Prediction of the Strength Properties of Carbon Fiber-Reinforced Lightweight Concrete Exposed to the High Temperature Using Artificial Neural Network and Support Vector Machine

+e artificial neural network and support vector machine were used to estimate the compressive strength and flexural strength of carbon fiber-reinforced lightweight concrete with the silica fume exposed to the high temperature. Cement was replaced with three percentages of silica fumes (0%, 10%, and 20%). +e carbon fibers were used in four different proportions (0, 2, 4, and 8 kg/m). +e specimens of each concrete mixture were heated at 20°C, 400°C, 600°C, and 800°C. After this process, the specimens were subjected to the strength tests. +e amount of cement, the amount of silica fumes, the amount of carbon fiber, the amount of aggregates, and temperature were selected as the input variables for the prediction models.+e compressive and flexural strengths of the lightweight concrete were determined as the output variables. +e model results were compared with the experimental results.+e best results were achieved from the artificial neural networkmodel.+e accuracy of the artificial neural networkmodel was found at 99.02% and 96.80%.


Introduction
Many researchers have investigated lightweight concrete since the 1990s.e new requirements of concrete have contributed to development of lightweight concrete [1].Lightweight aggregates have been classified as natural and artificial.Lightweight concrete can be made with higher strength, better tensile strength, and lower thermal expansion by using lightweight aggregate [2][3][4].e fire resistance of concrete is complicated.Concrete has different thermal characteristics [5].Researchers indicate that the addition of silica fume has caused high densities in the pore structure of concrete.It can result in big spalling because of the big pore pressure by steam when exposed at the elevated temperature [6].Since the evaporation of water starts at 80 °C in concrete, concretes with silica fume may show low performance when compared to concretes without silica fume at high temperatures [7].Lightweight aggregates have high resistance to fire.e heat capacity of lightweight concrete is less than that of heavier concretes [8].
Fiber composites are new materials based on carbon fiber, steel fiber, and polymer fiber.Cement composite containing carbon fiber is used as the functional materials as well as the structural materials [9].e carbon fiberreinforced concrete decreases the drying shrinkage and increases the flexural toughness [10,11].e carbon fiber increases the flexural strength of concrete because it increases the ability to sense its own strain, damage, and temperature [12].e loading capacity is increased when the steel fiber is used as an additive to the lightweight concretes.Furthermore, it has been decreasing the crack width of concrete [13,14].Due to the advantageous properties, fiber composites have been used in concrete structures [9][10][11][12][13][14][15][16][17][18][19].
Researchers have recently studied the ANN modeling for concrete properties [20][21][22][23][24][25][26][27].Zarandi et al. reported that a fuzzy polynomial neural network's model had been devised for prediction of the compressive strength of concrete [20].Topcu and Sarıdemir showed that the ANN could be a predicted model with a tiny error [21][22][23].Yeh indicated that the ANN was much more accurate than the regression analysis [24].Demir showed that ANNs have the strong potential for predicting elastic modulus of the concrete [25].Altun et al. concluded that the ANN predicts the compressive strength of lightweight concrete with steel fiber better than the multilinear regression technique [26].İnan et al. reported that the neuro-fuzzy model exhibits superior performance [27].
e support vector machine (SVM) has been intensively applied to solve classification and function-approximation problems [28].e SVMs use training and testing of data instances.While the neural network aims to minimize the errors, the SVMs seek to minimize the generalization error [29].SVMs have been recently applied in civil engineering.Lee et al. successfully estimated the concrete strength using the SVR algorithm [30].Chen et al. predicted the exposed temperature for concrete exposed to fire with SVM.ey reported that the correctness of the model increased by the increase of the parameters [31].Shi and Dong used to predict the strength of cement using SVM.ey reported that SVM had become a practical method for the estimation of cement strength [32].
An ANN or SVM model for strength properties (compressive strength and flexural strength) of lightweight concrete with silica fume and carbon fiber after elevated temperature has not been investigated.Because of this, the models have been devised to predict the strength properties of lightweight concrete with silica fume and carbon fiber exposed to elevated temperature.

Materials.
e cement was selected as ordinary Portland cement (CEM I 42.5).Silica fume was provided from Turkey.
e chemical analyses of these materials are shown in Table 1.e lightweight aggregate was selected for the concrete mixes.
e specific gravity of lightweight aggregate was found to be 2. e maximum grain size of lightweight aggregate was found to be 16 mm.e water absorption of lightweight aggregate became 23%.e properties of carbon fiber are presented in Table 2.

Specimens and Curing.
e mix proportions are shown in Table 3. e superplasticizer was used in mixtures.e mixtures with silica fume and carbon fiber were prepared.e silica fume rates were used as 0%, 10%, and 20%.Furthermore, the carbon-fiber ratios were used as 0, 2, 4, and 8 kg/m 3 .100 mm cubes were produced for the compressive strength.100 × 100 × 500 mm prisms were produced for the flexural strength.ree specimens were produced for each temperature.
e specimens were kept in laboratory at 21 ± 2 °C for 24 h.en, the specimens were kept in a water tank for curing up to 28 days.After curing, they were exposed to 400 °C, 600 °C, and 800 °C.ey were kept for 1 h for each temperature [33].e heating rate was 2.5 °C/minute [34,35].e specimens were allowed to cool naturally for one day.

Experimental Results
When the concrete is faced with high temperatures, the same effects may be occurring.e evaporable water passes outwardly at 100 °C.e dehydration of the hydrate calcium silicate begins at 180 °C.
e decomposition of calcium hydroxide occurs at 500 °C.More changes occur when the temperature is more than 500 °C [36].e decomposition of the hydrate calcium silicate begins around 700 °C.e changes of strength properties of concrete at the elevated temperatures are related to the concrete properties beside environmental factors [37].
e high-temperature damage on lightweight concrete with silica fume and carbon fiber was investigated.After the concrete specimens were exposed to 400 °C, 600 °C, and 800 °C, the compressive and flexural strength tests were performed.
e strength test results of specimens are given in Figures 1-6.As seen in Figures 1 and 4, compressive and flexural strengths of lightweight concretes without silica fumes showed slightly better performance, when the carbon-fiber percentage is 0.5.When silica fume percentage was increased, the compressive and flexural strengths of lightweight concretes produced with 1% carbon fiber showed a better performance.Furthermore, the compressive and flexural strengths of lightweight concretes containing 2% carbon fiber were decreased in all temperatures.It can be seen from Figures 1-6 that the compressive and flexural strengths of carbon fiber-reinforced lightweight concrete decreased at 400 °C.e experimental results showed that, for each temperature, a strength loss was caused.ese losses have been prevented with silica fume and carbon-fiber content.ese losses can be become due to the escape of free water, and the hydration water on concrete after exposed to the high temperatures.e dehydration decreases the strength, elastic modulus, and thermal conductivity of concrete [38].
e concrete cracks are increased at high temperatures.e C-S-H crystals grow.ey keep less space in the matrix at high temperatures.e high thermal stresses cause cracks.erefore, the concrete microstructures are negatively affected because of these factors [39].Since the high elastic modulus fibers restrict cracking inside the concrete, they can decrease the volume change of concrete.It has decreased the initiation and expansion of inner micro defects of concrete [40].Çavdar investigated the effect of types of fibers on the flexural strength of cementitious composites exposed to high temperature [41].e carbon fiber was not losing its properties, but the other fibers were deteriorated at high temperatures.Tanyildizi reported that carbon fibers

Prediction Model Based on Artificial Neural
Network.An artificial neural network (ANN) is based on the operations of biological neural networks.erefore, ANN can be defined as an emulation of biological neural systems [43].ANN has amazing capability in modeling the human brain [44].ANN is the occurrence of a large number of simple processing elements called neurons.An artificial neuron model is shown in Figure 7. W j is used as the weight for the input X j .Bias is b.Let X � (X 1 , X 2 , ...,X n ) be symbolized the n input to the neuron.Equation ( 1) is the output of the neuron.ey depend on the connection links.ere is a weight of each link.ere is an activation function.is function is used to calculate the   x j w j − b, V � f(u). ( When an ANN model is devised, many factors must be considered.After this, the activation function and the most suitable structure for the ANN model must be established.Furthermore, the number of layers and the units in each layer must be detected [45][46][47].Many kinds of ANN structure can be used.e multilayer feed-forward ANN is one of these. is structure is given in Figure 8.An artificial network model was designed with five input and two output parameters.e amount of cement, the amount of silica fumes, the amount of carbon fiber, and the amount of aggregates and temperature were selected as input variables.e compressive and flexural strengths of the carbon fiber-reinforced lightweight concrete were used as the model output variables.e back propagation learning algorithm was selected as a feed-forward algorithm for this study.Besides, it determined a single hidden layer neural network.e model is shown in Figure 9. ree different algorithms are selected.e algorithms are the Levenberg-Marquardt (LM) backpropagation, Polak-Ribiere conjugate gradient projection (CGP) backpropagation, and one-step secant (OSS) backpropagation algorithms.e number of neurons in the layer is found to be fifteen.144 experimental results have been obtained for artificial neural networks.ey were divided by max values to normalize.From these data, 72 data were selected for the training.e other 72 data were randomly selected for the testing.e ANN results are given in Figures 10-18.
e measured and predicted strengths are shown in Figures 11,12,14,15,17,and 18. e training performance is given in Figures 10,13    Advances in Civil Engineering predicted the compressive strength of carbon fiber-reinforced lightweight concrete exposed to the high temperature with R 2 of 0.9902, 0.9897, and 0.9886.However, Figures 12,15,and 18 show that the ANN models have predicted the flexural strength of carbon fiber-reinforced lightweight concrete exposed to the high temperature with R 2 of 0.9680, 0.9673, and 0.9671.e best learning algorithm for compressive and flexural strengths was Levenberg-Marquardt backpropagation because the maximum correlation coefficient (R 2 ) was obtained from the Levenberg-Marquardt backpropagation. is algorithm was commonly applied [26]. is was also shown in other articles in the literature [48][49][50][51][52][53][54].e artificial neural networks could be used to solve the complicated civil engineering problems [55].

Prediction Model Based on Support Vector Machine
Model.Support vector machines (SVMs) are used to obtain good generalization.Support vector machines attempt to minimize the generalization error [56].Support vector machines were introduced by Vapnik [57].e SVM theory can be explained as follows: e binary classification problem with its training set of N sample points is shown by x i is a sample value of the input vector x consisting of N training patterns.y i is the corresponding value of the desired model output. y is represented as a linear function.e function is shown by e coefficient is w.w is a one-dimensional array.e superscript "T" denotes transposed.ϕ(x) is a nonlinear transformation function.It is used to map the input space to a higher-dimension feature space.
e SVM is used to minimize the empirical risk variable.R emp is defined as y i −  y i     ε is Vapnik's ε-insensitive loss function defined as w and b are then estimated by minimizing the cost function J ε (w, ξ, ξ * ) defined by e constraints are given as follows: ξ i and ξ * i are positive slack variables.C is a positive real constant [31,[56][57][58][59].
In this study, a support vector model was designed by using five input and output parameters.e input variables were used as the amount of cement, the amount of silica fumes, the amount of carbon fiber, and the amount of aggregates and temperature.e model output variables were the compressive and flexural strengths of the carbon fiberreinforced lightweight concrete.Several parameters are needed to be known for the SVR algorithm.Firstly, it should be determined by three parameters.e parameters are C, error insensitive zone (ε), and kernel specific parameters (c) [56].e optimal values of the parameters were obtained after several trials with these data.e optimal values of C, ε, and c were 100, 1.1 ×10 −6 , and 0.9, respectively.Furthermore, data were divided by max values to normalize.72 data were used for training.Besides, the other 72 data were randomly used as the test data set.e SVM results are given in Figures 19 and 20  It can be seen from Figures 19 and 20 that the support vector model has predicted the compressive and flexural strengths of carbon fiber-reinforced concrete exposed to the high temperature with an R 2 of 0.9701 and 0.91602, respectively.Although the ANN model showed a good performance to estimate the compressive strength and flexural strength of carbon fiber-reinforced lightweight concrete exposed to the high temperature, it has a large number of controlling parameters.ese parameters need to be optimum.
is should be made an effort.e SVM model consists of three parameters [60].It can be said that few parameters are easy to identify.

Conclusions
e prediction models (ANN and SVM) were devised for the strength properties of carbon fiber-reinforced concrete with the silica fume exposed to the high temperature.e 0%, 10%, and 20% silica fumes of cement were used in this study.Furthermore, the carbon-fiber volume was used 0, 0.5, 1 and 2 % in concrete mixtures.After exposed to 20 °C, 400 °C, 600 °C, and 800 °C, the compressive and flexural strength tests of lightweight concrete were performed.
e high compressive and flexural strength results were obtained experimentally from the lightweight concrete with 20% silica fume and 1% carbon fiber for all temperatures.Furthermore, compressive and flexural strengths were decreased with 2% carbon fiber for all mixes and temperatures.SVM models predicted the compressive and flexural strengths with R 2 values of 0.9701 and 0.9160, respectively.ree ANN algorithms were tested in this study.e best learning algorithm was obtained as Levenberg-Marquardt algorithm for this study.e ANN prediction model showed a very good statistical performance because the correlation coefficients (R 2 ) between measured and predicted results for compressive and flexural strengths were 0.9902 and 0.9680, respectively.Finally, the results showed that the artificial neural network prediction model could be used with a high degree of accuracy and reliability for the strength properties of lightweight concrete with silica fume and carbon fiber after exposed to high temperature.

Figure 1 :Figure 2 :
Figure 1: Compressive strength results for carbon fiber-reinforced lightweight concrete without silica fume.

Figure 7 :
Figure 7: Artificial neuron model.W j represents the weight for input X j , and b is a bias.

Figure 16 :Figure 17 :Figure 18 :
Figure 16: Training performance for the one-step secant backpropagation algorithm of ANN.

Figure 19 :Figure 20 :
Figure19: Linear relationship between measured and predicted compressive strengths of support vector machine model.

Table 1 :
e chemical properties of cement and silica fume.
[42]ances in Civil Engineering could improve the high-temperature performance of the lightweight concrete[42].
Cement is ASTM Type I.

Table 2 :
e properties of carbon fiber.
Advances in Civil Engineering output of each neuron.Many kinds of activation function can be used.u � ∑ m j�0