Part 2: Coefficient method, Journal of the Civil
Engineering Division, IEB, 27, 41-57.
Alam, M.R. (2003), Computation of deflection of edge supported reinforced concrete slab under
service loads, M.Sc.Engg Thesis, BUET, Dhaka.
Branson, D.E. (1977), Deformation of concrete structures, New York: McGraw-Hill Book Co.
CEB-FIP (1993), CEB-FIP Model Code 1990- Design Code, Lausanne: Comite EuroInternational du Beton (CEB).
EC2 Editorial Group (1991), Design of concrete structures, Part 1, General rules and rules for
buildings, EuroCode2.
Flood, I. (2001), Neural networks in civil engineering: a review, Civil and structural engineering
computing: 2001, Saxe-Coburg Publications.
Flood, I. and Kartam, N. (1994), Neural networks in civil engineering. I: principles and
understanding, J. Comp. in Civ. Engrg, 8, 131.

128

T.R. Hossain et al. / Journal of Civil Engineering (IEB), 35 (2) (2007) 119-128

Flood, I. and Kartam, N. (1994), Neural networks in civil engineering. II: principles and
understanding, J. Comp. in Civ. Engrg, 8, 149.
Ghali, A. (1990), Deflection of prestressed two-way floor systems, ACI Structural Journal, 87,
60-65.
Hitchings, D. (1999), FE77 User Manual Version 2.54.
Hossain, T.R. and Alam, M.R. (2007), Deflection estimation of two-way edge-supported slabs,
Journal of Civil Engineering , IEB , 35, 1-15.
Hossain T. R., Uddin, M. T. and Siddiquee M.S.A. (2005), Deflection Estimation of Two-way
Edge-supported Reinforced Concrete Slabs using Artificial Neural Network in Third Annual
Paper Meet and International Conference on Civil Engineering, organized by Institution of
Engineers Bangladesh (IEB), held in March 2005.
Hossain, T.R. and Alam, M.R. (2004), ACI minimum thickness provisions and estimated
deflection for two-way edge-supported reinforced concrete slabs, Progress in Structural
Engineering, Mechanics and Computation, Taylor & Frances Group, London. : 1701-1706.
Hossain T. R. and Alam M. R. (2003), Adequacy of ACI Minimum Thickness Provisions For
Two-way Edge-supported Slabs, in Second Annual Paper Meet and International
Conference on Civil Engineering, jointly organized by Institution of Engineers Bangladesh
(IEB) and Bangladesh University of Engineering & Technology (BUET), held at BUET
during 17-18 July 2003.
Hossain, T.R. and Vollum, R.L. (2002), Prediction of slab deflections and validation against
Cardington data, Structures and Building 152, 235-248.
Hossain T.R. (1999), Numerical Modelling of Deflections of Reinforced Concrete Flat Slabs
under Service Load. PhD thesis. University of London.
Jeng, D. Cha, H and Blumenstein, M. (2003), Application of artificial neural network in civil
engineering problems, The International Internet-Processing-Systems-Interdisciplinaries
(IPSI-2003) Conference, Seveti-Stefan, Montenegro, October 5-11.
Kartam, N, Flood, I., Garrett, J.H. and Agarwal, G. (1997), Artificial Neural Networks for Civil
Engineers: Fundamentals and Applications, ASCE.
Nilson, A.H. (1997), Design of concrete structures. New York: McGraw-Hill Book Co.
Polak, M.A. (1996), Effective stiffness model for reinforced concrete slabs. Journal of Structural
Engineering, ASCE, 122, 1025-1030.
Shukla, S.N. and Mittal, M.K. (1976), Short-term deflection in two-way reinforced concrete slabs
after cracking, ACI Journal 73, 416-419.
Siddiquee,
M.S.A
(2007),
Artificial
Neural
Network
Simulator,
http://203.208.166.84/sid/download/
Thompson, D.P. and Scanlon, A. (1988), Minimum thickness requirements for control of twoway slab deflections, ACI Journal, 12-22.
Tam, K.S.S. and Scanlon, A. (1986), Deflection of two-way slabs subjected to restraint volume
change and transverse loads, ACI Journal 83, 737-744.
Uddin M.T. (2005), Deflection estimation of edge supported reinforced concrete slabs using
Artificial Neural Network. M. Engg. thesis, Department of Civil Engineering, Bangladesh
University of Engineering and Technology, Dhaka.