Nomenclatures

*A _{st}, A_{sb} *,area of top and bottom reinforcement, respectively

*B, D* width and depth of beam

*B _{f}, D_{f} * width and depth of flange

*B _{w}, D_{w} * width and depth of web

*E _{c}, E_{s} * modulus of elasticity of concrete and steel, respectively

*I _{e} * effective moment of inertia

*I _{g}, I_{cr} * moment of inertia of gross and fully cracked transformed cross section, respectively

*I _{j} ^{jth} * input parameter

*L* length of beam

*M _{cr}, M_{e} * minimum moment at which the cracking takes place at a cross-section in the beam and applied (elastic) moment, respectively

*O* _{1} output parameter

*bias* bias of hidden or output neuron

*d _{EXP}, d_{FEM}, d_{NN} * mid-span deflection from experiments, FEM, and neural network/explicit expression, respectively

*d _{t}, d_{b} * effective concrete cover at top and bottom. respectively

*f _{t} * tensile strength of concrete

cylindrical compressive strength of concrete at 28 days

*h _{k} ^{kth} * hidden neuron

*m* constant

*q* number of input parameters

*r* number of hidden neurons

*w, n* uniformly distributed load and modular ratio, respectively

*w _{cr} * minimum load at which the cracking takes place in the beam cracking

weight of the link between *I _{j} * and

*h*

_{k} weight of the link between *h _{k} * and

*O*

_{1}

ε*cr*, ε*u* cracking strain, and maximum tensile strain of concrete, respectively

ρ*t*, ρ*c* percentage tension and compression reinforcement, respectively

Subscript

*j* input neuron number

*k* hidden neuron number or function number

*o* output neuron number

Superscript

*ho* connection between hidden and output layers

*ih* connection between input and hidden layers

1 INTRODUCTION

Deflection is an important parameter to check the serviceability criteria of structure. The short term deflection is generally calculated using effective moment of inertia of entire span at service load. The equations for effective moment of inertia, available in literature, are mainly based on two approaches: (i) springs in parallel and (ii) springs in series (^{Kalkan, 2010}). The stiffnesses of the uncracked and cracked portions are averaged in the springs in parallel approach (^{Branson, 1965}; ^{Al-Zaid et al., 1991}; ^{Al-Shaikh and Al-Zaid, 1993}; ^{SAA-AS 3600, 1994}; ^{TS 500, 2000}; ^{CSA-A23.3, 2004}; ^{ACI 318, 2005}; ^{AASHTO, 2005}), whereas the flexibilities of the uncracked and cracked portions are averaged in the springs in series approach (^{Ghali, 1993}; ^{CEN Eurocode 2, 2004}; ^{Bischoff, 2005}; ^{Bischoff and Scanlon, 2007}; ^{Bischoff, 2007}).

Considering parallel springs approach first, the following equation of effective moment of inertia (*I _{e} *) in terms of fully cracked and uncracked moment of inertia was originally proposed by

^{Branson (1965)}for simply supported beams as

where, *M _{cr} * = minimum moment at which the cracking takes place at a cross-section in the beam;

*M*= applied (elastic) moment along the span;

_{e}*I*= moment of inertia of the gross cross section;

_{g}*I*= moment of inertia of the fully cracked transformed cross section and

_{cr}*m*= constant.

Eq. (1) was derived empirically based on the experimental test results of simply supported rectangular reinforced concrete (RC) uniformly loaded beams with tension reinforcement, ρ*t* = 1.65% and ratio of moment of inertia of the fully cracked transformed cross section and moment of inertia of the gross cross section, *I _{cr} */

*I*= 0.45 at maximum applied (elastic) moment equal to 2.5

_{g}*M*(

_{cr}^{Branson, 1965}).

Eq. (1) has been adopted in many international standards and codes (^{SAA-AS 3600, 1994}; ^{TS 500, 2000}; ^{CSA-A23.3, 2004}; ^{ACI 318, 2005}; ^{AASHTO, 2005}) to calculate *I _{e} * and therefore deflection, taking

*m*= 3. Some researchers (

^{Bischoff, 2005};

^{Gilbert, 1999};

^{Scanlon et al., 2001};

^{Gilbert, 2006}) found out that Eq. (1) with

*m*= 3 calculates effective moment of inertia accurately in case of medium to high tension reinforcement(ρ

*t*> 1%), while it overestimates effective moment of inertia for low tension reinforcement (ρ

*t*< 1%).

^{Al-Zaid et al. (1991)}experimentally proved that the value of* m* in Eq. (1)depends on the loading configurations and suggested *m* = 2.8 (in Eq. (1)) for uniformly distributed load when *M _{c} * > 1.5

*M*. The value of

_{cr}*m*was found to change from about 3 to 4.3 for moderately-reinforced concrete beams (ρ

*t*= 1.2%,

*I*/

_{cr}*I*= 0.34) in the range of

_{g}*M*<

_{cr}*M*< 1.5

_{c}*M*.

_{cr}^{Al-Shaikh and Al-Zaid (1993)}performed experiments on mid-span point loaded beams with varying reinforcement. The values of

*m*was found to vary from about 1.8 to 2.5 for lightly reinforced beams (ρ

*t*= 0.8%,

*I*/

_{cr}*I*= 0.22) in the range of 1.5

_{g}*M*<

_{cr}*M*< 4

_{c}*M*, while for the heavily reinforced beams (ρ

_{cr}*t*= 2%,

*I*/

_{cr}*I*= 0.44),

_{g}*m*varied in a range of 0.9 to 1.3. They also suggested

*m*= 3 - 0.8ρ

*t*incorporating reinforcement effect in Eq. (1)for point loaded beams.

^{Al-Zaid et al. (1991)};

^{Al-Shaikh and Al-Zaid (1993)}also proposed to calculate

*I*based on cracked length incorporating reinforcement and loading effects respectively.

_{e}Next, consider the springs in series approach. The models based on this approach (^{Bischoff, 2005}; ^{Bischoff and Scanlon, 2007}; ^{Bischoff, 2007}) take into account tension stiffening effect in concrete for calculating *I _{e} *. The deflections obtained by the expression proposed by

^{Bischoff (2005)}have been found in good agreement with experimental deflections for lightly reinforced beams (ρ

*t*< 1%) (

^{Gilbert, 2006};

^{Bischoff and Scanlon, 2007}).

^{Kalkan (2010)} found out that the expressions given by Eq. (1) and ^{Bischoff (2005)} estimate deflections of moderately-reinforced to highly-reinforced concrete beams (ρ*t* > 1%) accurately on using the experimental value of cracking moment which, however, is difficult to obtain for each and every case.

It is observed from the review that no single approach or model is directly applicable for the entire range of practical reinforcement. Therefore, development of an approach for rapid estimation of the mid-span deflections in uniformly distributed loaded RC beams considering entire practical range of reinforcement at service load is desirable. The approach should be simple to use requiring a minimal computational effort but must give accuracy that is acceptable for practical applications. The application of neural network can be such an alternate approach. For generation of training data for neural networks, finite element technique may be used.

Nowadays, neural networks are being extensively applied in the field of structural engineering. Some of the recent applications of neural networks in the field of structural engineering include prediction of time effects in RC frames (^{Maru and Nagpal, 2004}), prediction of damage detection in RC framed buildings after earthquake (^{Kanwar et al., 2007}), structural health monitoring (^{Min et al., 2012}; ^{Kaloop and Kim, 2014}), bending moment and deflection prediction in composite structures (^{Chaudhary et al., 2007}, ^{2014}; ^{Pendharkar et al., 2007}, ^{2010}, ^{2011}; ^{Tadesse et al., 2012}; ^{Gupta et al., 2013}), predicting the creep response of a rotating composite disc operating at elevated temperature (^{Gupta et al., 2007}), optimum design of RC beams subjected to cost (^{Sarkar and Gupta, 2009}), static model identification (^{Kim et al., 2009}), response prediction of offshore floating structure (^{Uddin et al., 2012}), prediction of deflection in high strength self-compacting concrete deep beams (^{Mohammadhassani et al., 2013a}; ^{2013b}) and prediction of energy absorption capability and mechanical properties of fiber reinforced self-compacting concrete containing nano-Silica particles (^{Tavakoli et al., 2014a}; ^{2014b}). These studies reveal the strength of neural networks in predicting the solutions of different structural engineering problems.

This paper presents an alternative approach for estimating effective moment of inertia which is neither spring in parallel nor spring in series approach. Neural network model is developed, at service load, for predicting effective moment of inertia (and deflection), in a RC beam considering entire practical range of tension and compression reinforcement, tension stiffening and flexural concrete cracking. The data sets for training, validating and testing are generated using finite element models. The finite element models have been developed in ^{ABAQUS (2011)} software and validated with the experimental results available in literature. Explicit expression has been obtained based on developed neural network model which can be used in design offices by practicing engineers. The proposed neural network/explicit expression has been validated for a number of simply supported and continuous RC beams. Sensitivity analysis has been performed to understand the influence of relevant parameters on effective moment of inertia.

2 FINITE ELEMENT MODEL AND ITS VALIDATION

The finite element model (FEM) has been developed using the ^{ABAQUS (2011)} software. The beam has been modelled using B21 elements (2-node linear Timoshenko beam element). Under service load, the stress-strain relationship of concrete is assumed to be linear in compression. Concrete has been considered as an elastic material in tension before cracking and softening behaviour is assumed after cracking (Figure 1). Further, at service load, the stress in reinforcement is assumed to be in the linear range. The steel reinforcement has been embedded into the concrete using "REBAR" option in which a perfect bond is considered between steel reinforcement and concrete. In order to consider cracking and tension stiffening, the smeared crack model has been used. Tension stiffening has been defined using post-failure stress-strain data proposed by ^{Gilbert and Warner (1978)}. A high shear stiffness has been assumed to neglect the shear deformations.

The results of FEM have been compared with the experimental results (mid-span deflections of the beam under increasing uniformly distributed load, *w*, after the cracking of the concrete) reported by ^{Al-Zaid et al. (1991)} for a simply supported beam (VB) with 2.5 m clear span (effective span = 2.62 m) and cross-sectional dimensions *B* × *D* = 200 × 200 mm (Figure 2). The other properties considered are: cylindrical compressive strength of concrete at 28 days, = 38.2 N/mm^{2}; modulus of elasticity of concrete, *E _{c} * = 2.96 × 10

^{4}N/mm

^{2}; modulus of elasticity of steel,

*E*= 2 × 10

_{s}^{5}N/mm

^{2}; tensile strength of concrete,

*f*= 3.47 N/mm

_{t}^{2}; cracking moment,

*M*= 5.2 kNm; area of top reinforcement,

_{cr}*A*= 78.54 mm

_{st}^{2}and area of bottom reinforcement,

*A*= 402.12 mm

_{sb}^{2}. The effective concrete cover at top (

*d*) and at bottom (

_{t}*d*) have been taken as 30 mm and 33 mm respectively.

_{b}

In order to define the smeared crack model, the absolute value of the ratio of uniaxial tensile stress at failure to the uniaxial compressive stress at failure is taken as 0.09. The strain at cracking, ε*cr* is taken as 0.00012 and in view of low/moderate tensile reinforcement, *A _{sb} * = 402.12 mm

^{2}(= 1.2%), the plastic strain is (ε

*u*- ε

*cr*) taken as 0.0004. For convergence, about 16 elements are required when cracking is considered (

^{Patel et al., 2014}). Results from the developed FEM and experiments are compared in Figure 3. Close agreement is observed between the results from FEM and experiments.

Next, the results have been compared with experimental results reported by ^{Washa and Fluck (1952)} for four sets of rectangular cross-sectional (Figure 2) simply supported beams: A1,A4; B1,B4; C1,C4; D1,D4 subjected to uniformly distributed loads at service load. Two beams in a set are identical. The cross-sectional properties, material properties, span lengths and uniform distributed loads of all four beams have been given in Table 1. Additionally, *E _{s} * has been assumed as 2 × 10

^{5}N/mm

^{2}.

*E*and

_{c}*f*are taken in accordance with

_{t}^{ACI 318 (2005)}. The mid-span deflections obtained from the FEM (d

_{FEM}) are in close agreement with the reported experimental deflections (d

_{EXP}) as shown in Table 1. The finite element models can therefore be used for generation of data sets.

3 SAMPLING POINTS AND DATA SETS

For development of neural network, significant parameters need to be identified. Eq. (1) shows that *I _{e} */

*I*explicitly depends on

_{g}*I*/

_{cr}*I*and

_{g}*M*/

_{cr}*M*. It is assumed that

_{a}*I*/

_{e}*I*depends on ρ

_{g}*t*and ρ

*c*(percentage compression reinforcement) also. The value of ρ

*c*however depends on ρ

*t*and ranges from 0.0 to ρ

*t*(

*n*- 1)/

*n*, where,

*n*= modular ratio. The value of

*I*/

_{cr}*I*in turn also depends on the combinations of ρ

_{g}*t*and ρ

*c*. Consider a typical beam cross-section as shown in Figure 2 (

*B*= 300 mm;

*D*= 700 mm;

*E*= 2.73 × 10

_{c}^{4}

*N*/

*mm*

^{2};

*E*= 2.00 × 10

_{s}^{5}N/mm

^{2};

*d*= 30 mm and

_{t}*d*= 33 mm). For this beam, the variations of ρ

_{b}*c*and

*I*/

_{cr}*I*with ρ

_{g}*t*are shown in Figure 4. The parameter

*M*/

_{cr}*M*depends on the load and moment only.

_{a}

Taking the above observations into account, ρ*t* is also considered as an input parameter alongwith *I _{cr} */

*I*and

_{g}*M*/

_{cr}*M*. The sampling points of the parameters considered for data generation are shown in Table 2. It may be noted that the combinations of sampling points take into account the different values of the ρ

_{e}*c*(fourth and left out parameter), corresponding to each value of ρ

*t*Considering the equation

*d*= 5

_{FEM}

_{w}L^{4}/384

*E*, the output parameter

_{c}I_{e}*I*/

_{e}*I*is obtained as 5

_{g}*wL*

^{4}/384

*E*.

_{c}d_{FEM}I_{g}

4 TRAINING OF NEURAL NETWORK

Neural network has been developed for the prediction of effective moment of inertia in RC beams. The neural network chosen is a set of multilayered feed-forward networks with neurons in all the layers fully connected in the feed forward manner (Figure 5). The training is carried out using the ^{MATLAB Neural Network toolbox (2009)}. Sigmoid function (logsig) is used as an activation function and the Levenberg-Marquardt back propagation learning algorithm (trainlm) is used for training. The back propagation algorithm has been used successfully for many structural engineering applications (^{Maru and Nagpal, 2004}; ^{Kanwar et al., 2007}; ^{Gupta et al., 2007}; ^{Pendharkar et al., 2007}; ^{2010}; ^{2011}; ^{Chaudhary et al., 2007}; ^{2014}; ^{Sarkar and Gupta, 2009}; ^{Gupta and Sarkar, 2009}; ^{Min et al., 2012}; ^{Tadesse et al., 2012}; ^{Mohammadhassani et al., 2013a}; ^{Gupta et al., 2013}) and is considered as one of the efficient algorithms in engineering applications (^{Hsu et al., 1993}). One hidden layer is chosen and the number of neurons in the layer is decided in the learning process by trial and error.

Different combinations of sampling points of the input parameters and the resulting values of the output parameters are considered in order to train the neural network. Each such combination of the input parameters and the resulting output parameters comprises a data set. The total number of data sets considered for the training, validating and testing of the network are 3444.

Normalisation factors are applied to input and output parameter to bring and well distribute them in the range. No bias is applied to the input and output parameters. Normalisation factors of 4, 2, 7 and 3 are applied to input parameters ρ* _{t}, I_{cr} */

*I*/

_{g}, M_{cr}*M*and output parameter

_{g}*I*/

_{e}*I*respectively.

_{g}70% data sets are used for training and the remaining data sets are divided equally in the validating and testing sets. For the training, several trials are carried out with different numbers of neurons in the hidden layer starting with a small number of neurons in the hidden layer and progressively increasing it, and checking the mean square errors (MSE) for the training, validating and testing. The number of neurons in the hidden layer is decided on the basis of the least mean square errors (MSE) for the training as well as validating and testing. Care is taken that the mean square error for test results should not increase with the number of neurons in hidden layer or epochs (overtraining). The final configuration (number of input parameters - number of neurons in the hidden layer - number of output parameters) of NN is 3-6-1. The responses of proposed neural network model to predict effective moment of inertia for training, validating, and testing are shown in Figures 6 (a)- ^{(c)} respectively. The proposed neural network model achieved good performance as the testing data points are mostly on equity line (Figure 6 (c)). The statistical parameters i.e. mean square error (MSE), root mean square error (RMSE), mean absolute percentage error (MAPE), average absolute deviation (AAD), correlation coefficient (R^{2)} and coefficient of variation (COV) (^{Sozen et al., 2004}; ^{Azmathullah et al., 2005}) of training, validating and testing data sets are shown in Table 3. All the parameters indicate a good agreement.

5 EXPLICIT EXPRESSION FOR PREDICTION OF EFFECTIVE MOMENT OF INERTIA

For the ease of practicing engineers and users, simplified explicit expression can be developed for the prediction of effective moment of inertia. The explicit expression requires the values of inputs, weights of the links between the neurons in different layers, and biases of output neurons (^{Tadesse et al., 2012}; ^{Gupta et al., 2013}).

As stated earlier, the sigmoid function (logsig) has been used as the activation function. The output *O* _{1} (Figure 5) may therefore be obtained as below (^{Tadesse et al., 2012}; ^{Gupta et al., 2013}):

where, *q* and *r* are the number of input parameters and the number of hidden neurons respectively; *bias* _{k} and *bias* _{o} are the bias of *k _{th} * hidden neuron (

*h*) and the bias of output neuron respectively; and are the weight of the link between

_{k}*I*and

_{j}*h*and the weight of the link between

_{k}*h*and

_{k}*O*

_{1}respectively. The weights of the links and biases of the output neurons for NN are listed in Table 4.

The value of *I _{e} */

*I*is equal to de-normalized output

_{g}*O*

_{1}. The effective moment of inertia

*I*may be obtained from Eq. (2) by putting the values of from Table 4 as

_{e}

where, *H* _{1}, *H* _{2}, *H* _{3}, *H* _{4}, *H* _{5} and *H* _{6} may be obtained from Eqs. (5)- ^{(10)} by using the weights and biases (Table 4) as

6 VALIDATION OF NEURAL NETWORK/EXPLICIT EXPRESSION

The developed neural network/explicit expression is validated for a number of simply supported and continuous beams with a wide variation of input parameters. The results (mid-span deflections), obtained from the proposed neural network/explicit expression are compared with the experimental results for simply supported beams available in literature and with the FEM results for continuous beams.

6.1 Simply supported beams

First, the results have been compared with experimental results reported by ^{Washa and Fluck (1952)} for sets of simply supported beams with rectangular cross-section (Figure 2): A2,A5; B2,B5; C2,C5; D2,D5 subjected to uniformly distributed loads, and designated, here, as VB1-VB4, respectively. Two beams in a set are identical. The details of the beams are given in Table 5. Additionally, *E _{s} * is assumed as 2.05 × 10

^{5}N/mm

^{2}.

*E*and

_{c}*f*are taken in accordance with

_{t}^{ACI 318 (2005)}. The mid-span deflections obtained from the proposed explicit expression (

*d*) are shown in Table 5 along with the reported experimental mid-span deflections (

_{NN}*d*). The values obtained from the proposed explicit expression are in reasonable agreement with the reported experimental values of mid-span deflections.

_{EXP}Next, the results have been compared with experimental results reported by ^{Yu and Winter (1960)} for simply supported beams with T cross-section (Figure 7): A-1; B-1; C-1; D-1; E-1; F-1 subjected to uniformly distributed loads, and designated, here, as VB5-VB10, respectively. The cross-sectional and material properties of the beams are given in Table 6. The mid-span deflections obtained from the proposed explicit expression (*d _{NN} *) are shown in Table 6 along with the reported experimental mid-span deflections (

*d*). Again, the values obtained from the proposed explicit expression are in reasonable agreement with the reported experimental values of mid-span deflections.

_{EXP}

The results (mid-span deflections) obtained from the proposed neural network/explicit expression need to be compared with the finite element results for lightly reinforced simply supported beams (ρ*t* ≤ 1%) also. Consider a 2.625 m long simply supported beam VB11 with rectangular cross-section (Figure 2) subjected to uniformly distributed load. The other properties are: *B* = 200 mm; *D* = 500 mm; = 27.9 N/mm^{2}; *E _{s} * = 2.05 × 10

^{5 }N/mm

^{2};

*A*= 400 mm

_{st}^{2};

*A*= 700 mm

_{sb}^{2};

*d*=

_{t}*d*= 35 mm.

_{b}*E*and

_{c}*f*are taken in accordance with

_{t}^{ACI 318 (2005)}. Mid-span deflections, for beam VB11 are obtained from the proposed explicit expression, FEM and

^{ACI 318 (2005)}for varying magnitude of uniformly distributed loads,

*w*and shown in Figure 8. The mid-span deflections obtained from the proposed explicit expression and FEM are close for the range of the load considered. The difference between FEM and proposed explicit expression is 2.91% as compared to 16.81% difference between FEM and

^{ACI 318 (2005)}at 4

*w*,(

_{cr}*w*= cracking uniformly distributed load).

_{cr}

Consider another simply supported beam VB12 subjected to uniformly distributed load with the same cross-sectional (Figure 2) and material properties as that of beam VB11 except *A _{sb} *. The value of

*A*is now assumed as 900 mm

_{sb}^{2}. The close agreement is observed between the mid-span deflections obtained from the proposed explicit expression, FEM and

^{ACI 318 (2005)}as shown in Figure 9.

6.2 Continuous beams

In order to validate the proposed explicit expression for a continuous beam, results from the explicit expression are also compared with FEM and ^{ACI 318 (2005)} results for a 12.2 m two equal span uniformly distributed loaded continuous beam (VB13) with rectangular cross-section (Figure 2). The other properties are: *B* = 152.4 mm; *D* = 203.2 mm; = 24.1 N/mm^{2}; *E _{s} * = 2.07x10

^{5}N/mm

^{2};

*A*=

_{st}*A*= 112 mm

_{sb}^{2};

*d*=

_{t}*d*= 25 mm.

_{b}*E*and

_{c}*f*are taken in accordance with

_{t}^{ACI 318 (2005)}.

The mid-span deflections obtained from the proposed explicit expression and FEM are close for the range of the load considered (Figure 10). The difference between FEM and proposed explicit expression is 5.34% as compared to 28.25% difference between FEM and ^{ACI 318 (2005)} at 5*w _{cr} *.

Similarly, another 12.2 m two equal span continuous beam VB14 with rectangular cross-section (Figure 2) subjected to uniformly distributed load has been considered. The cross-sectional and material properties are taken same as that of beam VB13 and only *A _{sb} * and

*A*are increased to 200 mm

_{st}^{2}. The close agreement is observed between the mid-span deflections obtained from the proposed explicit expression, FEM and

^{ACI 318 (2005)}as shown in Figure 11.

7 SENSITIVITY ANALYSIS

The proposed explicit expression shows satisfactory performance on validation with experimental results available in literature and FEM results. A sensitivity analysis is carried out next to capture the influence of individual input parameters on output parameter using the proposed explicit expression. The effect of input parameters ρ* _{t}, I_{cr} */

*I*/

_{g}, M_{cr}*M*along with additional parameters ρ

_{e}*on output parameter*

_{c}, n*I*/

_{e}*I*is studied. Only one parameter (the parameter under consideration) is varied at a time, keeping the other parameters constant.

_{g}7.1 Effect of ρ**t**

As stated earlier, ρ*t* has been considered as the input parameter in the present study. Figure 12 shows the variation of *I _{e} */

*I*with respect to ρ

_{g}*t*for various values of

*I*/

_{cr}*I*, keeping the value of

_{g}*M*/

_{cr}*M*constant as 0.5. Rich influence of ρ

_{e}*t*on

*I*/

_{e}*I*is seen in Figure 12. Though, the effect is significant for all values of ρ

_{g}*t*, the effect of lower values of ρ

*t*is more significant in case of higher

*I*/

_{cr}*I*.

_{g}

7.2 Effect of **I** **cr**/**I** **g**

The variation of *I _{e} */

*I*with respect to

_{g}*I*/

_{cr}*I*for various values of ρ

_{g}*t*is shown in Figure 13. The value of

*M*/

_{cr}*M*is kept constant as 0.5. The effect is significant only for lower values of

_{e}*I*/

_{cr}*I*in case of low ρ

_{g}*t*. However, the effect extends of the range considered for

*I*/

_{cr}*I*in case of higher values of ρ

_{g}*t*. The effect of

*I*/

_{cr}*I*is significant for all values of ρ

_{g}*t*.

7.3 Effect of **M** **cr**/**M** **e**

As stated earlier, *M _{cr} */

*M*has been considered as the input parameter affecting

_{e}*I*/

_{e}*I*. The variation of

_{g}*I*/

_{e}*I*with respect to

_{g}*M*/

_{cr}*M*for different values of ρ

_{e}*t*is shown in Figure 14 (a). The value of

*I*/

_{cr}*I*is kept constant as 0.5. Similarly, Figure 14 (b)shows the variation of the ratio

_{g}*I*/

_{e}*I*with respect to

_{g}*M*/

_{cr}*M*for different values of

_{e}*I*/

_{cr}*I*. The value of ρ

_{g}*t*is kept constant as 1.5. As expected, the effect of

*M*/

_{cr}*M*is significant during cracking (

_{e}*M*/

_{cr}*M*< 1) and the value of

_{e}*I*/

_{e}*I*increases with increase in value of

_{g}*M*/

_{cr}*M*up to 1.00. The effect is more for higher value of ρ

_{e}*t*.

7.4 Effect of ρ**c**

Figure 15 shows the variation of *I _{e} */

*I*with respect to ρ

_{g}*c*for different values of ρ

*t*. The value of

*M*/

_{cr}*M*is kept constant as 0.5. The value of

_{e}*I*/

_{e}*I*is found to increase with the increase in value of ρ

_{g}*c*. A significant variation is observed in case of higher value of ρ

*t*.

7.5 Effect of **n**

The variation of *I _{e} */

*I*with respect to

_{g}*n*for different values of ρ

*t*is shown in Figures 16 (a)-

^{(b)}for ρ

*c*= 0 and ρ

*c*= ρ

*t*(

*n*- 1)/

*n*respectively. The value of

*M*/

_{cr}*M*is kept constant as 0.5. The nature of plot changes from concave to convex with increase in ρ

_{e}*t*.

8 CONCLUSIONS

An explicit expression has been proposed for the prediction of effective moment of inertia (and deflection) considering concrete cracking, tension stiffening and entire practical range of reinforcement at service load. A set of three parameters (ρ* _{t}, I_{cr} */

*I*/

_{g}, M_{cr}*M*) has been identified that govern the change in

_{e}*I*/

_{e}*I*and therefore deflection. Using the sampling points of these parameters and the validated FEM, the data sets have been generated for training, validating and testing of neural network. The explicit expression has been developed from the trained neural network. The proposed explicit expression has been validated for a number of simply supported and continuous beams and it is found that the predicted deflections have reasonable accuracy for practical purpose. Sensitivity analysis has been carried out to capture the influence of individual input parameters on output parameter. The effect of the input parameters ρ

_{g}*/*

_{t}, I_{cr}*I*/

_{g}, M_{cr}*M*on output parameter

_{e}*I*/

_{e}*I*is studied using the proposed explicit expression . The lower values of ρ

_{g}*t*are found to have more significant effect on

*I*/

_{e}*I*. The effect of

_{g}*M*/

_{cr}*M*is found to be significant during cracking (

_{e}*M*/

_{cr}*M*< 1) and the value of

_{e}*I*/

_{e}*I*is found to increase up to 10.0 with increase in value of

_{g}*M*/

_{cr}*M*. The effect of ρ

_{e}*c*and

*n*is found to be less significant and can be incorporated through

*I*/

_{cr}*I*.

_{g}The methodology presented herein can be further developed for beams with point loads. The effect of shear deformation may be incorporated in future studies by considering span to depth ratio of beam as an input parameter. Similarly, age of loading and characteristic compressive strength of concrete can also be considered as input parameters to account for shrinkage cracking in future studies.