ANN-based strength prediction of concrete with nano silica, glass, and coir fibers

ABSTRACT

Ensuring improved strength and durability in concrete has become significant in modern infrastructure. This study presents Artificial Neural Networks for predictive analysis to investigate how glass fibers, coir fibers, and nanosilica affect the mechanical properties of concrete. Incorporating 3% nanosilica with varying proportions of glass, coir fibers (0.3 – 3%) revealed significant improvement in the concrete’s performance. At 1.5% concentration glass fiber increased compressive strength by 15%, flexural strength by 35.6%, and tensile strength by 43.5%. Similarly, 1.2% concentration, coir fibers improved these properties by 9%, 16.4%, and 32.5% respectively. The mean square propagation, testing, training, validation results are used to prepare the ANN model. For glass fibers, the results are 0.8772, 0.9734, 0.9828, and 0.9563, and the optimal validation results is 0.0107 at epoch 10; for coir fibers, the results are 0.9743, 0.9949, 0.9941, and 0.9909, and the optimal validation results is 0.0220 at epoch 8 correspondingly. To further validate the model’s reliability, additional statistical metrics are Willmott’s Index of Agreement was found to be 0.9953, the Nash–Sutcliffe Efficiency was 0.9814, and the percent bias was –0.0083%. These results confirm the robustness and practical applicability of the proposed ANN model in predicting concrete strength parameters with high fidelity.

Keywords:
Nano silica; Glass fiber; Coir fiber; Artificial Neural Network; Concrete Strength Prediction

1. INTRODUCTION

Concrete is one of the most popular building materials worldwide, owing to its range of applications, relatively low cost, and high strength. However, plain concrete has relatively low tensile strength and ductility, directly affecting its resistance to cracking. To combat these limitations, various approaches have been investigated, including the addition of natural and synthetic fibers into concrete mix. Coir fiber (derived from coconut husks) and glass fiber have proven to be good examples for enhancing the mechanical properties of concrete. Additionally, nano-silica, an extremely reactive pozzolanic material, has been shown to improve the microstructure and strength of concrete. This study aims to evaluate environmental sustainability requirements through nano-silica (as a sustainable additive) and effectiveness through coir fiber and glass fiber reinforcement, targeting an environmentally friendly construction material. Nano-silica has gained heightened significance due to its potential in enhancing the mechanical, durability, and microstructural characteristics of concrete. Due to its increased pozzolanic reactivity and filler effect, which refines the pore structure and densifies the interfacial transition zone, the strength development in concrete is enhanced, demonstrating improved mechanical attributes [¹,²,³]. Studies reveal that incorporating nano-silica between 2–3% showed a 20–25% increase in concrete properties. However, beyond the specified limit, the addition of nano-silica resulted in microcracking and increased porosity [⁴]. Furthermore, studies on concrete infused with glass fiber revealed that at optimum levels, both flexural and tensile responses are improved, with less pronounced impact on compression behavior [⁵,⁶,⁷,⁸]. The addition of glass fiber has considerable impact on concrete’s ability to resist cracking and withstand dynamic loading conditions [⁹]. At 1–3% addition by weight of cement, glass fiber boosted the mechanical properties of concrete to the desired level [¹⁰]. Without requiring any admixtures, 0.5–1% addition of glass fiber showed enhanced mechanical behavior [¹¹]. Conversely, coir fiber, being one of the most natural, cost-effective, and widely available materials, helps improve concrete’s mechanical behavior. Observations and previous studies reveal that 2–4% addition of coir fiber resulted in better performance under compression [¹²–¹³]. Furthermore, in combination with nano-silica, 0.25% addition of coir fiber increased concrete’s abrasion resistance [¹⁴]. However, being a natural fiber, it must be initially alkali-treated before incorporation into the concrete matrix [¹⁵]. Addition beyond the optimum value resulted in increased flexural response while depreciating other mechanical properties [¹⁶]. To make strength assessment more accurate, several analytical, empirical, and artificial intelligence methods are deployed [¹⁷]. In this scenario, strength prediction has been greatly assisted by machine learning approaches like Artificial Neural Networks (ANNs) [¹⁸]. Using ANN’s capability to identify and capture nonlinear relationships in large data sets, concrete strength prediction has been achieved with greater accuracy and minimal error, functioning better than other mathematical equations [¹⁹]. Several models have been deployed under ANN for acquiring accurate and reliable results [²⁰,²¹,²²,²³]. TAK et al. [²⁴] undertook the development of an advanced concrete strength prediction system by integrating gradient boosting and neutral attention mechanism which achieved a 27% margin reduction on the Mean Absolute Error against previous benchmarks and exhibited superior performance across diverse formulations, especially with high-performance mixes and early-age strength prediction leading to ratios of lower cement content optimized designs which retained the required strength properties and set groundwork for predicting properties of new concrete formulations with minimal experimental data [²⁴]. Yet another work focussed on the leveraging the efficiency of multiple Models like ANN, SVM and tree-based methods under a collective term “Hybrid ensemble surrogate models” which demonstrated remarkable prediction accuracy providing improved reliability and interpretability compared to the standalone models [²⁵]. Further studies integrating Random forest, neural Network and SVMs assisted in mapping the relationship between the mechanical properties of concrete and the agricultural waste content revealing the relations that traditional regression methods failed to capture enabling mix design predictions based on the performance requirements and available waste material [²⁶]. Data driven approaches have also been applied to determine the durability properties and self-healing mechanism in concrete where the feature extraction algorithm identifies the properties that affect the self-healing process and provides insights for mix design optimization [²⁷]. Additionally, the behaviour of fiber reinforced concrete at elevated temperatures also have been studied using ML algorithms and highlights the role of hybrid models in effective prediction accuracy for critical properties like residual strength, spalling behavior, and thermal conductivity changes in concrete [²⁸]. Another work focussed on the network included an output layer with one neuron representing the compressive strength and two hidden layers with ten and six neurons, respectively, using the ReLU (Rectified Linear Unit) activation function [²⁹]. This paper attempts using Artificial Neural Networks to predict the concrete strength, focusing on nano-silica incorporated with coir fiber and glass fiber. The work depicts the determination of optimum proportions of these ingredients while relying on the capacity of ANN to predict concrete strength characteristics. The TANNP technique greatly improves the accuracy of concrete compressive strength forecasts while successfully addressing the data-hunger problem that deep learning models inevitably face [³⁰]. AI and Machine Learning models are advanced regression approaches designed to overcome the liitations of traditional simple regression techniques [³¹]. The growth in data availability, coupled with the advancement of computer systems capable of processing datasets more quickly and accurately than humans, has significantly propelled developments in the field of artificial intelligence [³²]. The journal is structured in such a way that the materials and methods section elaborate the mechanical testing procedures and the Artificial Neural Network framework developed for strength prediction along with the materials used followed by the results and discussion section which elaborates the key findings and analysis of the results obtained. Finally, the conclusion section summarises the key findings of the work. Despite the growing use of AI in concrete prediction, limited studies have explored hybrid fibre-reinforced concrete combined with nano silica using ANN models. This study aims to develop a robust ANN model to predict the mechanical properties of sustainable mixes. The work addresses the gap in integrating natural and synthetic fibers with nano additives using a machine learning approach.

2. RESEARCH SIGNIFICANCE

Current methods for predicting concrete strength in sustainable additive applications are inaccurate and are heavily reliant on time-consuming physical testing that restricts practical implementations in a real-world setting. This research represents the current state of the art by integrating Artificial Neural Networks (ANN) with applied sustainable concrete materials and science to provide a very accurate strength prediction without having to go through an exhaustive experimental program. This study uniquely shows how machine learning can successfully determine the complex nonlinear interrelationships of nano silica (NS), glass fibers (GF), and coir fibers (CF) in concrete mixtures and provides a reliable computational method for optimizing concrete mixtures. These outcomes contribute greatly to available knowledge in geotechnical engineering and create a baseline method that reduces material testing costs, increases speed of mixture design optimization, and promotes further investigation into environmental beneficial materials while still meeting structural integrity and performance. This study aims to develop an ANN model to predict the mechanical properties of concrete incorporating 3% nano silica with varying glass and coir fiber content. The novelty lies in combining hybrid fibers with nano silica for optimized strength using machine learning. The novelty lies in the combination of these three additives and the use of ANN to simultaneously predict multiple strength metrics, which has not been explored in existing literature.

3. MATERIAL AND METHODS

3.1. Materials

Concrete mix of M20 Grade comprising of OPC 53 grade conforming to IS 12269:2013 with specific gravity of 3.15 and consistency of 32%, Zone II Fine aggregate conforming to IS 383:2016 specifications having specific gravity of 2.65 and fineness modulus of 2.85 and Crushed stone with nominal maximum size of 20 mm, conforming to IS 383:2016, with specific gravity of 2.70 was utilized as coarse aggregate for conventional concrete. Glass fibers of 12 mm length and 14 microns’ diameter, having specific gravity of 2.68 and tensile strength of 3500 MPa and Natural coir fibers of 50 mm length with specific gravity of 1.15 and tensile strength of 140 MPa were also used in prepping the samples required. The water absorption value of coir fibre was 8%. The hydroxyl group in coir fiber, which has a propensity to absorb moisture, increased the amount of water absorbed. Since the influence of water absorption had no effect on strength, the workability of concrete is especially impacted by the water absorption of coir fiber. Furthermore, nano silica is used in combination with the fibers for further enhancing the obtained strength results. It is a white colored powder and odorless powder, amorphous solid with a density of 1.648 g/cm³. Purity is 99.9%, particle size is 20 nano meter and silica content is 98.8%. Surface area is 202m²/g. They exhibit excellent thermal stability and low density, making them suitable for a wide range of applications. Nanosilica is mixed by two step processes involving dispersion and stabilization. Initially NS is mixed with water in a beaker and it is stirred well to disperse the particles evenly and then it is mixed into concrete mix. Due to surface modification, the colloidal form of the liquid, and dispersing agents, agglomeration is very minimal. Potable water conforming to IS 456:2000 specifications was used for mixing and curing, maintaining a water-cement ratio of 0.55 to achieve the target compressive strength of 20 MPa at 28 days. In the present study the mechanical attributes of varying proportions (0.3–3%) by volume of glass fiber and coir fiber in M20 mix were investigated in incremental proportions and the strength is predicted using ANN.

3.2. Test methods

The process begins with material selection, where 3% nano silica is used along with varying percentages of glass and coir fibers. Experimental testing is then conducted to measure 28-day compressive, flexural, and tensile strengths. Collected data is normalized and split into training (70%), validation (15%), and testing (15%) sets. The model architecture is designed using an Artificial Neural Network (ANN) with two hidden layers consisting of 64 and 32 neurons, respectively. Hyper parameters such as learning rate and early stopping are tuned, with the final model typically converging within 8–15 epochs. The model is trained on the training set and validated on the validation set. Performance is evaluated using measures such as R2, RMSE, a20 index, WI, NSE, and PBIAS. Plots of actual versus projected results are used to visualize the results. The model’s total R² was 0.9985, its RMSE was 0.0011 MPa, and it was 100% inside the a20 index. The work flow chart (Schematic diagram) is shown in Figure 1.

Figure 1
Work flow chart.

3.2.1. Mechanical properties

In the initial phase concrete samples (cube, cylinder and prism) are investigated for its mechanical behaviour maintaining the proportion of nano silica (After a detailed literature survey, the substitution percent of NS is fixed) at a constant 3% with varying proportions of glass fiber (0.3–3%) at 0.3 percent increment. Keeping the same nano silica content constant samples are prepared for coir fiber infusion at the same proportion as that of glass fiber and the mechanical behaviour is analysed for 28 days strength. Optimum proportion for the fibers is obtained and the percentage increase in the strength is noted after several trial mixes which is then used for strength prediction using Artificial Neural Network. The experimental data used for model development were obtained from laboratory testing conducted at KPR Institute of Engineering and Technology. A total of 110 samples were tested with varying fiber proportions and constant 3% nano silica. The statistical summary is shown in Table 1.

Table 1
Statistical analysis of data (Constant 3% nano silica with various proportions of Glass fibers and Coir fibers).

PROPERTY	MEAN (MPa)		STANDARD DEVIATION		MIN (MPa)		MAX (MPa)
	GLASS FIBERS	COIR FIBERS	GLASS FIBERS	COIR FIBERS	GLASS FIBERS	COIR FIBERS	GLASS FIBERS	COIR FIBERS
Compressive Strength	32.14	30.86	1.28	1.02	30.4	29.39	34.5	32.6
Flexural Strength	4.45	4.23	0.29	0.22	4.02	3.96	4.87	4.61
Split Tensile strength	3.41	2.89	0.24	0.20	3.09	2.63	3.76	3.22

3.2.2. Artificial neural network

The prediction of concrete strength, when made with multiple additives such as nano-silica and fibers, is highly complex, which dictates the demand for highly complex computations. Random Forest (RF) modelling with ANN presents a solid framework for accurate forecasting of the mechanical properties of fiber-reinforced concrete enhanced with nano-silica. This amalgamation of ensemble learning abilities of RF with ANN pattern recognition strengths would provide a robust methodology for the understanding of complex interactions between various constituents of concrete and their resultant properties. Random Forest (RF) modelling was selected because of its strong predictive power, particularly when working with complicated datasets and nonlinear interactions that are common in materials research, such performance prediction and concrete mix formulation. In order to improve accuracy and avoid overfitting, this supervised machine learning technique creates an ensemble of decision trees and mixes their outputs. The advantages are high accuracy, handles non linearity and interactions, robust to overfitting, feature importance and minimal processing required. And some of the limitations are Interpretability, computationally intensive, overfitting in some case and less effective for extrapolation.

Pseudo code for the proposed model is clearly mentioned in terms of input and output parameters and a step by step procedure for models.

Input:

• –

  Dataset with input features (e.g., cement, water, additives, etc.)

• –

  Target variable: Concrete strength (compressive / tensile / flexural)

Output:

• –

  Trained ANN model

• –

  Predicted strength values (test set)

• –

  Evaluation metrics: R², RMSE, WI, NSE, PBIAS, a20 Index

Step 1: Loaded and Pre-processed Dataset – The dataset was imported, identifiers (such as Mix ID) were removed, and input feature (X) were separated from the target output variable (y).

Step 2: Normalized Input Features – All input features (X) were normalized using Min-Max scaling to scale the values between 0 and 1.

Step 3: Split Dataset – The dataset was split into training (70%), validation (15%) and testing (15%) subsets.

Step 4: Defined ANN Architecture – An ANN was constructed with an input layer (neurons equal to the number of input features), two hidden layer (64 and 32 neurons respectively, both with ReLU activation) and an output layer with 1 neuron using linear activation.

Step 5: Compiled model – The model was compiled Mean Square Error (MSE) as the loss function and the Adam optimizer with a learning rate of 0.001.

Step 6: Trained Model – The model was trained on the training data for 100 epochs with a batch size of 1, using validation data for monitoring and applying early stopping to prevent overfitting.

Step 7: Generated Predictions – The trained model was used to generate predictions on the training, validation and test datasets.

Step 8: Evaluated Model Performance - The model’s performance was assessed by calculating R², RMSE, PBIAS, Willmott’s Index of Agreement (WI), Nash–Sutcliffe Efficiency (NSE), and the a20 Index (percentage of predictions within ±20% of actual values).

Step 9: Visualized Results - Model performance was visualized by plotting Actual vs. Predicted values with reference lines (y = x, y = 1.2x, y = 0.8x), a histogram of prediction errors, Radar chart and a Taylor Diagram for the test set.

With this new model, RF takes into consideration the parameters of constant nano silica with coir fiber content ranging from 0.3 to 3 percent, while glass fiber percentage ranged from 0.3–3 percent. In addition to that, this model also takes into consideration the conventional concrete mixture constituents. The architecture of this model is constituted of multiple decision trees, each intended to analyze a different dimension of the composition of concrete, however at the same time taking under consideration the synergy existing between the nano-silica pozzolanic reactivity and the mechanical bridging provided by the hybrid fibers. This capability of the RF algorithm particularly comes into play while assessing the complex interplay between nano-silica particles and the interfacial transition zone where both coir and glass fibers influence the development of microstructures.

Root Mean Square Error (RMSE): RMSE is commonly used by models and predictors to compare predicted values with actual observed values [³³].

$$RMSE=\sqrt{\frac{1}{n}{\displaystyle \sum _{i=1}^n{(y_e-y_0)}^2}}$$

y_e = predicted value;

y_o = observed value;

Coefficient of Determination (R²): R-squared measures how well the regression line represents the data by analyzing the variance of the sample points around the fitted line [³³].

$$(R=){\left(\frac{{\displaystyle {\sum }_{i=1}^n(y_e-{\overline{y}}_e)(y_0-{\overline{y}}_0)}}{{\displaystyle {\sum }_{i=1}^n{(y_e-{\overline{y}}_e)}^2{(y_0-{\overline{y}}_0)}^2}}\right)}^2$$

y_e = predicted value;

y_o = observed value;

_o = mean of the observed value;

_e = mean of the observed value;

Nash-Sutcliffe Efficiency (NSE): NSE is a normalized metric that compares the variance of the residuals to the variance of the observed data [³³].

$$NSE=1-\frac{{\displaystyle {\sum }_{i=1}^n{(y_{obs}-y_{predict})}^2}}{{\displaystyle {\sum }_{i=1}^n{(y_{obs}-{\overline{y}}_{obs})}^2}}$$

The integration of ANN allows the model to have a better prediction with added accuracy by including a multilayer perceptron structure, usually divided into three, that operates in the following sequences: input layers for calculating material proportions, hidden layers that capture complex interrelationships, and output layers which generate numerous strength parameters. This hybrid modelling shows better performance in the prediction of compressive strength, tensile strength, and flexural strength in comparison to traditional techniques of regression.

Artificial Neural Networks (ANN) were chosen for their shown capacity to simulate complicated, non-linear relationships, which are common in material strength prediction problems involving several interacting variables such as nano silica, glass fiber, and coir fiber. Compared to other machine learning models, ANN offers flexibility in architecture and strong generalization capability when properly tuned. A comparison table that summarizes important hyperparameters before and after training, such as learning rate, number of hidden layers, neurons per layer, activation functions, and training epochs, has been provided to increase clarity. This Table 2 supports the argument that ANN is the best method for this study and aids in visualizing the optimization process. The experimental study considered Nano silica as constant input material as 3% of the cement weight across all concrete mixes. Glass fiber and Coir fiber was used as a variable input, ranging from 0.3% to 3% of the concrete volume in increments of 0.3%. The output parameters evaluated were the 28 – day compressive strength, split tensile strength and flexural strength, all measured in MPa.

Table 2
Comparison table summarizing hyper parameters before and after training.

HYPERPARAMETER	INITIAL SETTING	FINAL OPTIMIZED VALUE
Hidden Layers	2	2
Neurons	64-32-1	64-32-1
Activation Function	ReLU	ReLU
Learning Rate	0.001	0.001
Optimizer	Adam	Adam
Loss Function	MSE	MSE
Batch Size	1	1
Epochs	100	8–15*

The methodology was structured to assess how varying proportions of glass and coir fibers influence concrete strength when combined with 3% nano silica. The experimental design followed standardized mechanical testing procedures, and the ANN model was selected for its ability to capture complex nonlinear relationships. Each step, from mix design to model training, was justified based on prior research and optimized through trial evaluations.

4. RESULTS AND DISCUSSION

4.1. Compressive Strength

Figure 2 demonstrates the concretes behavior under compression and it was noted that highest strength was obtained at 1.5% for glass fiber and 1.2% for coir fiber respectively. A 15% increase was noted in the strength of concrete under compression at 1.5% for glass fiber which gradually depreciated with further increment. In case of coir fiber, 9% increase in strength parameters was observed which again gradually depreciated with further incremented proportion of coir fiber in concrete.

Figure 2
Compressive strength results.

Seventy percent of the data is used to train the model. The model’s prediction of compressive strength based on experimental values for training is mean square propagation. For testing, training, and validation, the values are 1, 1, 1, and 1 for overall propagation, and the optimal validation results is 0.1633 at epoch 15 for glass fiber, as shown in the graph. For coir fiber, the results are 0.9947, 0.9994, 0.9998, and 0.9986, and the optimal validation results is 0.1634 at epoch 14 respectively. Figures 3 and 4 display the regression graph and the mean squared error plot for glass fiber respectively, while Figures 5 and 6 display the regression graph and the mean squared error plot for coir fiber respectively. Figure 7 represents the compressive strength analysis for coir and glass fiber reinforced concrete. Figure 7 (a) illustrates the strong correlation between experimental and ANN – predicted compressive strength values of coir – glass fiber reinforced concrete, showing excellent agreement with minimal deviation. Figures 7(b-d) further evaluate model performance using the Radar chart, Taylor diagram and Histogram plot. The ANN model achieved high regression values for both coir and glass fiber datasets, with optimal validation MSEs of 0.16347 and 0.16336, respectively. The close clustering of data points around the ideal line confirms the model’s robustness and predictive accuracy. These results validate the ANN’s effectiveness in reliably forecasting concrete strength across varying fiber combinations.

Figure 3
Neural network performance analysis - train performance vs. test performance vs. Validation performance for Glass fiber reinforced concrete with constant nano silica content under compression.

Figure 4
The plot depicting the mean squared error (MSE) with different converging patterns over the epochs for glass fiber under compression.

Figure 5
Neural network performance analysis - train performance vs. Test performance vs. Validation performance for coir fiber reinforced concrete with constant nano silica content under compression.

Figure 6
The plot depicting the Mean Squared Error (MSE) with different converging patterns over the epochs for coir fiber under compression.

Figure 7
Compressive strength analysis of coir and glass fiber reinforced concrete, (a) Experimental vs Predicted values, (b) Model performance Radar chart, (c) Taylor diagram and (d) Histogram plot.

4.2. Flexural strength

Figure 8 demonstrates the concretes behavior under flexure and it was noted that highest strength was obtained at 1.5% for glass fiber and 1.2% for coir fiber respectively. A 35.6% increase was noted in the strength of concrete under flexure at 1.5% for glass fiber which gradually depreciated with further increment. In case of coir fiber, 16.4% increase in strength parameters was observed at 1.5% which again gradually depreciated with further incremented proportion of coir fiber in concrete. Figure 8 illustrates the variation in flexural strength with increasing fiber content. The highest strength was observed at 1.5% glass fiber and 1.2% coir fiber, showing 35.6% and 16.4% improvement, respectively compared to control. Beyond these points, strength declined, indicating an optimum dosage for fiber reinforcement.

Figure 8
Flexural strength results.

For flexural strength, ANN models were also created. The mean square propagation, testing, training, and validation results are as follows: for glass fibers, the values are 0.8772, 0.9734, 0.9828, and 0.9563; the optimal validation results is 0.0107 at epoch 10; and for coir fibers, the values are 0.9948, 0.9949, 0.9966, and 0.9951; the optimal validation results is 0.0077 at epoch 8. Figures 9 and 10 display the regression graph and the mean squared error plot for glass fiber respectively, while Figures 11 and 12 display the regression graph and the mean squared error plot for coir fiber respectively. Figure 13 represents the flexural strength analysis for coir and glass fiber reinforced concrete. Figure 13 (a) represents the experimental vs predicted values in coir and glass fibers. Figures 13(b-d) shows the model performance using Radar chart, Taylor diagram and Histogram plot of coir fibers and glass fibers in evaluating the flexural strength.

Figure 9
Neural network performance analysis - train performance vs. Test performance vs. Validation performance for glass fiber reinforced concrete with constant nano silica content under flexure.

Figure 10
The plot depicting the Mean Squared Error (MSE) with different converging patterns over the epochs for glass fiber under flexure.

Figure 11
Neural network performance analysis - train performance vs. Test performance vs. Validation performance for coir fiber reinforced concrete with constant nano silica content under flexure.

Figure 12
The plot depicting the Mean Squared Error (MSE) with different converging patterns over the epochs for coir fiber under flexure.

Figure 13
Flexural strength analysis of coir and glass fiber reinforced concrete, (a) Experimental vs Predicted values, (b) Model performance Radar chart, (c) Taylor diagram, (d) Histogram plot.

4.3. Split tensile strength

Figure 14 demonstrates the concretes behavior under tension and it was noted that highest strength was obtained at 1.5% for glass fiber and coir fiber respectively. A 43.5% increase was noted in the strength of concrete under flexure at 1.5% for glass fiber which gradually depreciated with further increment. In case of coir fiber, 32.5% increase in strength parameters was observed at 1.5% which again gradually depreciated with further incremented proportion of coir fiber in concrete.

Figure 14
Split tensile strength.

The mean square propagation, testing, training, and validation results are used to prepare the ANN model. For glass fibers, the results are 0.8803, 0.9591, 0.9536, and 0.9437, and the optimal validation results is 0.0071 at epoch 8; for coir fibers, the results are 0.9743, 0.9949, 0.9941, and 0.9909, and the optimal validation results is 0.0220 at epoch 8. Figures 15 and 16 display the regression graph and the mean squared error plot for glass fiber respectively, while Figures 17 and 18 display the regression graph and the mean squared error plot for coir fiber respectively. The high correlation observed across taining, validation and testing sets indicates excellent model generalization. The near - unity R² values confirm that the ANN effectively captures the nonlinear relationship between fiber content and tensile strength. Figure 19 represents the split tensile strength analysis for coir and glass fiber reinforced concrete. Figure 19 (a) represents the experimental vs predicted values in coir and glass fibers. Figures 19(b-d) shows the model performance using Radar chart, Taylor diagram and Histogram plot of coir fibers and glass fibers in evaluating the split tensile strength.

Figure 15
Neural network performance analysis- train performance vs. Test performance vs. Validation performance for glass fiber reinforced concrete with constant nano silica content under tension.

Figure 16
The plot depicting the Mean Squared Error (MSE) with different converging patterns over the epochs for glass fiber under tension.

Figure 17
Neural network performance analysis- train performance vs. Test performance vs. Validation performance for coir fiber reinforced concrete with constant nano silica content under tension.

Figure 18
The plot depicting the Mean Squared Error (MSE) with different converging patterns over the epochs for coir fiber under tension.

Figure 19
Split tensile strength analysis of coir and glass fiber reinforced concrete, (a) experimental vs predicted values, (b) model performance radar chart, (c) taylor diagram, (d) histogram plot.

In addition to typical metrics like R² and RMSE, we also generated the a20 index, which assesses the percentage of predictions within ±20% of the experimental data. At 0.0013 MPa, 0.0001 MPa, and 0.0001 MPa for training, validation, and testing, respectively, the RMSE values (in MPa) were extremely modest, suggesting that there was little error between the expected and actual values. Our ANN model achieved a flawless a20 index of 100.00% across all datasets, suggesting that every prediction was within ±20% of the actual experimental value. With a flawless a20 index and a low RMSE, this demonstrates the model’s high accuracy and dependability and qualifies it for use in predicting concrete strength in practical situations.

The ANN model developed in this study offers notable advantages, including high predicyion accuracy, reduced experimental effort and effective handling of nonlinear relationships between input parameters. However, its performance depends on the quality and range of training data, and the model may not generalize well beyond the scope. Additionally, as with most neural networks, interpretability remains limited due to its balck- box nature. Despite these constraints, the model serves as a reliable and efficient tool for predicting concrete strength.

Careful hyperparameter adjustment was essential to our ANN model’s performance. Table 3 represents the hyper parameter for different models.The complex interactions between concrete components and the consequent mechanical properties were too complex to represent with our original architecture, which had a single 16-neuron buried layer. Increasing the number of hidden layers to two (64 and 32 neurons, respectively) greatly improved the ability to recognize patterns.

Table 3
Hyper parameters for different models.

HYPERPARAMETER	INITIAL SETTING	FINAL OPTIMIZED VALUE	PURPOSE
Number of hidden layers	1 (16 neurons)	2 (64 + 32 neurons)	Increased complexity for better learning ability
Activation function	ReLU	ReLU	Nonlinearity and computational efficiency
Optimizer	SGD	Adam	Faster convergence and stability
Learning rate	0.01	0.001	Optimized training speed and stability
Batch size	1	4	Stable updates, balancing memory and performance
Epochs	100	150	To ensure convergence
Early stopping patience	None	20	Prevents overfitting

For improved training stability, we moved to the Adam optimizer while maintaining the effective ReLU activation function. For more accurate adjustments, we decreased the learning rate from 0.01 to 0.001, and for balanced processing, we increased the batch size from 1 to 4 samples. To achieve a compromise between thorough learning and efficiency, training cycles were extended to 150 epochs with automatic halting when no further improvement occurred. Figure 16 compares the prediction accuracy of ANN models developed for both glass and coir fiber-reinforced concrete. The coir based model achieved the highest accuracy, with R² values exceeding 0.99 and minimal RMSE, while the glass fiber model showed slightly lower performance. This variation is due to better data consistency and optimized fiber interaction in the coir based mixes, enhancing the model’s learning efficiency.

In addition to standard metrics such as R^2 and RSME, model performance was assessed using Willmott’s Index(WI), Nash-Sutcliffe Efficiency(NSE), and Perfect Bias(PBIAS), which are widely accepted in model validation [²⁷–²⁹]. These metrics offer deeper insights into the model’s agreement with experimental results, robustness and error bias is shown in Table 4.

Table 4
Evaluation matrices for different models.

MODEL TYPE	COEFFICIENT OF DETERMINATION (R2)	ROOT MEAN SQUARED ERROR (RMSE)	WILLMOTT’S INDEX (WI)	NASH–SUTCLIFFE EFFICIENCY (NSE)	PERCENT BIAS (PBIAS)
ANN- glass fiber	0.9563	0.0107	0.9814	0.9422	+1.73
ANN coir-fibre	0.9951	0.0077	0.9927	0.9841	+0.94

The strength results obtained in this study align well with existing literature but show noticeable improvements in prediction accuracy. For instance, achieved an MAE reduction of 27% using hybrid models, while the present ANN model achieved a a20 index and R^2 above 0.99 for coir based mixes [²⁴]. Similarly, predicted compressive strength of scc with an RMSE of 0.0043 MPa, while our ANN model showed comparable precision with RMSE as low as 0.0001 MPa. Additionally, the integration of hybrid fibers with nano silica in this study provides a unique, sustainable and efficient approach not commonly addressed in earlier models [²⁹].

Future studies can explore hybrid fiber – reinforced concrete using other natural and industrial fibers, or extend the model to predict long-term durability parameters. Integrating deep learning architectures such as CNN or LSTM may further enhance prediction accuracy. The approach can also be adapted for other grades of concrete or sustainability-focused applications. While all prediction curves appear to follow a similar direction, this consistency reflects the model’s ability to learn the general behavior of the data effectively. To ensure fairness and avoid overfitting, separate models were developed for each strength type and fiber variation. Each was tested independently and showed strong performance across key metrics, confirming that the uniform trend results from stable learning rather than biased prediction. This consistent direction across datasets highlights the model’s reliability in capturing the underlying patterns of the material behavior.

5. CONCLUSION

The mechanical properties and the strength predictions using Artificial neural network of the concrete by adding constant 3% nano silica with glass fibers and coir fibers were examined in the present investigation. The forthcoming conclusions were made with respect to the results obtained.

1. The dataset consists of glass, coir fibers incorporated with nano silica, with 70% used for training, and 15% for validation and 15%, for testing.

2. Regression values of 1, 0.9437, and 0.9563 for compressive, split tensile, and flexural strengths were found by combining a constant 3% nano silica with different percentages of glass fibers. Regression values of 0.9986, 0.9909, and 0.9951 for compressive, split tensile, and flexural strengths were found by combining a constant 3% nano silica with different percentages of coir fibers.

3. Across all datasets, the ANN model had a perfect a20 index of 100.00%, indicating that each prediction was within ±20% of the experimental value.

4. A higher regression coefficient denotes a more accurate model for making predictions. It’s interesting to note that the model’s projected values closely matched the experimental values for each strength.

5. This study uniquely combines nano silica with both coir(mean) and glass fibers(synthetic) fibers to develop a predictive ANN model for concrete strength. The model achieved excellent accuracy, offering a reliable alternative to traditional trial-based testing. These findings contribute a novel data-driven approach to sustainable concrete design.

6. BIBLIOGRAPHY

• ^[1] ABHILASH, P.P., NAYAK, D.K., SANGOJU, B., et al., “Effect of nano-silica in concrete; a review”, Construction & Building Materials, v. 278, pp. 122347, 2021. doi: http://doi.org/10.1016/j.conbuildmat.2021.122347.
  » https://doi.org/10.1016/j.conbuildmat.2021.122347
• ^[2] SAID, A., ZEIDAN, M., BASSUONI, M., et al., “Properties of concrete incorporating nano-silica”, Construction & Building Materials, v. 36, pp. 838–844, 2012. doi: http://doi.org/10.1016/j.conbuildmat.2012.06.044.
  » https://doi.org/10.1016/j.conbuildmat.2012.06.044
• ^[3] HAMADA, H., SHI, J., YOUSIF, S., et al., “Use of nano-silica in cement-based materials: a comprehensive review”, Journal of Sustainable Cement-Based Materials, v. 12, n. 10, pp. 1286–1306, 2023. doi: http://doi.org/10.1080/21650373.2023.2214146.
  » https://doi.org/10.1080/21650373.2023.2214146
• ^[4] KHAN, K., AHMAD, W., AMIN, M., et al., “Nano-silica-modified concrete: a bibliographic analysis and comprehensive review of material properties”, Nanomaterials, v. 12, n. 12, pp. 1989, 2022. doi: http://doi.org/10.3390/nano12121989. PMid:35745327.
  » https://doi.org/10.3390/nano12121989
• ^[5] YUAN, Z., JIA, Y., “Mechanical properties and microstructure of glass fiber and polypropylene fiber reinforced concrete: An experimental study”, Construction & Building Materials, v. 266, pp. 121048, 2021. doi: http://doi.org/10.1016/j.conbuildmat.2020.121048.
  » https://doi.org/10.1016/j.conbuildmat.2020.121048
• ^[6] TASSEW, S., LUBELL, A., “Mechanical properties of glass fiber reinforced ceramic concrete”, Construction & Building Materials, v. 51, pp. 215–224, 2014. doi: http://doi.org/10.1016/j.conbuildmat.2013.10.046.
  » https://doi.org/10.1016/j.conbuildmat.2013.10.046
• ^[7] AHMAD, J., GONZÁLEZ-LEZCANO, R., MAJDI, A., et al., “Glass fibers reinforced concrete: overview on mechanical, durability and microstructure analysis”, Materials, v. 15, n. 15, pp. 5111, 2022. doi: http://doi.org/10.3390/ma15155111. PubMed PMID: 35897549.
  » https://doi.org/10.3390/ma15155111
• ^[8] OROUJI, M., ZAHRAI, S., NAJAF, E., “Effect of glass powder & polypropylene fibers on compressive and flexural strengths, toughness and ductility of concrete: an environmental approach”, Structures, v. 33, pp. 4616–4628, 2021. doi: http://doi.org/10.1016/j.istruc.2021.07.048.
  » https://doi.org/10.1016/j.istruc.2021.07.048
• ^[9] HILLES, M.M., ZIARA, M.M., “Mechanical behavior of high strength concrete reinforced with glass fiber”, Engineering Science and Technology, v. 22, n. 3, pp. 920–928, 2019. doi: http://doi.org/10.1016/j.jestch.2019.01.003.
  » https://doi.org/10.1016/j.jestch.2019.01.003
• ^[10] RUBEN, N., VENKATESH, C., DURGA, C., et al., “Comprehensive study on performance of glass fibers-based concrete”, Innovative Infrastructure Solutions, v. 6, n. 2, pp. 112, 2021. doi: http://doi.org/10.1007/s41062-021-00490-4.
  » https://doi.org/10.1007/s41062-021-00490-4
• ^[11] GNANASUNDAR, V., “Experimental analysis of glass fibre in concrete”, Sustainable Materials and Smart Practices, v. 23, pp. 246–254, 2022. doi: http://doi.org/10.21741/9781644901953-28.
  » https://doi.org/10.21741/9781644901953-28
• ^[12] KATMAN, H.Y.B., KHAI, W.J., BHEEL, N., et al., “Fabrication and characterization of cement-based hybrid concrete containing coir fiber for advancing concrete construction”, Buildings, v. 12, n. 9, pp. 1450, 2022. doi: http://doi.org/10.3390/buildings12091450.
  » https://doi.org/10.3390/buildings12091450
• ^[13] KORE, S., “Sustainable production of concrete using coir fibres”, IOP Conference Series: Earth and Environmental Science, v. 795, pp. 012006, 2021. doi: http://doi.org/10.1088/1755-1315/795/1/012006.
  » https://doi.org/10.1088/1755-1315/795/1/012006
• ^[14] GUPTA, M., KUMAR, M., “Effect of nano silica and coir fiber on compressive strength and abrasion resistance of Concrete”, Construction & Building Materials, v. 226, pp. 44–50, 2019. doi: http://doi.org/10.1016/j.conbuildmat.2019.07.232.
  » https://doi.org/10.1016/j.conbuildmat.2019.07.232
• ^[15] YAN, L., CHOUW, N., HUANG, L., et al., “Effect of alkali treatment on microstructure and mechanical properties of coir fibres, coir fibre reinforced-polymer composites and reinforced-cementitious composites”, Construction & Building Materials, v. 112, pp. 168–182, 2016. doi: http://doi.org/10.1016/j.conbuildmat.2016.02.182.
  » https://doi.org/10.1016/j.conbuildmat.2016.02.182
• ^[16] AMIRTHARAJ, J., “Strength enhancement of concrete using coir fiber”, International Journal for Research in Applied Science and Engineering Technology, v. 7, n. 4, pp. 1176–1183, 2019. doi: http://doi.org/10.22214/ijraset.2019.4210.
  » https://doi.org/10.22214/ijraset.2019.4210
• ^[17] ANANDARAJ, S., ROOBY, J., RAVINDRAN, G., et al., “Strength prediction using ANN for concrete with Marble and Quarry dust”, In: International Conference on Intelligent Computing and Communication for Smart World (I2C2SW), Erode, India, 2018. doi: http://doi.org/10.1109/I2C2SW45816.2018.8997326.
  » https://doi.org/10.1109/I2C2SW45816.2018.8997326
• ^[18] GUPTA, R., KEWALRAMANI, M., GOEL, A., “Prediction of concrete strength using neural-expert system”, Journal of Materials in Civil Engineering, v. 18, n. 3, pp. 462–466, 2006. doi: http://doi.org/10.1061/(ASCE)0899-1561(2006)18:3(462).
  » https://doi.org/10.1061/(ASCE)0899-1561(2006)18:3(462)
• ^[19] HOSSAIN, K., GLADSON, L., ANWAR, M., “Modeling shear strength of medium- to ultra-high-strength steel fiber-reinforced concrete beams using artificial neural network”, Neural Computing & Applications, v. 28, n. S1, pp. 1119–1130, 2017. doi: http://doi.org/10.1007/s00521-016-2417-2.
  » https://doi.org/10.1007/s00521-016-2417-2
• ^[20] NGUYEN, Q.H., LY, H., NGUYEN, T., et al., “Investigation of ANN architecture for predicting shear strength of fiber reinforcement bars concrete beams”, PLoS One, v. 16, n. 4, pp. e0247391, 2021. doi: http://doi.org/10.1371/journal.pone.0247391. PubMed PMID: 33798200.
  » https://doi.org/10.1371/journal.pone.0247391
• ^[21] TRAN, V.Q., MAI, H., NGUYEN, T., et al., “Investigation of ANN architecture for predicting the compressive strength of concrete containing GGBFS”, PLoS One, v. 16, n. 12, pp. e0260847, 2021. doi: http://doi.org/10.1371/journal.pone.0260847. PubMed PMID: 34860842.
  » https://doi.org/10.1371/journal.pone.0260847
• ^[22] LIN, C., WU, N., “An ANN model for predicting the compressive strength of concrete”, Applied Sciences, v. 11, n. 9, pp. 3798, 2021. doi: http://doi.org/10.3390/app11093798.
  » https://doi.org/10.3390/app11093798
• ^[23] ASTERIS, P., SKENTOU, A., BARDHAN, A., et al., “Predicting concrete compressive strength using hybrid ensembling of surrogate machine learning models”, Cement and Concrete Research, v. 145, pp. 106449, 2021. doi: http://doi.org/10.1016/j.cemconres.2021.106449.
  » https://doi.org/10.1016/j.cemconres.2021.106449
• ^[24] TAK, M.S.N., FENG, Y., MAHGOUB, M., “Advanced machine learning techniques for predicting concrete compressive strength”, Infrastructures, v. 10, n. 2, pp. 26, 2025. doi: http://doi.org/10.3390/infrastructures10020026.
  » https://doi.org/10.3390/infrastructures10020026
• ^[25] ASTERIS, P., SKENTOU, A., BARDHAN, A., et al., “Predicting concrete compressive strength using hybrid ensembling of surrogate machine learning models”, Cement and Concrete Research, v. 145, pp. 106449, 2021. doi: http://doi.org/10.1016/j.cemconres.2021.106449.
  » https://doi.org/10.1016/j.cemconres.2021.106449
• ^[26] ZHU, W., HUANG, L., MAO, L., et al., “Predicting the uniaxial compressive strength of oil palm shell lightweight aggregate concrete using artificial intelligence‐based algorithms”, Structural Concrete, v. 23, n. 6, pp. 3631–3650, 2021. doi: http://doi.org/10.1002/suco.202100656.
  » https://doi.org/10.1002/suco.202100656
• ^[27] HUANG, Z., CUENCA, E., FERRARA, L., “Optimized data-driven method to study the self-healing and durability of ultra-high-performance concrete”, Engineering Applications of Artificial Intelligence, v. 143, pp. 110043, 2025. doi: http://doi.org/10.1016/j.engappai.2025.110043.
  » https://doi.org/10.1016/j.engappai.2025.110043
• ^[28] ALKAYEM, N.F., SHEN, L., MAYYA, A., et al., “Prediction of concrete and FRC properties at high temperature using machine and deep learning: A review of recent advances and future perspectives”, Journal of Building Engineering, v. 83, pp. 108369, 2024. doi: http://doi.org/10.1016/j.jobe.2023.108369.
  » https://doi.org/10.1016/j.jobe.2023.108369
• ^[29] ALZABEN, N., MAASHI, M., NOURI, A.M., et al., “Artificial neural network to predict the compressive strength of high strength self-compacting concrete made of marble dust”, Revista Materia, v. 29, n. 3, pp. e20240329, 2024. doi: http://doi.org/10.1590/1517-7076-rmat-2024-0329.
  » https://doi.org/10.1590/1517-7076-rmat-2024-0329
• ^[30] ZHANG, Y., YUAN, X., “Prediction of high-performance concrete compressive strength using Decision Tree-Guided Artificial Neural Network Pretraining approach”, Engineering Applications of Artificial Intelligence, v. 156, pp. 110828, 2025. doi: http://doi.org/10.1016/j.engappai.2025.110828.
  » https://doi.org/10.1016/j.engappai.2025.110828
• ^[31] SAMANTARAY, S., SAHOO, A., “Groundwater level prediction using an improved ELM model integrated with hybrid particle swarm optimisation and grey wolf optimisation”, Groundwater for Sustainable Development, v. 26, pp. 101178, 2024. doi: http://doi.org/10.1016/j.gsd.2024.101178.
  » https://doi.org/10.1016/j.gsd.2024.101178
• ^[32] SAMANTARAY, S., SAHOO, A., YASEEN, Z.M., et al., “River discharge prediction based multivariate climatological variables using hybridized long short-term memory with nature inspired algorithm”, Journal of Hydrology, v. 649, pp. 132453, 2025. doi: http://doi.org/10.1016/j.jhydrol.2024.132453.
  » https://doi.org/10.1016/j.jhydrol.2024.132453
• ^[33] MAHAKUR, V., MAHAKUR, V.K., SAMANTARAY, S., et al., “Prediction of runoff at ungauged areas employing interpolation techniques and deep learning algorithm”, HydroResearch, v. 8, pp. 265–275, 2025. doi: http://doi.org/10.1016/j.hydres.2024.12.001.
  » https://doi.org/10.1016/j.hydres.2024.12.001

Publication Dates

• Publication in this collection
  03 Oct 2025
• Date of issue
  2025

History

• Received
  31 Jan 2025
• Accepted
  18 Aug 2025