An Artificial Neural Network Based Prediction of Mechanical and Durability Characteristics of Sustainable Geopolymer Composite

School of Civil Engineering, Vellore Institute of Technology, Chennai Campus, Tamil Nadu, India Department of Civil Engineering, Indian Institute of Technology, Guwahati, Assam, India Department of Civil Engineering, Sree Vidyanikethan Engineering College, Tirupati, Andhra Pradesh, India Department of Mechanical Engineering, Saveetha School of Engineering, SIMATS, Chennai, Tamil Nadu, India Department of Civil Engineering, Copper Belt University, Kitwe, Zambia


Introduction
Ordinary Portland Cement (OPC) is commonly used as a traditional binding material in all concreting projects. e manufacturing of OPC consumes a tremendous amount of energy and disperses a huge proportion of carbon dioxide into the Earth's atmosphere. To mitigate carbon dioxide emissions, a new promising binder known as geopolymer was introduced [1]. Numerous researches have been carried out on the e ective and comprehensive utilization of di erent industrial waste materials in the manufacturing process of geopolymer concrete [2]. Its manufacturing process includes the formulation of binders from the alumina and silica rich sources acquired from the industrial byproducts or low-cost materials such as y ash (FA), ground granulated blast furnace slag (GGBS), metakaolin (MK), rice husk ash (RHA), high magnesium nickel slag (HMNS), palm oil fuel ash (POFA), waste glass powder (WGP), red mud, etc. using an alkali activator solution [3][4][5].
e presence of binding material in geopolymer binder is supplemented by industrial/agricultural wastes that comprises pozzolanic characteristics comparable to OPC and abundant in alumina and silica proportions [6][7][8]. In order to extract the silica and alumina sourced from the raw materials, the alkali-activated solution is employed as a catalyst which contains a mixture of sodium hydroxide (NaOH) and sodium silicate (Na 2 SiO 3 ) solutions [9,10]. FA and RHA are industrial byproducts of thermal power stations and rice husk burning, respectively. e principal objective of producing a geopolymer composite from an industrial byproduct is to promote a sustainable alternative for conventional Portland cement concrete by significantly lowering greenhouse gas emissions and industrial waste disposal concerns [11][12][13]. Earlier studies showed that the effectiveness and usage of higher molarity of alkaline solution significantly influence the early strength of the geopolymer concrete [3,14,15]. Literature reported that the required mechanical properties of the geopolymer concrete specimens could be achieved in the ambient curing conditions [16,17]. On the other hand, FA-GGBS based geopolymer binders produced excellent mechanical and microstructural characterizations even after the exposure to elevated temperatures [18,19]. e addition of copper slag in the FA type geopolymer concrete resulted in higher compressive strength results [20,21]. Partial incorporation of RHA with FA in geopolymer concrete resulted in an increase in durability and mechanical strength properties [22,23]. Incorporation of RHA as a source material in the slag-based geopolymer concrete resulted in greater compressive and split-tensile strength results [24].
ANN is based on machine learning framework that simulates a network of biological neural networks. It can be used extensively in the domain of science and engineering to overcome extremely complex problems [25,26]. e ANN framework outperforms other techniques in aspects of nonlinear connection among input parameters [27]. According to recent findings, the ANN structure can be used successfully in the construction and building materials stream to estimate their strength properties with precision [28][29][30]. Khademi et al. employed multiple linear regression (MLR), artificial neural network, and adaptive neurofuzzy inference system (ANFIS) techniques to estimate the 28-day compressive strength of concrete [31]. Apart from mechanical strength, other important parameters like mix design [32], cement content [33], replacement level of recycled coarse aggregates [34], drying shrinkage of concrete [35], slump values [36], etc. can also be predicted with the help of neural networks along with experimental results. Several studies described that the compressive, split-tensile, and flexural (mechanical) properties of FA-based geopolymer matrix are predicted with the application of the ANN framework [37,38].
Although the usage of FA in geopolymer production is significantly reported in the kinds of literature, the use of RHA and fibers is scanty. is experimental investigation is aimed at exploring the influence of fiber and RHA in the fly ash based geopolymer mortar, since the potential use of geopolymer mortar as a repair material for the strengthening of structures. In addition to this ANN framework was developed using Levenberg-Marquardt (LM) Algorithm in MATLAB-2018a to estimate the mechanical and durability strength results of fiber incorporated RHA-FA-based geopolymer mortar.

Materials and Sample Preparation.
In this study, the materials procured for the geopolymer mortar preparation were FA and RHA. e FA and RHA obtained from Kolkata were used as the source materials. Table 1 presents the chemical compositions of the geopolymeric precursor products acquired from X-ray fluorescence (XRF) analysis. Locally resourced river sand with specific gravity 2.5 was used as fine aggregate. e mixture of commercially available sodium hydroxide (flakes type) and sodium silicate (liquid gel type) sourced and supplied by Sharma brothers, India, was employed as an alkaline activator solution. e alkali activator solution was produced by blending sodium silicate solution with a molar ratio (SiO 2 /Na 2 O) of 2.65 and sodium hydroxide. e specific gravity and molar mass of the sodium silicate solution and sodium hydroxide pellets employed were 1.52 and 2.14 and 123 g/mol and 38.8 g/mol, respectively. e source materials present in the FA and RHA geopolymer mortars were enhanced by the alkali activator solution. Commercially available PP fibers and sulfuric acid were used. Figure 1 illustrates the visual appearances of the geopolymeric source materials (RHA and FA) and PP fibers used in this investigation.
A partial replacement of FA was carried out using RHA (0%, 10%, and 20%) with the addition of polypropylene fiber of 0.0%, 0.1%, and 0.3% by volume and mixed thoroughly with alkaline activator solution to obtain uniform slurry. Fine aggregate was then introduced to the slurry in the ratio of 1 to FA to obtain geopolymer mortars. Geopolymer specimens were prepared in two layers using 70 × 70 × 70 mm cubes and vibrated for about two minutes in table vibrator to remove the entrapped air present in mortars. e geopolymer mortar specimens thus prepared were cured in the hot air oven for about 24 hours at the temperature of 110°C and then kept in the ambient conditions until further testing.

Experimental Approach.
e various mix proportions considered for the experimental investigations on FA-RHA geopolymer mortar influenced with polypropylene fibers are listed below. e mixes with varying % of RHA and the  Table 2. e strength reported was the average of three identical specimens. To achieve the preferred workability in all mixture proportions, the alkaline to binder ratio and sodium silicate to sodium hydroxide fraction were selected to 0.5 and 2.5, respectively. For all of the mix composition, the quantities of fine aggregates, binder content, sodium silicate solution, and sodium hydroxide flakes selected were 600 kg/m 3 , 600 kg/m 3 , 257.15 kg/m 3 , and 102.85 kg/m 3 , respectively.
Universal Testing Machine (UTM) was used for calculating the uniaxial compressive strength of geopolymer mortars at 28 days as per IS 516 (1959) provisions. e flexural strength characteristics of fiber reinforced FA-RHA based geopolymer mortar specimens were ascertained in accordance with IS 516 (1959) standards using a Universal Testing Machine of 1000 kN capacity [39]. Prism samples of size 40 × 40 × 160 mm were casted and examined for flexural performance after 28 days. e method monitored for the determination of water absorption of geopolymer samples was in accordance with ASTM C 642 standards [40]. After measuring the weights of 28-day-old geopolymer mortar samples, they were dried at 110°C for 24 hours before being immersed in water. e specimens were then removed from the water and wiped clean and directly weighed in saturated surface dry conditions to find an increase in weight.
Existing literature reports proved that geopolymer binders were acid resistant, providing them a promising and alternative construction material for the sewer environment.
is study examines the durability of FA-RHA based geopolymer mortars subjected to 10% sulfuric acid concentration for 56 days and tested for its strength according to ASTM C 643 standards [41]. e rate of capillary rise absorption by mortar cube is ascertained by the sorptivity test. e samples are initially coated with waterproof enamel paint on all sides except the bottom and top surfaces, so as to allow capillary uptake of water only from the bottom. e specimens are then conditioned at 110°C for 24 hours to obtain constant mass. Test samples are made to rest on supports (a supporting wire mesh in the present case) in a manner such that only the lowest 2 to 5 mm of the cube is underwater. e rise in the mass of the sample with time is noted. en water uptake per unit area of concrete surface I (g/mm 2 ) is plotted with the square root of time for the suction periods (t). Hence I � C + St 1/2 where I � increase in mass per unit area (g/mm 2 ); t � time, measured in minutes at which the mass is determined; S � sorptivity in g/mm 2 / min 0.5 ; C � a constant.

Prediction of Strength and Durability Characteristics Using ANN
ANN is indeed a massively simultaneous computing intelligence processing architecture which operates equivalent to biological neural systems [42]. It also has the ability to comprehend and extrapolate mostly from provided   information and intended to deliver appropriate responses even though the group of input variables comprises an inconsistency or is ambiguous [37,43]. It comprises several interlinked engineered neuron-like structures, each of which delivers a distinct response (Y) from most of the inputs (X j ) across equation (1) [44]. e activation function (f ) is associated with the sum of input parameters procured from the sum function and determines the neuron's output. Phrase (H) illustrates the amount of the input parameters that can be anticipated using equation (2), and "b" is the bias coefficient, which is applied to influence the activation function.
Since the ANN framework constitutes three components, it can be regarded as a Multilayer Perception (MLP) structure, as illustrated in Figure 2. e first layer (input layer) contains three independent variables (RHA/FA ratio, different molarities, and percent of fibers) that are used for entering data. e second layer is regarded as the hidden layer or computational layer, whereas the third layer is recognized as the output layer, from which ANN model estimates compressive, flexural, water absorption, acid resistance, and water sorptivity values.
Different variables such as RHA/FA ratio, varying concentrations of NaOH solution, and percentages of polypropylene fibers have a significant impact on the strength and durability characteristics of geopolymer mortar mixes [45,46]. Hence, the RHA/FA ratio, different molarities, and percentages of polypropylene fibers were preferred as input parameters for the geopolymer mortar mixes, and the target variables were compressive strength (CS), flexural strength (FS), water absorption (WA), water sorptivity (WS), and acid resistance (AR) of geopolymer mortar specimens.
e overall amount of hidden compartments and the number of neurons in every hidden compartment in the ANN structure could be ascertained through implementing the handful of assessments throughout the training and testing period until the desired outcomes are achieved with negligible error values. e LM algorithm was implemented in ANN model with feedforward backpropagation technique to estimate the durability and mechanical properties of geopolymer mortar using an ANN model with two hidden layers and five neurons in each layer. Out of 27 experimental test results, 19 were selected for training, 4 for testing, and 4 for validation phase. e limits for input and output responses considered for this study are listed in Table 3. e accuracy of the output responses recorded from the Input layer Hidden Layers Output layer Mix ID  G10  G15  G23  G11  G14  G20  G19  G26  G22  G21  G16  G24  G25  G12  G18  G17  G27  G1  G8  G6  G4  G9  G2  G7  G5 G4  G5  G6  G7  G8  G9  G10  G11  G12  G13  G14  G15  G16  G17  G18  G19  G20  G21  G22  G23  G24  G25  G26 G16  G12  G26  G18  G17  G21  G15  G24  G11  G25  G20  G27  G22  G23  G19  G14  G10  G6  G5  G2  G8  G4  G3 G7 G9 G1 Mix ID  Acid resistance at 56 days      [15,49]. A higher concentration of NaOH solution enhances the solubility of Al and Si ions from the precursor materials, resulting in the generation of relatively strong Si-O-Al, C-A-S-H, and N-A-S-H gels which resulted in the increase in strength properties. However, the different RHA substitution levels have no effect on the development of flexural strength in FA-based geopolymer mortars. e increased proportion of RHA results in a significant concentration of unreacted RHA granules in the geopolymer mixture, resulting in a relatively weak and less ductile geopolymer matrix. e enhanced quantity of SiO 2 disruptions the interaction of Si and Al particles ultimately results in a lesser density geopolymer binder with lower flexural strength [50].  G25  G22  G21  G19  G18  G17  G16  G15  G10  G14  G26  G27  G12  G11  G23  G20  G24  G13  G3  G5  G9  G2  G6  G4  G7  G8 G22  G21  G20  G26  G19  G18  G17  G24  G10  G11  G12  G13  G14  G15  G16  G23  G27  G25  G6  G4  G8  G2  G3  G1  G7  G5  G9   Mix ID   20   25   30   35   40   45 Compressive sterngth in 10% H 2 SO 4 (MPa) Figure 12: Comparison of experimental and predictive acid resistance results. Figure 5 demonstrates the water absorption test results after 240 mins for oven cured FAbased geopolymer mortar specimens with varying levels of RHA (0%, 10%, and 20%), polypropylene fibers (0%, 0.1%, and 0.3%), and NaOH (5 M, 10 M, and 15 M). From Figure 5, it can be observed that G7, G8, G9, G16, G17, G18, G25, G26, and G27 geopolymer mixes with 0.3 percent PP fiber incorporation had lower water absorption than the other mixes. e behavior of the PP fibers restricts the formation of microcracks, which reduces the water absorption capacity of the mortar mixes. Furthermore, the PP fiber's nonabsorbability (hydrophobicity) nature contributed to a decrease in waster absorption capacity [51]. Moreover, introducing 20% RHA replacement levels to FA-based geopolymer mortar resulted in increased water absorption test result compared to other combinations.   Advances in Civil Engineering geopolymer mortar mixes constituting 80% FA and 20% RHA showed higher rate of water absorption due to the inferior properties of RHA particles such as higher water absorption capacity than FA [50].

Prediction of Strength and Durability Properties Using ANN
e percentages of error values for the strength and durability characteristics of geopolymer mortars obtained from ANN model were listed in Table 4. As seen in Table 4, it can be stated that the maximum percentages of error observed for the geopolymer mortar mixes under compressive, flexural, water absorption, sorptivity, and acid resistance test results were found to be 2.94%, 4.0%, 5.40%, 6.36%, and 6.15%, respectively. e error values obtained from (3) are negligible as the error percentage for all the predicted values is less than 10 percentage. According to the preceding sentence, the ANN framework could be utilized to estimate the mechanical and durability characteristics of fiber influenced FA-RHA-based geopolymer mortars. e comparison between the experimental and the predicted values of compressive strength results for fiber influenced FA-RHA based geopolymer mortars is expressed in Figure 8. Figure 9 depicts the correlation among the predicted and experimental flexural strength results. In case of water absorption test results, the variation between experimental and predicted values is illustrated in Figure 10. Consequently, Figures 11 and 12  e cumulative coefficient of correlation (R) for compressive strength results at the stage of training, validation, testing, and the association of three levels in the ANN framework was measured as 1, 0.97189, 0.95547, and 0.97808, as seen in Figure 13. For flexural outcomes, the calculated R values throughout training, validation, testing, and the combination of three-phased convergence were computed from Figure 14 as 0.98670, 0.99819, 0.98673, and 0.97317, respectively. Figure 15   Advances in Civil Engineering the water sorptivity and acid resistance results in the form of R values. R values larger than 0.9 explicitly indicate a strong association among the observed and simulation outcomes across all instances [37,38]; the developed ANN structure, which has been performed using measured data, precisely anticipated the intended outputs. Data Availability e data used to support the findings of this study are included within the article. Should further data or information be required, these are available from the corresponding author upon request.

Conflicts of Interest
e authors declare that there are no conflicts of interest.