Applicability of Artificial Neural Networks to Predict Mechanical and Permeability Properties of Volcanic Scoria-Based Concrete

Numerous volcanic scoria (VS) cones are found in many places worldwide. Many of them have not yet been investigated, although few of which have been used as a supplementary cementitiousmaterial (SCM) for a long time..e use of natural pozzolans as cement replacement could be considered as a common practice in the construction industry due to the related economic, ecologic, and performance benefits. In the current paper, the effect of VS on the properties of concrete was investigated. Twenty-one concretemixes with three w/b ratios (0.5, 0.6, and 0.7) and seven replacement levels of VS (0%, 10%, 15%, 20%, 25%, 30%, and 35%) were produced. .e investigated concrete properties were the compressive strength, the water permeability, and the concrete porosity. Artificial neural networks (ANNs) were used for prediction of the investigated properties. Feed-forward backpropagation neural networks have been used. .e ANN models have been established by incorporation of the laboratory experimental data and by properly choosing the network architecture and training processes. .is study shows that the use of ANN models provided a more accurate tool to capture the effects of five parameters (cement content, volcanic scoria content, water content, superplasticizer content, and curing time) on the investigated properties. .is prediction makes it possible to design VS-based concretes for a desired strength, water impermeability, and porosity at any given age and replacement level. Some correlations between the investigated properties were derived from the analysed data. Furthermore, the sensitivity analysis showed that all studied parameters have a strong effect on the investigated properties. .e modification of the microstructure of VS-based cement paste has been observed, as well.


Introduction
Concrete is the most widely used building material around the world because of the economic and widespread availability of its constituents, its versatility, its durability, and its adaptability [1].Ordinary Portland Cement (OPC) concrete is a composite material, and its constituents are cement mixed with water, fine-grained aggregate (sand), and coarsegrained aggregate consisting of natural gravel or crushed stones [2].
e considerable amount of carbon dioxide (CO 2 ) liberated during the production of Portland cement, the most commonly used hydraulic cement, is of a greater concern.On average, about 1 tonne of CO 2 is liberated per tonne of Portland cement produced [1].e use of mineral admixtures such as pozzolans in concrete is now widespread due to many economic, ecological, and performance-related benefits [3,4].e term "pozzolan" is originally from the town of Pozzuoli, northeast of Naples in Italy, where pozzolanic deposits from Vesuvius' volcano were found [5].
Pozzolanic materials can be classified as natural and artificial pozzolans.Natural pozzolans could be considered the first cementitious materials used for the production of artificial stones, ancient mortars, and concretes, 3000 years ago [6].Natural pozzolans may be further subdivided into two main groups: (i) those derived from volcanic rocks (volcanic scorias, pumices, etc.); (ii) others derived from rocks and earths [7].An overview on the use of volcanic scoria as cement replacement can be found in the paper recently published by the author [8].Although there are numerous works on using natural pozzolan as a substitute for cement, few studies investigating on volcanic scoria have been reported in the literature.
Many new prediction methods were developed in the last decades to investigate the compressive strength and durability of concrete containing supplementary cementitious materials (SCMs).However, there is no prediction model for concretes containing volcanic scoria as cement replacement.Some previous efforts to develop a predictive model for VS-based concrete used statistical methods that did not seem sufficient for different compositions of concrete.e objective of this work is to develop an empirical model to predict the compressive strength, porosity, and water permeability of concrete containing VS as cement replacement using ANNs. is prediction could be considered useful for concrete mix designers.For this purpose, twenty-one concrete mixes have been prepared with seven replacement levels and three different w/b ratios.Multiple linear regression (MLR) analysis has been used for comparison.In addition, some relationships among the investigated concrete properties have been established.Some estimation equations have been derived, as well.
e importance of this paper is to encourage countries having ample sources of scoria to investigate their potential use as cement replacement and thus making a greener concrete.e current paper focusses not only on one aspect but also it deals with both strength and durability of concrete containing VS as cement replacement.Moreover, technical information on the Syrian pozzolan is scarce, and this paper is part of the first comprehensive research that addresses the investigation of Syrian volcanic scoria.Furthermore, the paper might be particularly interesting for the regions where volcanic scoria is abundant, such as Harrat al-Shaam, which covers parts of Syria, Jordan, and KSA.Regions of similar geology may also get benefits from this study.

Background
2.1.Artificial Neural Networks.Artificial neural networks (ANNs) are algorithms simulating the human neurons.ey are forms of artificial intelligence, which attempts to simulate the networks of the nerve cell (neurons) of the biological central nervous system [9].An artificial neuron, also called a unit or a node, takes several input connections (dendrites in the biological neuron) which are assigned certain weights (analogous to synapses).e unit then computes the sum of the weighted inputs and applies an activation function (analogous to the cell body in the biological neuron).e result of the unit is then passed on using the output connection (axon function) [10].
In the recent year, such intelligent expert systems have been successfully applied in many fields of engineering, among which, they have been extensively used for predicting the compressive strength of blended cement-based concrete [11][12][13][14][15][16][17][18][19][20][21][22][23][24][25][26][27].It was proved that the ANN-based strength prediction model could be successfully used to predict the strength of concrete for various mix ingredients and at different curing times.Bilim et al. [15] concluded in their study that ANNs can be an alternative approach for predicting compressive strength of ground-granulated blast furnace slag-(GGBFS-) based concrete.Saridemir [16] has established ANN models for predicting the compressive strength of concretes containing metakaolin and silica fume developed up to 180 days of curing.e results have shown that ANNs have a great ability for predicting the compressive strength of metakaolin and silica fume-based concretes.Udhayakumar et al. [17] concluded also that a neural network-based strength model could be used successfully to find out the strength development of fly ash-based concrete with the curing time.Chithra et al. [18] have carried out a comparative study between ANN and MLR for predicting the compressive strength of concretes containing nanosilica and copper slag.
e results have revealed that, in terms of the regression coefficient R 2 and MSE, ANN models have provided better results than MLR.Other studies have used ANNs for predicting some durability-related properties of concrete [28][29][30][31][32][33][34][35].ey concluded that ANN models could be used effectively in predicting the concrete durability properties.
ANNs are able to model nonlinear relations between a set of inputs and corresponding outputs.e data set used to develop the ANN models are divided into subsets (i.e., training set, testing set, and validation set).e present paper deals with the prediction of the compressive strength, water permeability, and porosity of concrete using ANNs.MLR analysis has been used for comparison.e predicted concrete properties have been plotted versus the experimental results obtained in the laboratory.
e neural network uses the backpropagation (BP) procedure.e backpropagation learning algorithm, designed to train a feed-forward network, is an effective learning technique used to exploit the regularities and exceptions in the training sample [36,37].Backpropagation neural network (BPNN) which will be used in this research can be considered the most fundamental and widely used method among the ANN methods.It generally consists of multiple layers: an input layer, one or more hidden layers, and an output layer Hidden layers may contain a large number of hidden neurons (processing units).Activation propagation is forwarded from the input layer toward the output layer, and then the algorithm compares the network outputs with known targets as it is a supervised learning algorithm and propagates the error backward.Weights and biases are updated based on calculated errors in order to meet the target.e most common activation function is the sigmoid function, which is a continuously differentiable function that satisfies the relation, as follows [9]: where α is a constant used to control the slope of the semilinear region [16].e activation function is applied to bind the network input and output of the different layers to a specific range that the network can efficiently handle.e logistic sigmoid activation function with a scaling range between 0 and 1 was found to be the best settings for the present application.
By repeating the procedure described above until the error is acceptably small or no marked improvement is noted, the final output can be obtained [38].

Multiple Linear Regression. Multiple linear regression
(MLR) is a statistical method whose general purpose is to 2 Advances in Civil Engineering generate relationship between several independent variables and a dependent variable [39].In MLR, the predicted value of a single dependent variable Y is a linear transformation of one or more independent variables X such that the sum of squared deviations of the observed and predicted Y is a minimum.With five independent variables, as in the present paper, the prediction of Y is expressed by the following formula: where b i values are called regression weights and are calculated in a way that minimizes the sum of squared deviations and X i are the independent variables.
e compression test was conducted on 150 mm cubic concrete specimens (Figures 2(a) and 2(b)) in accordance with ISO 4012.Water permeability measured in terms of depth of water penetration was carried out as per standard EN 12390-8 (Figure 2(c)).Porosity measurements were conducted using the vacuum saturation method in accordance with RILEM CPC 11.3 [40].Table 1 and Figure 3 summarize the chemical composition of the materials used in the concrete mix and the aggregate grading with their physical properties.All of the experimental results were employed in the prediction process using ANNs.

Network Architecture.
Artificial neural network is a powerful data-modeling tool that is able to capture and represent complex input and output relationships.e design of the ANN model requires identifying the network architecture (i.e., number of input neurons, output neurons, hidden layers, and neurons in each hidden layers) and the network settings (activation function and learning rate).Artificial neural networks consist of at least three layers, i.e., an input layer, one or more hidden layer/layers, and an output layer.
ree ANN models have been established: ANN1 for predicting the concrete compressive strength, ANN2 for predicting the concrete water permeability, and ANN3 for predicting the concrete porosity.e adopted network architecture consists of the following: (i) Five neurons (N i � 5) in the input layer, which represent the variables of cement content (CC: kg), volcanic scoria content (VC; kg), water content (W; kg), superplasticizer content (SP; kg), and curing time (t; day) (ii) One neuron in the output layer, which represents the value of the corresponding compressive strength (MPa), water permeability (mm), or concrete porosity (%).
(iii) One hidden layer with 16 or 15 neurons in ANN1 and ANN3 models, respectively, and two hidden layers with 9 and 7 neurons in the ANN2 model.
Determining the optimum number of the hidden layer neurons is an important issue in order to predict accurately a characteristic using ANNs.
e choice of one or two hidden layers is a common practice because of the ability of these networks to approximate any nonlinear function and map any unknown relationships between the input and output variables.Four-layer ANNs (i.e., two hidden layers) have superior fitting capabilities over three-layer ANNs (i.e., one hidden layer); however, three-layer ANNs are computationally faster and have better generalization capabilities [41,42].at is why three-or four-layer ANNs were selected for the present application.
No reasonable theory on how many hidden layer neurons need to be employed for a particular problem has been established.us, the best approach to find the optimum number of hidden neurons is to start with a fewer number of neurons and then slightly increase the number of neurons.In the current work, the network models were developed as follows: (i) ANN1 was built for predicting the compressive strength.e number of hidden neurons that was selected is 16 (∼3Ni) [43] in one hidden layer.e learning rate and momentum were 0.6 and 0.3, respectively.is selection was made upon experimenting all the possibilities of increasing hidden neurons from 10 to 22, learning rate from 0.1 to 0.9, and momentum from 0.1 to 0.9 for each.e experimental data sets are 525, 314, and 316 samples for compressive strength, water permeability, and concrete porosity, respectively.(ii) ANN2 was built for predicting the concrete water permeability.e number of hidden neurons that was selected is 9 in the first hidden layer and 7 in the second layer.e learning rate and momentum were 0.5 and 0.9, respectively.
is selection was made upon experimenting all the possibilities of increasing hidden neurons from 7 to 30 in one hidden layer, learning rate from 0.1 to 0.9, and momentum from 0.1 to 0.9 for each.e experimental data set is 314 samples.(iii) ANN3 was built for predicting the concrete porosity.e number of hidden neurons that was selected is 15 (�3Ni) in one hidden layer.e learning rate and momentum were 0.3 and 0.4, respectively.is selection was made upon experimenting all the possibilities of increasing hidden neurons from 7 to 22, learning rate from 0.1 to 0.9, and momentum from 0.1 to 0.9 for each.e experimental data set is 316 samples.e architecture of the ANN models is shown in Figures 4-6      Advances in Civil Engineering regression analysis has also been carried out using MATLAB software.e validity of the constructed models was evaluated using the following criteria: (i) Root mean squared error (RMSE): the error that arose during the training, testing, and validating in ANN models can be evaluated by the RMSE, which can be calculated using the following function: (ii) e RMSE is used in evaluating the average vertical distance between the data points and the corresponding points on the fitting straight.When the RMSE value is smaller, the ANN model is better.(iii) Mean absolute percentage error (MAPE): the average of the absolute percentage error can be calculated using the following function: (iv) R-square coefficient (R 2 ): the R-square coefficient is the absolute fraction of variance of a model.R 2 can be calculated using the following function: When R 2 is closer to 1, this means, there is a closer relationship between output and targeted output.As expected, all concretes showed an increase in strength with curing time.Concretes specimens containing VS0 had higher compressive strengths at any curing time compared to VS-based binder concretes.In addition, the compressive strength of VS-based concretes decreased with the volcanic scoria replacement level for all curing times.is could be explained by (i) the reduction of cement content in the mix with the increase of VS content; i.e., the dilution effect and (ii) the slowness of the pozzolanic reaction [40].However, due to the continuation of this reaction and the formation of a secondary C-S-H, a greater degree of hydration was achieved resulting in strengths after 90 and 180 days which were comparable to those of VS0-based specimens [40].Furthermore, the compressive strength decreased significantly with the increase in w/b ratio.

Water Permeability and Concrete Porosity.
Results of water penetration depth and porosity tests are illustrated in Figures 10 and 11, respectively, for all binder types, curing age, and w/b ratios.Water penetration depth can be considered as an indication of permeable and impermeable concrete.A depth of less than 50 mm classifies the concrete as impermeable and a depth of less than 30 mm as impermeable under aggressive conditions [44].None of concretes even with w/b � 0.5 was found to be impermeable before 28 days of curing.However, concretes containing VS20 to VS35 can be considered as impermeable after 28 days of curing and as impermeable under aggressive environments after 90 days of curing, according to Neville [44].In addition, none of concretes with w/b � 0.7 was found to be impermeable under aggressive environments even after 180 days curing.Porosity of all mixes decreased with curing time.Furthermore, porosity of the concretes containing binders with high  e reduction in the water permeability and porosity could be attributed to the pozzolanic reaction between the glassy phase in volcanic scoria and the CH liberated from hydration of C3S and C2S [40]. is was also confirmed by the SEM and EDX analysis.

Correlations between Compressive Strength and
Permeability-Related Properties.Some correlations between compressive strength and each of water permeability and porosity of the investigated concretes are plotted in Figures 12 and 13. ese correlations were calculated for the entire population of test results.As shown in Figures 12 and  13, there are no reasonable relationships between the compressive strength and each of permeability-related properties of VS-based binder concrete.is supports the fact that strong concrete does not always ensure durable concrete.For instance, VS25-based binder concrete has lower compressive strengths, but higher water impermeability compared to VS0-based binder concrete at almost all curing times.

Correlations between Water Permeability and Porosity.
e relationship between the water permeability and porosity is given in Figure 14.Definite correlations with a regression coefficient (R 2 ) of 0.92 were observed between the water permeability and porosity, such that one can be predicted from the knowledge of the other.e correlation between the fitted parameters can be graded excellent when R 2 ≥ 0.85 [45].Such similar relationship may need to be developed for other types of concrete aggregates, different volcanic scoria types, and other curing conditions.Advances in Civil Engineering   Advances in Civil Engineering prediction of the compressive strength, the water permeability, and the concrete porosity, respectively.e validation has been carried out with also 15% of the samples.RMSE values of 0.3899, 4.3322, and 0.4859 and R 2 values of 0.9998, 0.9981, and 0.9995 were obtained for prediction of the compressive strength, the water permeability, and the concrete porosity, respectively.e results mentioned above show that training has given the best performance by providing a very low MSE and R 2 value very close to one.Furthermore, testing and validation have given an excellent performance by providing a very low RMSE and R 2 value approximately equal to one.In contrast to the ANN results (Table 4), RMSE values of 6.11, 4.39, and 17.5, MAPE values of 20.922, 52.3229, and 59.212, and R 2 values of 0.748, 0.82, and 0.829 were obtained by the MLR analysis for predicting the compressive strength, the water permeability, and concrete porosity, respectively.Furthermore, the obtained DW values for ANN models were between 1.5 and 2.5, consistent with the ideal range of values.However, the MLR models had very low DW values, which ranged from 0.026 to 0.157.e DW values signify the incidence of null and positive autocorrelation for ANN and MLR models, respectively.

ANN and MLR
Figures 15-17 clearly show that the goodness-of-fit of the ANN models is superior when compared to the MLR models.In addition, the statistical values shown above demonstrate that the prediction of the compressive strength, water permeability, and porosity of VS-based concrete with ANN models is highly accurate.
e results related to compressive strength are in well agreement with those obtained by Chithra et al. [18], Ferhat Bingol et al. [46], and Saridemir [47].However, as no or very little investigations on the prediction of water permeability and porosity of concretes containing natural pozzolans were found in the literature, the results could be comparable with other concrete durability-related properties, such as chloride ion permeability [31] and sulfate attack [30].
Furthermore, it is worth mentioning that the accuracy of ANN models in terms of RMSE, MAPE, R 2 , and DW can be arranged in the following order: compressive strength ⟶ concrete porosity ⟶ water permeability.e sequence in MLR models is different from that in ANNs and can be written as follows when R 2 is taken into account: concrete porosity ⟶ water permeability ⟶ compressive strength and can be written as follows when RMSE and MAPE are taken into account: compressive strength ⟶ water permeability ⟶ concrete porosity.
e equations obtained by MLR analysis are as follows: where f c is the compressive strength (MPa), WPD is the water penetration depth (mm), Po is the concrete porosity (%), t is the curing time (day), CC is the cement content (kg),     Advances in Civil Engineering VC is the volcanic scoria content (kg), W is the water content, and SP is the superplasticizer content.

Sensitivity Analysis of ANN Models.
In order to assess the relative importance of the input variables, sensitivity analysis was conducted based on the Garson equation [48].Garson [48] proposed an equation based on the partitioning of connection weights: where I j is the relative importance of the jth input parameter on the output; Ni and Nh are the numbers of input and hidden neurons, respectively; w is connection weights; the superscripts i, h, and o refer to input, hidden, and output layers, respectively; and the subscripts k, m, and n refer to input, hidden, and output neurons, respectively.Table 5 shows the weights between input and hidden layers (W1) and weights between hidden and output layers (W2).Table 6 shows the relative importance of the input parameters (curing time, cement content, VS content, water content, and superplasticizer content).It can be noted that all parameters have a strong effect on the investigated properties.
As clearly seen in Table 5, curing time was found to be the most influential parameter with a relative importance of 32.2%, 33.52%, and 51.80% for compressive strength, concrete porosity, and water permeability, respectively.e higher relative importance of curing time can be attributed to the significant effect of this parameter on the permeability-related properties of concretes containing volcanic scoria as cement replacement.As mentioned earlier, a significant gradual improvement in permeability-related properties can be obtained with an increase in curing time, particularly at ages ≥ 28 days [38].In addition, it is to be noted that other parameters have also considerable effects on the output values, particularly those forming the "w/b" ratio, i.e., cement content, VS content, and water content.

Microstructural Investigation.
Scanning electron micrographs (SEM) of 7 day-, 28 day-, and 90 day-cured VS30based pastes are shown in Figure 18. Figure 18(a) clearly shows a porous and noncompacted structure in seven daycured VS30-based paste.However, a denser structure can be clearly observed after 28 and 90 days of curing (Figures 18(b) and 18(c)).is can be attributed to the continuation of the cement hydration, and formation of cementitious phases, such as (C-S-H) and (C-A-S-H) through the pozzolanic reaction between the glassy phase in volcanic scoria and CH released during the hydration of calcium silicates (C 3 S and C 2 S) [49].

Conclusion
In this study, artificial neural networks were used for the prediction of 2, 7, 28, 90, and 180 days compressive strength, Advances in Civil Engineering water permeability, and porosity of concretes containing volcanic scoria as cement replacement.MLR was also used for comparison.Based on the results obtained, the following conclusions can be drawn: (1) Volcanic scoria may be more efficient in terms of its contribution to one property rather than another one.For example, volcanic scoria makes a significant contribution to reducing permeability of concrete despite the lower compressive strength.
(2) In terms of concrete durability, the binders containing ≥ 25% VS were better than the control binder and globally gave the best performance among all the replacement levels.Consequently, a significant reduction in OPC consumption and greenhouse gas  Comparison between ANN models and MLR models in terms of RMSE, MAPE, R 2 , and DW statistics showed that ANNs provide better results than those of MLR in prediction of all investigated properties.For instance, R 2 values of 0.748, 0.82, and 0.829 were obtained by the MLR analysis for predicting the compressive strength, the water permeability, and concrete porosity, respectively, while these values, by contrast, were close to one in all ANNs models.(7) Sensitivity analysis showed that all studied parameters in this work (curing time, cement content, VS content, water content, and superplasticizer content) have considerable effects on the properties of concrete containing VS as cement replacement.However, curing time was found to be the most influential parameter with relative importance of more than 30%.(8) Investigating the volcanic scoria cones that have not been exploited yet is highly recommended.In addition, making more sustainable and durable concrete using volcanic scoria is highly encouraged.

Figure 1 :
Figure 1: Macrograph of (a) the investigated volcanic scoria and (b) the EDX analysis.

Figure 2 :
Figure 2: Photos of (a) the cubic concrete specimens, (b) the compression test, and (c) the water permeability set-up.
(v) Durbin-Watson statistic (DW): this statistic criterion is used to verify the existence of multicollinearity.It varies between 0 and 4. e acceptable range of 1.5 to 2.5 indicates that the established models are unaffected by problems related to multicollinearity.

Figure 6 :
Figure 6: ANN3 model for predicting the porosity of VS-based concrete.

Figure 12 :
Figure 12: Correlation between water permeability and compressive strength of VS-based concretes.

Figure 14 :
Figure 14: Correlation between water permeability and porosity of VS-based concretes.

Figure 17 :
Figure 17: Predicted concrete porosity versus experimental concrete porosity for models ANN1 and MLR1.
. Levenberg-Marquardt backpropagation was used as a training function and hyperbolic tangent sigmoid was used as a transfer function.e learning rate, which identifies the amount of adjustments to connection weights Advances in Civil Engineering during training, was determined based on the network performance.e learning rate and momentum were adjusted for each ANN to get the best performance.e data set was divided into three subsets as follows: 70% for training, 15% for testing, and 15% for validating.
e correlation coefficient (R) obtained for training, testing, validating, and overall data for each ANN are shown in Table 2. e artificial neural networks have been established using MATLAB software, Neural Network Toolbox.e

Table 1 :
Chemical composition of the ingredients used in the concrete mix.

Table 2 :
Correlation coefficients (R) of artificial neural network models.
Analysis.Figures 4-6 show the structure of ANN models developed to predict the values of compressive strength, water permeability, and porosity of VSbased concrete.e calculated statistical values are tabulated in Table 3. e training has been carried out with 70% of the samples.RMSE values of 0.3066, 1.9783, and 0.3129, MAPE values of 1.2106, 2.6357, and 1.4998, and R 2 values of 0.9999, 0.9995, and 0.9998 were obtained for prediction of the compressive strength, the water permeability, and the concrete porosity, respectively.e testing has been carried out with 15% of the samples.RMSE values of 0.3931, 2.4439, and 0.4050, MAPE values of 1.5470, 4.8220, and 2.0528, and R 2 values of 0.9998, 0.9990, and 0.9997 were obtained for

Table 3 :
Results obtained for ANN models.

Table 4 :
Results obtained for MLR models.

Table 5 :
Weight matrix and weights between input and hidden layers (W1) and between hidden and output layers (W2) for the ANN1 model.

Table 6 :
e relative importance of input parameters.obtained.erefore,incorporation of more than 25% VS can be adopted in the design of ecoconcrete.(3)Incorporation of volcanic scoria in concrete has significant effects on the properties of concrete, particularly durability-related properties.e water penetration depth and the porosity of VS-based cement concrete demonstrated better performance as compared to plain concrete, especially at curing age of 28 days and longer.(4) A comparison between ANN and MLR methods depicts that ANNs can be used to predict the investigated concrete properties, effectively.(5) e values predicted by ANN models are close to the experimental results.Statistical values, such as RMSE, MAPE, R 2 , and DW which are used to evaluate the feasibility of the ANN and MLR models, have demonstrated that ANN models are all accurate methods for predicting the compressive strength, water permeability, and porosity of VS-based concrete.eir higher accuracy and applicability have made them a desirable substitute for the conventional regression methods.MLR models are less accurate than the ANN ones.erefore, by adopting ANN models, there is no need to go through timeconsuming and costly laboratory tests to obtain the investigated properties of VS-based concrete.(6) Results have indicated that ANN models are not only practical for predicting the compressive strength but also highly efficient for predicting the water permeability and porosity of VS-based concrete.