Predictive Assessment and Compressive Strength Test of Geopolymer Concrete Mix as a Partial Replacement to Conventional Concrete Mix with Artificial Neural Network Application

1. Introduction

The construction industries in the Philippines mostly utilize concrete when it comes to infrastructure projects. However, there are many downsides to conventional concrete (CC) mixes utilized nowadays. The materials employed in some of the current construction materials and the waste produced are not environmentally friendly. The global carbon footprint has been significantly impacted by the manufacture of ordinary portland cement (OPC), leading to ongoing environmental issues [1]. Conventional concrete mixes (CCMs) are among the most popular building materials utilized worldwide to develop structures, including buildings, infrastructures, roads, and bridges. Despite being a good building material, ordinary portland cement production produces a lot of CO2 [2]. On the other hand, several studies have shown that geopolymers have remarkably lower Carbon dioxide emissions [3]. According to the study of Bagci et al., geopolymer concrete (GPC) cylinders also has higher compressive strength and low shrinkage quick setting, acidity and heat resistance, low thermal conductivity, and other desirable properties [4].

In 1979, Davidovits initially proposed geopolymers made of semi-crystalline, three-dimensional aluminosilicate components. When a highly concentrated aqueous alkali hydroxide-silicate solution is applied to an aluminosilicate raw material, including slag, fly ash, calcined clay, etc., geopolymer, an amorphous, 3-D inorganic polymer, is created. Compared to Portland cement, geopolymers produce 80% less CO2 [5]. Geopolymer innovation is still in its early stages. Because these combinations are sensitive to several factors, using them correctly requires training.

Geopolymer concrete mix (GCM) has the potential to exhibit better mechanical properties than traditional concrete, including higher compressive strength, lower permeability, and greater durability [6]. However, the production of GCM is still experimental, and predictive assessments must be needed to determine its mechanical properties. Artificial Neural Networks (ANNs) have shown great potential for modeling and predicting complex systems, including the mechanical properties of concrete. ANNs can be used to model the relationships between various factors and the mechanical properties of GCM, including its constituent materials, curing conditions, and mixing proportions [7].

The proposed research aims to develop an ANN-based predictive model to assess the compressive strength of the GCMs. Using ANNs for the predictive assessment of GCM provides an efficient and reliable method for determining its mechanical properties. The developed ANN model can optimize the GCM design, reducing the need for expensive and time-consuming experimental testing. Moreover, the model can be used to assess the effects of various factors on the mechanical properties of GCM and optimize its production process [8].

As shown in Figure 1, the research started by identifying the mechanical properties that affect the compressive strength of the GCM, including the proportion of cementitious materials, aggregates, alkaline activators, etc. These factors are used as input variables in the ANN model, while the compressive strength is the output variable.

1.1. Objectives

This section may be divided by subheadings. It should provide a concise and precise description of the experimental results, their interpretation, as well as the experimental conclusions that can be drawn.

Furthermore, this research aims to fulfill and accomplish the following specific objectives:

• Justify how ANNs can contribute to predictively analyzing and determining the compressive strength of geopolymer concrete;
• Refer to the related literature on various geopolymer material properties and compressive strength;
• Conduct experiments for prototyping geopolymer concrete and compressive strength tests;
• Use neural networks to effectively determine the compressive strength of geopolymer concrete aligned to the experimental data; and
• Analyze the approaches of neural networks in determining the compressive strength of GPC.

1.2. Conceptual Framework

As shown in Figure 2, the conceptual framework exhibits the initial to final phases of the research that the study undertook in achieving the objectives, desired result, and conclusion.

2. Materials and Methods

2.1. Research Materials

Research materials are essential for the preparation and testing of GPC. The materials include alkaline activators, supplementary cementitious materials, aggregates, and other additives required to produce the GPC mix. The equipment consists of tools for mixing, casting, curing, and testing concrete specimens. These research materials are employed to investigate GPC’‘s’ properties, strength, and performance, enhancing our understanding and advancement of this alternative construction material. It also includes sodium hydroxide, sodium silicate, superplasticizer, fly ash (Class F), water, mixed aggregate (sand and gravel), cooking spray, and a wooden spoon [9].

2.2. Data Gathering Procedures

The research problem and objectives are determined to ensure that the data gathered is written in an organized manner and taken into account during the experiment. This aids in determining what data needs to be collected and what techniques are needed to gather it. Then, the researchers determine the sample size, which, in this study’‘s’ case, involves identifying the GPC mix and CC mix samples that will be tested.

Next, the appropriate data collection methods must be chosen based on the research problem and objectives. In this study, experimentation was considered a data-gathering method. Then, the data collection plan must be developed, which includes a detailed description of the data collection methods, data sources, and data collection tools to be used.

The data collection tools and methods were developed based on the sources reviewed in the literature to ensure their effectiveness and reliability. This aids in identifying any issues that may arise during data collection and addressing them before starting the actual experimentation. During the actual experimentation, the collected data must be accurate and consistent. After collecting the data, the researchers analyzed them using the proposed statistical method, ANNs, to determine the geopolymer and CC mix’‘s’ predictive assessment and compressive strength.

2.3. Research Design

Various data sources are utilized to evaluate the properties and performance of the GPC mix. These sources include water absorption tests, concrete curing tests, and concrete compressive strength tests. Each of these tests provides specific information related to the particle size distribution, porosity, moisture absorption, grading, curing conditions, and load-bearing capacity of the GPC. The data obtained from these tests helps assess the workability, durability, and overall strength of the GPC, contributing to a comprehensive understanding of its suitability as an alternative to CC. The study aims to gain insights into the characteristics and behavior of GPC, facilitating its further development and application in the construction industry.

2.4. Geopolymer Concrete Experiment

The experimentation to manufacture GPC was based on relevant references, such as articles containing information regarding GPC prototyping [9,10,11]. GPC is said to be the most improved type of concrete after OPC concrete. To prove this, experiments justifying the advantages of GPC, such as greater strength and durability, are necessary. Moreover, in conducting several tests, prototyping the GPC sample is required to produce cylinders.

The researchers conducted an experiment to produce 14 GPC samples and 2 CC samples. Table 1 below shows the density in kg/m3 of ingredients and components in each concrete sample from GPC1 to CC2.

2.5. GPC and CC Casting

The fine aggregates were then measured according to the designated percentage. The mix percentage of each cylinder is shown in Table 2. These mix ratios were strategically chosen to evaluate their combined properties, aiming to find an optimal balance between strength, durability, and sustainability. Specifically, it ranges from pure geopolymer (100% GP Mixture) to pure (100% OPC Mixture), with various intermediate ratios (e.g., 70%/30%, 50%/50%, 40%/60%, 15%/85%). These variations were selected based on a combination of factors, including preliminary experimental results, and practical considerations such as workability and setting time.

2.6. Water Absorption Test

To assess the porosity as well as the permeability of the mixture, water absorption tests are carried out on the concrete samples. These tests aid in evaluating the concrete’‘s’ efficiency and longevity under actual conditions, particularly in areas where humidity is a concern. The concrete specimens were then weighed after removing from the molds weighing their dry weight before submerging them in water for 28 days, this guarantees that water is completely absorbed into the pores of the concrete. After 28 days, the researchers proceed to surface drying, the specimens are weighed afterward. The percentage of water that the concrete absorbed throughout the submerged period is shown by the weight increase [12]. Generally, to compute the Water Absorption, Equation (1) is used:

WA% = ((Weight of water absorbed)/(Dry weight of the sample))*100%

(1)

Water absorption is determined by dividing the weight of water absorbed by the sample’‘s’ dry weight, stated as a percentage.

2.7. Compressive Strength Test

This is performed to determine GPC’‘s’ load-bearing capacity and strength characteristics. It involves subjecting cured concrete specimens to increasing compressive loads until failure occurs. The maximum load applied and the corresponding failure point are recorded to calculate the compressive strength. This data is crucial for assessing GPC’‘s’ structural performance and suitability in practical applications. The step-by-step procedure for this test is according to the ASTM C39 and other relevant references [13,14].

Figure 3 shows how ASTM’‘s’ C39 is done, illustrating the fracture patterns that occur when a concrete sample is compressed.

2.8. Artificial Neural Network

This ANN system aims to provide a reliable and accurate prediction model for concrete compressive strength evaluation. The statistical treatment involved rigorous analysis and optimization of the ANN architecture, training parameters, and input variables to ensure optimal performance and generalizability. By implementing appropriate statistical techniques, such as data preprocessing, feature selection, and model evaluation, the researchers aimed to create an ANN system that produces accurate predictions of concrete compressive strength and serves as a foundation for future researchers in the field [15].

The statistical treatment allows the identification of significant variables and the determination of their respective contributions to the overall predictive performance of the ANN model.

Developing this ANN system with a robust statistical treatment enhances its potential for broader applicability and generalization. It provides a reliable framework that can be used as a benchmark for future studies on concrete compressive strength prediction using ANNs. [16] The formulation of this working ANN system contributes to advancing research in the field, facilitating further investigations and enabling the comparison and validation of future models and methodologies. [17]

The validity of the strength model was verified using actual experimental data and experimental data from various sources. Test results were compiled for concrete, including fly ash, coarse and fine aggregates, GGBFS, and cement. It was decided to ensure these mixtures presented all the data needed for such an assessment and were a corresponding group controlling all the key factors affecting GPC strength. Given that 16 samples were produced in the experiment for the study, the current researchers also examined 1047 concrete samples from multiple studies and added them. [18,19,20,21,22] As shown in Appendix A, the following data are used as the input and output parameters for the ANN model.

The researchers utilized MATLAB to develop and assess the neural network model for this study. The independent variables included Fly Ash, Sand, Gravel, GGBFS, Portland Cement, Superplasticizers, Sodium Hydroxide, Water, and Sodium Silicate. The dependent variable was the concrete samples’’ Compressive Strength (MPa).

In MATLAB, the researchers implemented various neural network architectures to determine the most accurate model for predicting the compressive strength of GPC. The researchers ended up using the Neural Network Fitting of the software. The data was split into training, validation, and testing sets in the following proportions: 70% training data, 10% validation data, and 20% test data. This ensured the model’‘s’ integrity and accuracy. Training involved adjusting the network parameters to minimize prediction errors, while validation helped run the model to prevent overfitting. The testing phase evaluated the model’‘s’ performance on unseen data, ensuring its reliability.

Critical parameters for the neural network included a layer size of 15 using Bayesian Regularization. Bayesian Regularization was employed to improve generalization by reducing the risk of overfitting, especially with complex models.

The key performance parameters include Mean Squared Error (MSE), which measures the average squared difference between predicted and actual values, and the R-value, indicating the strength of the linear relationship between these values. Training state parameters consist of the gradient, reflecting the rate of change of the loss function; Mu, related to Bayesian Regularization to prevent overfitting; the number of parameters, representing the model’‘s’ complexity; sum squared parameters, which help promote simpler models by penalizing large weights; and validation checks, ensuring consistent performance without overfitting. Additionally, experimental parameters involved various ratios of geopolymer concrete (GPC) and Ordinary Portland Cement (OPC) mixes, such as 100% GPC, mixtures with different proportions of GPC and OPC, and 100% OPC. These parameters collectively contribute to evaluating and optimizing the ANN model’‘s’ ability to predict the compressive strength of geopolymer concrete mixes.

3. Results

3.1. Casting Observation and Outcomes of the Specimens

Proportions were made to test each component’‘s’ effectiveness in increasing the concrete’‘s’ compressive strength. Regarding proportionality, GPC1-GPC6 followed the 3-3-1 mix ratio, while GPC8-GPC14 was based on the 3-2-1 mix ratio. Considering all the factors, the following observations are made.

According to Table 2 in geopolymer concrete (GPC) and Conventional Concrete (CC) casting, GPC5-GPC6 containing 30% geopolymer Mixture and 70% Conventional Mixture and GPC8 containing 60% geopolymer mixture and 40% Conventional mixture, have the highest compressive strength observed in Table 4. These observations aid the researchers to associate the mixing proportions to its respective compressive strengths and identify the most effective proportions in achieving the desired results.

Figure 4 shows the outcome of the concrete cylinders GPC 1, GPC 3, GPC 5, GPC7, and CC1 after 36 hours of casting.

The ratio of cement to aggregates affects the strength of concrete. Too much water or cement can weaken concrete and make it fragile. A proper mix ratio guarantees that the concrete reaches its optimal strength. Concrete constructions can last longer by using the right mix ratios, especially in tough settings. For instance, a mix that contains more cement might be less vulnerable to chemical deterioration. [23]

3.2. Water Absorption Test Results

Table 3 shows the lightness and heaviness of concrete cylinders and the outcomes of the Water Absorption test by weight and percentage.

When all factors are considered, water absorption tests generate important information that aids in evaluating the general performance and overall quality of the concrete, guaranteeing that it satisfies the requirements for the purpose for which it is designed.

3.3. Compressive Strength Test

This study also aims to achieve a 20-30 mpa compressive strength in each sample because it is under the recommended concrete strength [24,25]. While some failed to achieve this, three samples out of 16 achieved a higher compressive strength than others. This aided the researchers to observe what factors to consider for higher compressive strength. Table 4 shows the number of days and the compressive strength of each sample tested in the lab.

Adding OPC to GPC’‘s’ mixture is essential because it can change its strength, setting time, and durability. Thus, compressive strength testing is conducted to comprehend the impacts of the cement additive and improve the mix design for applications.

It was clear from evaluating the samples and assessing the data that they satisfied the requirement and had appropriate handling, consistent material composition, and precise testing methods. On the other hand, there was evidence of variability in one or more of these areas in the samples that did not reach the criteria. Moreover, a thorough assessment of the material procurement, mix design, and testing procedures is advised to find and fix any strength issues. Doing this guarantees that every sample satisfies the requirements and enhances the consistency and dependability of compressive strength.

3.4. Artificial Neural Network: Fitting

Several graphs and results were considered to evaluate the model’‘s’ performance comprehensively:

A.

Performance. The performance plot demonstrates that the ANN model effectively learns and predicts the compressive strength of GPC. As shown in Figure 5, the best training performance is 5.3408 at epoch 658, which shows that the model has been trained with minimal errors.

B.

Error Histogram. This provided a visual representation of the distribution errors between the predicted and actual compressive strength values. As shown in Figure 6, this helped understand the model‘s’ accuracy and identify any significant outliers. The highest distribution is between -1.015 to 0.3587. The errors are mostly small and clustered around zero, indicating the model’s predictions are accurate and reliable.

C.

Regression. Figure 7 displays the correlation between the predicted and actual compressive strength values for the training, validation, and sets. High correlation coefficients (close to 1) indicated predictive solid performance. In this case, the training set resulted in R=0.98373, Test to R=0.88164, and All set to R=0.96982. The training set shows a very strong correlation between predicted and actual values during training. The test set, while lower than the training set, still shows a good correlation.

D.

Training State. Figure 8 shows the progress of the training process, including the following:

a.

Gradient (0.49901 at epoch 659). This suggests that the training process is converging well and the model is reaching its optimal state. This is important for ensuring that the model is not overfitting or underfitting and is learning effectively.

b.

Mu (50000000000 at epoch 659) indicates strong regularization, ensuring that the model generalizes well to unseen data by avoiding overfitting, which keeps it smart and balanced.

c.

Num Parameters (110.7105 at epoch 659). The value here indicates a balanced model complexity suited for the dataset size and problem.

d.

Sum Squared Parameters showing 183.1633 at epoch 659 ensures that the model is not overly complex, which is important for consistent performance across different datasets.

e.

Validation Checks (0 at epoch 659). Zero validation checks imply that this model consistently performed well on the validation set, indicating a stable training process without overfitting.

Following this, the data is ready for the neural network to be trained. To guarantee that the data is appropriate for training, this involved preprocessing procedures such as feature selection and scaling.

The study demonstrated that neural networks, when properly trained and validated, could effectively predict the compressive strength of GPC based on the given independent variables. The use of MATLAB provided a platform for the development, training, and assessment of these neural network models, contributing to the advancement of predictive analytics in concrete technology.

4. Conclusions

In this study, GPC and CC specimens were made by applying varying proportions of fly ash, OPC, and GGBFS. NaOH (14 Molarity) and Na2Si2O3 were the alkaline solutions employed. The samples were cured for 28 days at room temperature before tests were performed to determine their compressive strength and water absorption. The data was analyzed, and the following conclusions were made.

Increasing GGBFS enhanced the strength and prolonged the binding time at room temperature. The binding period was almost three days for GPC1, GPC2, GPC3, and GPC4, but only twenty-four hours for GPC5-GPC6 and CC1-CC2. When designing a GPC mix, determining the mass percentage of water to geopolymer solids is crucial; adding too much water will weaken the concrete.

GPC proves to have better workability when properly mixed with OPC. In conclusion, GPC Mix is more effective as an additive to Ordinary Concrete Mix. Not only are GCMs more eco-friendly, but they are also less expensive than OPCs. Incorporating this study will indeed promote sustainable construction and improve the construction industry‘s’ resourcefulness.

Moreover, ANNs have been proven effective in assessing and predicting the compressive strength of concrete. Based on our analysis, the Mean Squared Error (MSE) varies from approximately 5 to approximately 33, which indicates a gap. However, other performance metrics suggest that the model is in a good state, and the high R-value justifies the selection of these results. To further enhance the model‘s’ accuracy and reduce errors, the researchers recommend conducting additional training. Given the data set contains over a thousand samples, more extensive training can help minimize errors. Future studies should pay close attention to the values throughout the training process to ensure consistent and reliable performance. ANN models help evaluate and forecast concrete strength based on various input parameters since the models can capture complex links and nonlinear patterns in the data.