Engineering properties of concrete incorporating waste glass as natural sand substitution with tin can fiber: experimental and ANN application

Disposal of waste material is causing environmental issues all around the world. Waste glass and tin can are two such products that impose detrimental effects on the environment. One feasible approach is to utilize these discarded materials as constituents of concrete. This study investigates the performance of concrete prepared with glass waste as fine aggregate (GWA) and tin can fiber (TCF) in terms of workability, fresh density, compressive, and splitting tensile strength as desired concrete properties. An Artificial Neural Network (ANN) approach has been adopted for developing predictive models to simulate the impact of these waste materials on concrete properties. Experimental findings demonstrate that the worability of concrete mix decreases with increasing GWA and TCF contents in the mix. Increasing TCF percentages results in an increase in fresh density of concrete; however, a declining pattern in fresh density has been observed with increasing GWA percentage in the concrete mix. Compressive strength at all curing ages shows positive trends till 20% GWA and 1% TCF incorporation in concrete. Similar pattern has been observed for splitting tensile strength up to 10% GWA and 0.5% TCF. Constructed models are feasible for prediction of the desired concrete properties as verified by various statistical parameters. Sensitivity analysis reports that TCF has a greater influence on concrete properties than GWA.


Introduction
The rapid urbanization, population growth, excessive industrial, manufacturing, and construction activities have resulted in increase of waste materials [1]. Each year, 2.12 billion tons of waste are being dumped without recycling or any type of treatment [2]. As majority of the waste are non-biodegradable, these wastes are causing serious pollution in the surrounding environment [3][4][5]. Waste glass and discarded tin cans, generated enormously in daily basis, are directly disposed of to the landfills. Waste glass disposal is considered a central environmental problem around the world. Annually, glass waste Fiber reinforced concrete is a composite material produced by the addition of discrete reinforcing fibers in concrete [41][42][43]. Incorporation of fiber in concrete, resolve the brittle behavior of concrete, improves the energy absorption capacity and ductility of concrete [44,45]. FRC can develop compressive strength and tensile strength over 150 MPa and 8 MPa respectively [46]. Tin can as fiber reinforcement for production of fiber reinforced concrete have gained popularity in recent years [1,47,48]. Compressive strength of concrete having tin fiber is prominent up to 1% addition [47]. Meanwhile, 1.0% inclusion of tin fiber in concrete have 8.2% higher flexural strength than conventional concrete [47].
Therefore, models need to be developed between input variable as GWA and TCF and output variable as workability, fresh density, compressive strength, and splitting tensile strength at 7, 28, and 56 days. Nowadays, ANN has received ample priority for modeling and predicting concrete properties. Topçu and Saridemir [49] used artificial neural network for simulation of characteristics of concrete for incorporation of waste AAC aggregate, and found good accuracy of ANN model. Ince [50] reported viable and promising results of an ANN-based two-parameter model (TPM) of concrete. Similarly, good coherence with experimental result of compressive strength was registered by Kewalramani and Gupta [51] using ANN.
This research intends to explore the effect of GWA and TCF on experimental values of the workability, fresh density, compressive, and splitting tensile strength of concrete at 7, 28, and 56 days as well as modelling of these values using ANN. This study also wishes to perform sensitivity analysis of the variables on the properties of concrete. To our knowledge, this is the maiden paper focusing on modeling of concrete having glass waste and tin fiber using ANN. In addition, this research is unique in terms of finding relative effect by sensitivity analysis.

Properties of raw materials
The properties of sand, stone, glass and fiber are displayed in Table 1. Natural sand and waste glass have been incorporated as fine aggregates in concrete. Waste glasses were achieved from mechanical crushing of window glasses of different vehicles. The gradation curve of glass and sand have been shown in Fig. 1. The curve were compared with the ASTM (C33) specified lower and upper gradation curve and found that particle distribution curve of sand and glass fall within the allowable limit [52]. The condensed milk can waste were processed to make tin can fiber having 20 mm length and 2 mm width. All characteristics of crushed glass waste have been found lower than sand. As binding agent in concrete mixing, Ordinary Portland Cement (ASTM Type I) was chosen to use. The used materials has been displayed in Fig. 2.

Mix proportioning
There were two types of concrete mixes created in this experimental work ( Table 2). The control concrete mix (M0) containing cement (400 kg/m 3 ), stone (1200 kg/m 3 ), sand (600 kg/m 3 ) and water (200 kg/m 3 ). The rest mixes were prepared with GWA of 10%, 20%, and 30% as a partial substitution of sand and with the addition of TCF from 0.5 to 1.5% in step of 0.5%. In all mixes, a constant ratio of 1:1.5:3 (M20) has been maintained for concrete production. M20 is mostly used for reinforced cement concrete structures and has a water-cement ratio of 0.5. Since the workability behavior of concrete having waste glass is not well known and the flow of control concrete containing M20 grade is well established hence M20 was taken in in this study to avoid unexpected workability of concrete incorporating glass waste.

Preparation of specimens
Molds having a height of 200 mm and a diameter of 100 mm were greased properly to guarantee no loss of water during the casting process and to avoid concrete adhesion with molds for ensuring effortless demolding. The concrete was cast in two layers of 100 mm each where every layer was given 25 blows using a 5-mm tamping rod. Figure 3 shows the preparation of concrete sample preparation. The molds were left to dry for 24 h at 25 • C before the concrete specimens were demolded from them. After that, concrete cylinders were submerged in fresh water immediately to ensure proper curing for 7, 28, and 56 days. Table 3 exhibits the adopted experimental standards for this study.

Artificial neural network (ANN)
ANN is a mathematical modeling tool that mimic the nervous systems of human brain (Fig. 4). It has machine learning, adaptive learning, abilities to recognize patterns and self-organization. A artificial neural network contains neurons, input, hidden (one or many) and output layers, weight, bias, sum, activation and learning function. Neurons  are processing unit which are grouped in the layers (Fig. 5). Each connection between neurons in the layers has a coefficient assigned to it called weight. Within the same layer, there are usually no interconnections between units. Sum function calculates the product of inputs and weight along with addition of bias (Eq. 1). Activation processes determines the cell output through processing of the weighted sum (Eq. 2). ANN learns autonomously from the given training process and forms a connection between input and output variables [53]. It can produce meaningful outcomes even when the provided data have errors or are incomplete [54].  where net j represents weighted sum of j neuron for the input obtained from previous layer, w ij represents weight value between j neurons in the preceding layer, x i stands for output of the i neuron, b represents bias value.

Stone (kg/m 3 ) Sand (kg/m 3 ) GWA (%) GWA (kg/m 3 ) TCF (%) TCF (kg/m 3 ) Water (kg/m 3 )
where out j represents output of j neuron, and ∝ stands for a constant which controls the slope value within the semi-linear region. Because log-sigmoid is utilized as an activation function between the input and hidden layers in ANN, data normalization was required. Normalized sets of actual data could improve learning speed, and the training process's performance, accuracy, and stability [55]. Normalized values were measured using the following equation. Equation (3) where X act signifies experimental value, X min signifies the lowest value, X max signifies highest value, and X norm signifies normalized value.

Sensitivity analysis
Details about the target outputs are contained in input parameters. More input parameters usually imply more detail and simulation capacity. However, few parameters might be immaterial that can deceive the training process and degrade the function of learning algorithm. The determination of effect of variables on the responses in ANN is unlike the classic statistical models because of relative complexity of neural network [56]. Contribution of the independent variables on the target concrete properties were measured by following equation depending on the value of weights, Eq. (4) where N r=1 w rj implies the summation of the associated weights between the hidden neuron j . and input neurons N . IF denotes the proportion of influences of the input variable x i on the output y k [57].

Model performance evaluation
In this research, the ANN models were trained on training data (60%), applied to validation data (20%), and evaluated final model fitted on training data set (20%). The model's simulation accuracy was determined through statistical parameter coefficient of correlation (R), coefficient of determination (R 2 ), mean absolute error (MAE), mean square error (MSE), and root mean square error (RMSE) using Eqs. (5)-(9) respectively.
Here a i is experimental data, p i is simulated data, a and p are mean of experimental and simulated data, respectively, and N represents the number of total samples.

Experimental result
In order to investigate the characteristics of concrete having tin can fiber and waste glass with different ratios tests of workability, fresh density, compressive and splitting tensile strength at 7, 28, and 56 days were carried out. The outcomes of the tests have been displayed in the Figs. 6, 7, 8, 9, 10, and 11. The outcomes of the workability test for incorporation of GWA and TCF are presented in Figs. 6A, 7B, and 8C. The slump height for control concrete was found 98.75 mm. In Fig. 6A, the slump height gets decreased for augmentation of TCF owing to low weight as a consequence of water absorption reduction [48]. Similar lessening of the slump height can be visualized in Fig. 7B for inclusion of GWA from 10 to 30%. This reduction of workability can be imputed to the sharp edges of GWA which lessen the fluidity of concrete mixes [2]. Conversely, in Fig. 8C, workability of all mix proportions containing   GWA and TCF was found to be less than control batch. The results showed that for the combination of 0.5% TCF with 10-30% GWA (from mix M1-M3), the slump value was within 57-45 mm. The slump further decreases as TCF increases within the range of GWA. According to the Fig. 8C, the combination of 0.5%TCF and 10-30% GWA showed a higher slump compared with other amounts of GWA and TCF. Figures 6a, 7b, and 8c demonstrate the concrete's density values immediately after mixing. Fresh density gets significantly reduced with increasing replacement of GWA ( Fig. 7b). This trend is explained by the distinction of particles density between river sand and waste glass aggregate [58]. The similar result was reported by Adaway and Wang [31], Taha and Nounu [59], Topçu and Canbaz [60] which supports the claim that fresh concrete with GWA has a smaller density than conventional concrete. Meanwhile, Fig. 6a shows an increment of density of concrete with an increase of TCF. A similar result has been obtained by Wijatmiko et al. [48]. In Fig. 8c, the range of fresh density of concrete for mixes 1-9 was observed to be from 2557 kg/m 3 to 2500 kg/m 3 . These values of the density were below the density of conventional concrete (Mix M0). Mix 7 having 10% GWA with 1.5% TCF showed highest strength in comparison with other mixes having these two variables. At any content of TCF, the decrement of density was found with the rising GWA from 10 to 30%. The variations of concrete compressive strength having different proportions of TCF and GWA have been presented in Figs. 9A, 10B, and 11C. It can be noticed from Fig. 9A that rise of the strength happen from 13.5 to 14.9 MPa at 7-day, 23.6 to 25.9 MPa at 28-day and 29.6 to 31.9 MPa at 56-day for inclusion of TCF up to 1% but beyond this percentage the strength decreases due to diminishing of concrete's bonding strength by high amount of TCF content [47]. On the other hand, for inclusion of GWA up to 20% the strength shows accretion (Fig. 10B) owing to filler behavior of glass particles [2]. However, following 20% GWA, genesis of voids occur which resulted curtailment of the strength [2]. Moreover, improvement of strength can be seen in the graph of early stage (28 days) compared to later stage (56 days) due to silane effect that supplies more free water in the mix, that is essential to accelerate the hydration process. Moreover, the rise in the 7 days' compressive strength of glass waste concrete mixtures could be attributed to the pozzolanic effect of waste glass. Whereas, less increment of 56 days' strength might be due to reduced dispersion that lessen the hydration process [61]. Taher et al. [8] stated that decrement in compressive strength might be a result of the poor geometry of waste glass owing to a homogeneous distribution of aggregate.
The investigation from Fig. 11C reveals that the compressive strength at all ages displayed similar trends. According to the plot, the highest compressive strengths were recorded to be 15.90 MPa, 26.5 MPa, and 33.1 MPa at 7, 28, and 56 days, respectively. These maximum responses referred to the 20% GWA with 1% TCF at 7 and 28-day respectively, whereas 20% GWA with 0.5% TCF at 56 days. Overall, the 7, 28, and 56 days strength curve show similar trend. Figures 9a, 10b, and 11c demonstrate the influence of GWA and TCF on splitting tensile strength at the three ages. With the addition of 0.5% TCF, the strength increases by 1.13% at 7-day, 1.16% at 28-day and 1.04% at 56-day in comparison with control concrete (Fig. 9a). As the percentage of TCF increases after 0.5%, decrement of the strength was noticed. Whereas, in case of GWA (Fig. 10b), results observed in the range of 0-10% substitution level reflect rise in the splitting tensile strength. However, from 10% replacement of GWA, strength tends to drop due to smooth surface of glass particles [59].
The combined influence of GWA and TCF on splitting tensile strength at all ages can be seen in Fig. 11c. In comparison with control concrete, the higher strength was observed at mix 1-3 which contained 0.5% TCF with 10-30%. Turgut and Yahlizade [24] found similar higher strength than control concrete for 10-30% GWA replacement among all mixes. Mix 2 (10% GWA with 0.5% TCF) yielded the peak results of 2.3 MPa, 3.4 MPa, and 3.6 MPa at 7, 28, and 56 days respectively. The drop in the strength is mainly because to the existence of silane on the surface of the glass that could lessen the adhesion between the cement matrix and glass. Moreover, silane accelerate slippage and decrease the interlocking in the concrete mixtures [61].

Performance of ANN models
Back-propagation neural network having 16 neurons in one hidden layer was adopted to develop ANN model of the mentioned responses in this study. The input was divided into three sets of category in ANN where 70% were assigned for training sets and, testing and validation sets took 15% each.  [62]. In all cases, correlation coefficient (R) values were found close to 1 indicate outstanding performance of the models [63]. The registered MSE, RMSE, and MAE values for workability were 0.003, 0.052, and 0.025 respectively which demonstrate less deviation in prediction [64]. Similar results can be seen in the models constructed for the other responses. Figures 12, 13, 14, and 15 suggest that the developed ANN models can learn the correlation between input and output as the predicted outcomes in all models were close to that of the target outcomes. These figures also demonstrates that ANN model is adequate of simplifying between independent variables and responses variables with predictions of high accuracy [65]. From the figures, it can be inferred that the normally distributed errors in the ANN models as data were nearly distributed on straight lines [64].

Sensitivity analysis of the variables
The major objective of the sensitivity analysis was to explore the comparative influence of GWA and TCF on workability, fresh density, compressive strength and splitting tensile strength of concrete at 7, 28, and 56 days. Table 5 shows mean influence of the independent variables on the desired responses. The outcomes of sensitivity analysis seem to demonstrate that TCF have more influence than GWA on the responses except for fresh density. However, almost equal effect of both parameters can be observed in case of compressive strength at early days. Influence of TCF on the workability, 56-day compressive and splitting tensile strength was 67.38%, 64.33%, and 38.49%respectively, which  are double in magnitude than GWA. Similarly, TCF has four times importance than GWA on 28-day compressive strength and splitting tensile strength. On the other hand, GWA can be removed as its importance in predicting 7-day splitting tensile strength is close to zero (1.26%). The above discussed results implies that inclusion of glass waste and tin fiber have positive effect on the strength of concrete. The results were verified through statistical analysis which indicate that the simulated models are authentic and can be used for navigation. Hence, the promising results for inclusion of glass waste and tin can solve the following potential problems. Firstly, protect the environment since both wastes are non-biodegradable. Secondly, reduce the demand of natural resources as excessive extraction causes serious problems on associated ecosystem, landscape, water tables, and riverbeds. Lastly, virgin lands can be saved that are supposed to be sued for landfill of these waste materials.

Conclusions
In this research, ANN models were developed for the simulation of workability, fresh density, compressive strength, and splitting tensile strength at 7, 28, and 56 days of concrete having tin can fiber and glass waste aggregate. Also, sensitivity analysis of the models was executed to investigate the relative influences of GWA and TCF on each response. However, optimization of variables as well as validation of the predicted results have not been done. To incorporate the bending effect in overall performance of concrete having GWA and TCF contents, the flexural strength could have been considered. The following conclusions have been obtained from above discussion.
• For inclusion of GWA and TCF, workability of fresh concrete have a tendency to drop continuously. • Fresh density of concrete increases with rising TCF but decreases with rising GWA.
• Compressive strength at all ages shows positive trends till 20% GWA and 1% TCF. • In each case of splitting tensile strength, up to 10% GWA and 0.5% TCF strength gets increased. • The constructed models simulated the properties of concrete with good and high values of R 2 . Other parameters also indicate the applicability of the developed model for accurate prediction. • Sensitivity analysis of the models seem to reveal that TCF have more effect than GWA on the properties of concrete.
As a part of recommended future work, further research can be conducted by using various admixtures to increase workability and splitting tensile strength, which would result in sustainable concrete production.