Introduction

The use of concrete has grown substantially in recent years, driving up the need for cement as a primary binding agent. Worldwide, cement demand continues to climb, with annual production exceeding 2.8 billion tons to support expanding infrastructure projects—a trend that shows a consistent yearly growth rate of around 3%1,2,3. However, cement manufacturing comes with a major environmental drawback: producing just one ton of cement releases roughly one ton of carbon dioxide (CO₂) into the air. With concrete being such a fundamental construction material, finding sustainable alternatives has become a pressing priority4,5.

One promising solution is geopolymer, an innovative binder that can replace traditional cement. Geopolymer concrete (GPC) is created by mixing alumino-silicate materials with an alkaline activator, forming a hardened structure with a notably smaller carbon footprint than ordinary cement-based concrete6,7. Beyond its environmental benefits, GPC also reduces problems like shrinkage cracks, which often occur with high cement content. The chemical structure of geopolymer consists of interconnected alumina and silicate chains, represented by the formula \(M_{n} \left[ { - \left( {S{\text{i}}O_{2} } \right)z - AlO_{2} } \right]n \cdot wH_{2} O\), where ‘n’ indicates polymerization and ‘M’ stands for alkali ions such as potassium (K⁺) or sodium (Na⁺)8.

Various industrial and natural materials can serve as raw ingredients for GPC, including fly ash, rice husk ash, red mud, silica fume (SF), ground granulated blast furnace slag (GGBS), and natural zeolites such as metakaolin9,10. Among these, GGBS is especially useful because it helps form a dense, high-strength microstructure in alkali-activated concrete, even under normal curing conditions11. This compact structure enhances the material’s mechanical performance, making it more durable. Additionally, natural zeolitic materials such as metakaolin are valued for their rich chemical makeup and non-crystalline nature12. GGBS stands out due to its widespread availability and high silica-alumina content, offering a practical way to repurpose industrial waste while improving concrete sustainability.

Fly ash is a fine, powdery byproduct generated from coal combustion in thermal power plants. Classified as either Class F (low calcium) or Class C (high calcium), it is widely used as a SCM in concrete due to its pozzolanic and cementitious properties13. Fly ash improves concrete workability by reducing water demand due to its spherical particle shape14. It undergoes a pozzolanic reaction with calcium hydroxide (Ca(OH)₂) in cement, forming additional calcium silicate hydrate (C-S–H) gel, which enhances long-term strength and durability15. Studies show that replacing 15–30% of cement with fly ash can improve sulfate resistance, reduce heat of hydration, and mitigate alkali-silica reaction (ASR)16. Using fly ash in concrete reduces CO₂ emissions by decreasing cement consumption, contributing to sustainable construction. It also repurposes an industrial waste product, minimizing landfill disposal. Research indicates that fly ash-modified concrete exhibits lower permeability, enhancing resistance to chloride ingress and corrosion in reinforced structures. Fly ash concrete typically has slower early strength development, requiring longer curing times. Variability in fly ash composition can affect performance, necessitating strict quality control. Additionally, declining coal use in some regions has reduced fly ash availability, prompting research into alternative SCMs. Fly ash remains a key material in sustainable concrete production, offering durability, environmental, and economic benefits17.

Silica fume, also known as microsilica, is a byproduct of silicon and ferrosilicon alloy production. It consists of ultrafine, amorphous silicon dioxide (SiO₂) particles, typically less than 1 µm in diameter. Due to its high pozzolanic activity and fine particle size, silica fume is widely used as a supplementary cementitious material (SCM) in concrete to enhance mechanical and durability properties18,19. Silica fume reacts with calcium hydroxide (Ca (OH)₂) in cement to form additional calcium silicate hydrate (C-S–H) gel, improving concrete’s compressive strength, bond strength, and abrasion resistance. Studies indicate that incorporating 5–15% silica fume by weight of cement significantly increases early and long-term strength while reducing permeability20. This makes it particularly useful in high-performance concrete (HPC), ultra-high-performance concrete (UHPC), and marine structures where durability is critical. Silica fume reduces chloride ion penetration, carbonation, and sulfate attack, enhancing the longevity of reinforced concrete structures. Its fine particles fill micro-voids, refining the pore structure and reducing water absorption. Research has shown that silica fume-modified concrete exhibits superior resistance to alkali-silica reaction (ASR) and corrosion of reinforcing steel21,22. Despite its advantages, silica fume increases concrete’s viscosity, requiring superplasticizers for workability. Proper curing is essential to prevent plastic shrinkage cracking due to its high reactivity23.

Ground Granulated Blast Furnace Slag (GGBS) is a byproduct of iron production in blast furnaces. When molten slag is rapidly quenched with water and finely ground, it forms a glassy, granular material with latent hydraulic properties. GGBS is widely used as a partial replacement for Portland cement in concrete, offering technical and environmental benefits. GGBS reacts with water and alkalis in cement to form cementitious compounds, contributing to long-term strength development24,25. Research indicates that replacing 30–70% of cement with GGBS enhances durability by reducing permeability, chloride ingress, and sulfate attack. Its slow hydration rate lowers the heat of hydration, making it ideal for mass concrete applications26,27. Additionally, GGBS improves workability and reduces the risk of thermal cracking. The use of GGBS significantly reduces the carbon footprint of concrete, as its production requires less energy compared to Portland cement. By utilizing an industrial byproduct, GGBS supports sustainable construction practices and reduces landfill waste. Past studies have demonstrated that GGBS-modified concrete exhibits superior resistance to chemical aggression and extends the service life of structures in harsh environments28. The primary limitation of GGBS is its slower early strength development, which may delay formwork removal in construction projects.

Microfibers are short and fine synthetic fibers typically ranging from 0.1 to 1 mm in diameter and 5 to 30 mm in length. When added to concrete to enhance mechanical and durability properties by controlling crack formation at the micro-level29,30. Microfibers improve tensile strength, impact resistance, and fracture toughness by bridging micro-cracks that develop during plastic shrinkage and hardening. Past research indicated that adding 0.1–0.5% volume fraction of microfibers reduces early-age plastic shrinkage cracking by up to 80%31. Unlike macrofibers, microfibers do not significantly affect workability when used in appropriate dosages. These helps to enhance abrasion resistance and reduce permeability, improving durability against freeze–thaw cycles and chemical exposure32,33. Microfiber-reinforced concrete is widely used in thin overlays, shotcrete, precast elements, and high-performance flooring31. Steel microfibers are particularly effective in industrial floors and blast-resistant structures, while synthetic fibers (e.g., polypropylene) are preferred for fire resistance due to their melting behaviour, which creates escape pathways for vapour pressure. Excessive fiber content can lead to balling effects and reduced workability, requiring adjustments in mix design34. Dispersion uniformity is critical to prevent clumping. Long-term performance under sustained loads and environmental exposure remains an area of ongoing research. A comparative summary of recent ML studies on SCM-based concrete has been presented in Table 1.

Table 1 Summary of recent studies on ML applications in SCM-based concrete.
Full size table

Research gap and objectives

Despite the widespread use of supplementary cementitious materials (SCMs) like silica fume, fly ash, and GGBS in concrete, there remains a lack of robust, data-driven models to accurately predict their influence on chloride permeability, particularly when combined with microfibers. Conventional experimental approaches for assessing the RCPT results are time-intensive, costly, and limited in capturing the complex, nonlinear interactions among mix constituents. To address this gap, this study aims to:

  1. 1.

    Developing and comparing advanced soft computing algorithms—AdaBoost, AVOA, CatBoost, and LightGBM for predicting RCPT values, integrating SCMs with microfiber-based geopolymer concrete.

  2. 2.

    Identifying the most influential input variables governing chloride permeability through SHAP interpretability analysis.

  3. 3.

    Conduct a comparative evaluation between experimental and computational durability assessment methods, highlighting the advantages of ML-based prediction in terms of time efficiency, reproducibility, and sustainability.

  4. 4.

    Establishing a reliable computational framework to optimize durable and sustainable GPC mixes, minimizing reliance on empirical testing.

Materials and mix design

Materials

The current study utilized class F type fly ash (FA), ground granulated blast furnace slag (GGBS), and silica fume (SF), which procured from industrial sources in the Jamshedpur region. These materials conform to relevant Indian standards, their chemical composition, fineness as per IS 3812 and IS 12,089, respectively, depending on their origin, processing method, and parent industrial process. The chemical compositions of FA, GGBS, and SF are provided in Table 2. The physical properties of these SCMs are summarized in Table 3. On the other hand, microfibers (MF) were procured from Chirag Industries Ltd., Nagpur, conforming to IS 14,889 (Part I): 2000 specifications. The fibres possessed a length of 12 mm, a diameter of 20 µm, a tensile strength of 350 MPa, and a modulus of elasticity of about 4.0 GPa. The specific gravity was 0.91, with a melting point around 160–165 °C, ensuring stability under typical geopolymer curing conditions.

Table 2 Chemical composition of SCMs from the Jamshedpur region.
Full size table
Table 3 Physical properties of FA, GGBS and SF.
Full size table

Mix design

To ensure reproducibility, the detailed mix proportions of all experimental combinations used for model development are presented in Table 4. The mixtures were designed by partially replacing ordinary Portland cement (OPC) with varying proportions of FA, GGBS, SF and adding with of MF content. The alkaline activator composition and binder-to-aggregate ratios were maintained constant throughout all mixtures. The mix design for the geopolymer concrete (GPC) mixes was developed in accordance with IS 10,262:2019.

Table 4 Mix composition details used for GPC mixes.
Full size table

Methodology and sampling procedure

GPC mixes were prepared by dry-mixing FA, GGBS, and SF in varying proportions for 2 min to ensure uniform blending of the binders. A total of 200 specimens were cast based on different mix combinations for subsequent testing and model development. A 12 M NaOH solution was then added to the dry mix to achieve a workable consistency, followed by 5 min of mechanical mixing. The fresh geopolymer concrete was cast into cube molds (150 mm × 150 mm × 150 mm) in three layers, each compacted using a vibrating table for 30 s to eliminate air voids40. The specimens were sealed with plastic sheets to prevent moisture loss, cured at 80 °C for 24 h in an oven to accelerate geo-polymerization, and subsequently demolded and ambient-cured (25 ± 2 °C) until testing at 28 days. The resistance of concrete specimens to chloride ion penetration was evaluated using RCPT test in accordance with ASTM C120241. Figure 1. (a) depicts a systematic test arrangement of concrete samples into circular disc specimens, and (b) testing of concrete samples using the RCPT cell apparatus. The test was carried out with a cylindrical disc specimen having a diameter of 100 mm and a thickness of 50 mm. During testing, one side of the specimen was exposed to a 3% sodium chloride (NaCl) solution, while the opposite side was exposed to a 0.3N sodium hydroxide (NaOH) solution. A direct current (DC) voltage of 60 V was applied across the specimen for a duration of 6 h. The current passing through the specimen was recorded at 30-min intervals. The total charge passed (Q) was calculated using Eq. (1) as specified in ASTM C1202. This standard classifies the charge passed (C) during a 6-h test into five distinct durability classes is summarized in Table 5.

$$Q = \mathop \sum \limits_{i = 1}^{n} \frac{{\left( {I_{i} + I_{i + 1} } \right)}}{2} \times {\Delta }t$$
(1)
Fig. 1
Fig. 1
Full size image

(a) Arrangement of concrete sample into circular disc size (b) Testing of concrete sample by RCPT cell apparatus.

Table 5 ASTM C1202 classification of chloride permeability.
Full size table

where, \(Q\) = total charge passed (C), \(I_{i}\) and \(I_{i + 1}\) = current (Amperes) at the start and end of each time interval and \({\Delta }t\) = time interval every 30 min in Amperes.

Overview of the dataset for ML models

The dataset employed in this study comprises 200 experimental data points obtained from laboratory investigations on GPC mixes incorporating varying proportions of FA, GGBS, SF, and MF. These materials were selected to enhance the durability and sustainability of GPC to improving resistance to chloride ion penetration, as evaluated through the RCPT. Each dataset represents a mix combination defined by its binder composition, activator characteristics, and accompanied by the measured RCPT values. For model development, the entire dataset was randomly divided into two subsets following a 70:30 ratio, where 70% were used for model training, and 30% were utilized for testing and validation.

Dataset scaling and normalization

Prior to model development, it is essential to preprocess the raw dataset to ensure that all input variables contribute equally to the learning process42. The dataset used in this study comprises experimental results of GPC mixes incorporating varying proportions of FA, GGBS, SF, MF, with corresponding to RCPT values as the target output. Since these features possess different magnitudes and units, scaling and normalization were applied to enhance model training stability and convergence speed. To eliminate bias due to variable magnitude differences, Min–Max normalization was employed to scale all input features to a common range of [0,1]. All input and out variables were scaled to a consistent range using Min–Max normalization as expressed in Eq. (2). Table 6 summarized the descriptive statistics, scaling, and normalization range of input and output variables.

$$P_{scaled} = \frac{{P - P_{min} }}{{P_{max} - P_{min} }}$$
(2)
Table 6 Statistical summary and normalization range of input and output variables.
Full size table

where, \(P\) is the original value, \(P_{{{\text{min}}}}\) and \(P_{{{\text{max}}}}\) is the minimum and maximum values of the feature and \(P_{{{\text{scaled}}}}\) is the normalized value.

Where; C: Coulombs

Multicollinearity assessment and exploratory pair plot analysis

Figure 2 presents the Spearman correlation matrix illustrating the interrelationship among the input variables with respect to the output parameter. It was observed that, FA and SF exhibit moderate positive correlations with RCP, indicating that an increase in their proportion tends to influence chloride ion permeability. GGBS shows a relatively weaker correlation, suggesting a lesser but still noticeable effect on RCP. The correlation between microfiber content and RCPT values are also moderate, implying that fibre inclusion slightly alters the permeability characteristics by improving pore refinement and crack resistance. Furthermore, a strong positive correlation was observed between SF and MF, implying a synergistic relationship in mixes containing both, likely due to improved particle packing and matrix densification.

Fig. 2
Fig. 2
Full size image

Spearman correlation matrix showing the interrelationship between input and output variables.

Figure 3 depicts the pair plot analysis illustrating the pairwise relationships and individual data distributions of input parameters used in the GPC mixes. The diagonal plots represent the probability density distribution of each variable, whereas the off-diagonal scatter plots display their bivariate relationships. A positive association was noticed between GGBS and FA, suggesting that in several mix combinations, higher GGBS content coincides with increased fly ash levels, indicating their complementary contribution to the binder system. In contrast, MF and SF show a relatively scattered distribution with respect to GGBS and FA, implying direct correlations with the remaining variables.

Fig. 3
Fig. 3
Full size image

Pair plot illustrating the correlation and distribution patterns among input variables.

Development of predictive models using machine learning algorithms

Adaptive boosting

Adaptive Boosting (AdaBoost) is an ensemble ML algorithm which aims to improve the prediction power of a model based on the weighted addition of multiple weak learners, such as decision trees. The reward and penalty process are operated through attaching different weights to the training set, and these weights are updated after each iteration, so that misclassified samples will possess larger weights that force the following classifier to focus on the difficult cases. The prediction of the AdaBoost ensemble model is the accumulated weighted sum of the predictions of individual weak learners; that is, the more decisions which is right, the more important the weak learners are in the ensemble. AdaBoost is especially good at decreasing bias and variance, making it a durable method in classification, and (less so) in regression. It is affected by noise and outliers due to the fact of raising the weights of the wrongly classified samples (e.g., the mislabeled examples). Nevertheless, AdaBoost has received great attention, owing to its simplicity, generality, and effectiveness in boosting model performance without high parameter tuning.

African vultures optimization algorithm

African Vultures Optimization Algorithm (AVOA) is an emerging metaheuristic optimization algorithm that has been recently introduced by mimicking the foraging and social habits of the African vultures in real life. Formulated to grow under intricate optimization tasks, AVOA imitates the intelligent behaviours of vultures in seeking food, responding to the competition for living resources, and cooperating while living in harsh environments. In the algorithm, the vs are mathematically formulated to balance two important optimization stages: the exploration one (the search space is thoroughly covered to delay convergence to a local optimum) and the exploitation one (the search is focused in the vicinity of good solutions). The vultures are characterized as adaptable birds that have incredibly good vision and move in very energy-efficient ways, which are also considered in the algorithm by means of making it possible to control the behaviour of the search agents (vultures) in a dynamic way (hungry time, number of other vultures, etc.). Vultures in AVOA are dispersed at first in the search space and fly towards promising regions (using a certain degree of randomness reserved for less populated areas and individuals close to potential solutions). The algorithm presents important activities, as dynamic competitive and cooperative foraging, with subjects that can compete for the best solution regions or cooperate in organizations to exploit rich areas thoroughly. Their strategies depend on the level of hunger: starving vultures explore the space widely, whereas fed vultures engage in exploiting high-quality solutions located in their surroundings. Such a proportion evolving does help optimize the balance of exploring and exploiting during the search, which makes the algorithm dynamic and able to jump out of local optimal solutions. In addition, AVOA introduces strategies for enhancement near Pareto solutions, helping speed up the convergence while keeping the diversity in the search population. Due to its simple architecture, limited control parameters, and powerful global search ability, AVOA is very appealing for solving a variety of optimization problems, such as, e.g., engineering design, machine learning parameter tuning, feature selection, scheduling, and real-world decision-making scenarios. Similar to other bio-inspired algorithms, AVOA focuses on adaptability, robustness, and cooperation that are analogized directly from the survival strategies of animals.

Categorical gradient boosting

Categorical gradient Boosting (CatBoost) is a very powerful, high-performance gradient boosting algorithm that Yandex developed, capable of dealing more efficiently with categorical features than ML models. The name CatBoost stands for Categorical Boost, which is its core offering – a near-black-box approach for categorical features that bypasses the traditionally time-consuming feature engineering (using one-hot-encoding in particular). CatBoost also works as a sequence of decision trees, and corrects the accumulated error by the previous trees via gradient descent, like other boosting types. Figure 4 illustrates the Methodological architecture of the CatBoost-based modelling approach. The prediction in CatBoost can be expressed in Eq. 3.

$${\text{y}}_{p} = \mathop \sum \limits_{n = 1}^{N} \gamma_{n} \user2{ }h_{n} \left( {x_{i} } \right)\user2{ }$$
(3)
Fig. 4
Fig. 4
Full size image

Methodological architecture of CatBoost-based modelling approach.

where, \({\text{y}}_{p}\) is the predicted output, \(N\) is the total number of trees, \(h_{n} \left( {x_{i} } \right)\) represents the prediction of the \(n\)-th tree for sample \(i\), and \(\gamma_{n}\) is the learning rate to the \(n\)-th tree.

Light gradient boosting machine (LightGBM) regressor

LightGBM Regressor is an accurate and highly efficient gradient boosting framework for fast regression problems. LightGBM Regressor adds decision trees to an ensemble sequentially, where each added tree is trained to minimize the residuals of the current ensemble, with respect to some loss function. One major difference in LightGBM is the histogram-based algorithm, in which continuous features can be binned into discrete ones, thus greatly lowering the model complexity and accelerating the learning process. Figure 5 illustrates the methodological architecture of the LightGBM regressor-based model.

Fig. 5
Fig. 5
Full size image

Methodological architecture of LightGBM regressor-based model.

Another important feature is to use the new tree growing method "Leaf-wise" instead of the traditional "Level-wise" one. In the leaf-wise method, LightGBM grows by splitting the leaf that decreases the loss the most, resulting in deeper and usually more accurate trees, and maintains the number of splits with the maximum depth parameter to avoid overfitting. The prediction in LightGBM can be expressed in Eq. 4.

$${\text{y}}_{p} = \mathop \sum \limits_{n = 1}^{N} f_{m} \left( {x_{i} } \right)$$
(4)

where, \({\text{y}}_{p}\) is the predicted value for the sample \(i\), \(N\) is the total number of trees, \({f}_{m}({x}_{i})\) represents the output of the \(n\)-th tree.

Evaluation metrics for model accuracy

The performance and accuracy of the developed predictive models were assessed using standard statistical metrics, namely the coefficient of determination (R2), root mean square error (RMSE), mean absolute error (MAE), and standard deviation (SD). To evaluate the predictive performance and reliability of the developed models for estimating the RCP values of GPC incorporating GGBS, SF, FA, and MF. Equations 5, 6, 7 and 8 represent the statistical evaluation metrics employed to assess the performance of the developed predictive models.

$$R^{2} = 1 - \frac{{\mathop \sum \nolimits_{i = 1}^{n} (y_{a} - {\text{y}}_{p} )^{2} }}{{\mathop \sum \nolimits_{i = 1}^{n} (y_{a} - \overline{y})^{2} }}$$
(5)
$${\text{RMSE}} = \sqrt {\frac{1}{n}\mathop \sum \limits_{i = 1}^{n} (y_{a} - {\text{y}}_{p} )^{2} }$$
(6)
$${\text{MAE}} = \frac{1}{n}\mathop \sum \limits_{i = 1}^{n} \left| {y_{a} - {\text{y}}_{p} } \right|$$
(7)
$${\text{SD}} = \sqrt {\mathop \sum \limits_{i = 1}^{n} \frac{{\left( {e_{i} - e} \right)2}}{n - 1}}$$
(8)

where, \(y_{{\text{a}}}\) is the observed values, \({ }y_{{\text{p}}}\) is the predicted values, \(\overline{y}\) is the mean of observed values, and \(n\) is the number of data samples, and ei is the prediction error.

SHapley additive exPlanations (SHAP) analysis for model interpretability

SHAP analysis is a powerful model interpretability technique that helps to determine how each input feature contributes to a machine learning model’s prediction43,44. It is based on cooperative game theory, attributes the prediction output to each feature by calculating Shapley values that reflect the average marginal contribution of a variable over all possible feature combinations45,46. Equation 9 represents the mean SHAP values for the determination of feature importance.

$${\text{SHAP Sensitivity}}_{j} = \frac{1}{n}\mathop \sum \limits_{i = 1}^{n} \left| {\phi_{j}^{\left( i \right)} } \right|$$
(9)

where, \(\phi_{j}^{\left( i \right)}\) = SHAP value of feature \(j\) for instance \(i\), \(n\) = total number of samples, \(j\) = specific feature index.

Results and discussions

Model calibration through hyperparameter tuning

Hyperparameter tuning is a vital stage in the development of robust and high-performing ML models. Unlike internal parameters that are learned during the training phase, hyperparameters determine the overall learning strategy, structure, and convergence behaviour of a model. Proper calibration of these hyperparameters is essential to achieve a balance between bias and variance, ensuring enhanced generalization and predictive reliability47. Table 7 presents the optimised hyperparameter values using the grid search technique and tenfold cross-validation (CV) to improve the predictive model accuracy.

Table 7 Optimized Hyperparameter Tuning for ML models.
Full size table

Model validation using the cross-validation technique

To ensure the robustness and generalization ability of the developed predictive models, a tenfold CV technique was adopted. This CV was applied to evaluate the predictive accuracy of the developed ML models. In this approach, the entire dataset was divided into 10 mutually “folds” of approximately equal size. During each iteration, one-fold was used as the validation set, while the remaining 9 folds were employed for model training. This process was repeated 10 times, ensuring that every data point was used once for validation and 9 times for training. The final model performance was then obtained by averaging the results from all folds48,49. This approach minimizes bias and variance in performance estimation, preventing overfitting and ensuring that the models can effectively generalize to unseen data. The performance of each model in every fold was assessed using statistical metrics including R2, RMSE, MAE, and SD. Consistently high R2 and low error values across folds indicated that the developed models were stable and capable of accurately predicting the RCPT values of geopolymer concrete containing GGBFS, SF, FA and MF. Figure 6 depicts model validation results using 10-CV.

Fig. 6
Fig. 6
Full size image

Model validation results using 10-CV.

Performance of the adaptive boosting model results

Statistical outcomes of RCPT using AdaBoost are represented in Table 8. The training rate of the model also performed optimally, with an RMSE of 82.2632, an MAE of 65.9044, an SD of 269.1698, and an R2 of 0.9282. These indices suggest that the model could predict the RCPT values with high accuracy, according to the low values of RMSE and MAE, which means a small error between the outputs predicted when measured against the actual outputs50.

Table 8 Statistical outcomes of RCPT values using AdaBoost.
Full size table

Where; C: Coulombs

The large R2 suggests that 93% of the variation from the RCPT data could be explained by the AdaBoost model. This is good evidence that the model is learning the training data and is a good fit to the underlying patterns. The value of the standard deviation shows how the variability of the RCPT data confirms that the AdaBoost model was stable and consistent in construction when concrete mixes differ. During the testing phase, the AdaBoost technique proved to be consistent with application on a new dataset not observed before. Testing results revealed an RMSE of 106.3207, MAE of 82.0423, a standard deviation of 264.1722 and R2 of 0.8922. Although the RMSE and MAE slightly increase due to the model transition from training to real testing documents, overall, the results were still very satisfactory. An R2 of 0.8922 indicates that almost 89% of the variance in the RCPT can still be predicted well, which indicates the generalization power of the model. The tight correlation between training and test performances indicates that our AdaBoost model did not overfit the training data and can generalize well to predict a response of new mixes. These findings are of practical importance when assessing contractability. Because the AdaBoost model can predict chloride ion penetration in concrete as a function of mix design only, the model has the potential to lead to substantial savings in time and material demands for experimental laboratory drying tests. The design process of a durable concrete is faster, and the amount of waste material, as well as the testing charge, is reduced. Additionally, the use of AdaBoost in this context contributes to greener construction by allowing for the best use of SCMs, waste materials from industry.

Silica fume, fly ash and GGBS not only improve the performance of concrete but also reduce the usage of Portland cement, which in turn reduces the carbon emissions of concrete production51. Moreover, microfibers enhance the strength of the material and the resistance of the said material to cracks, so that durability is also increased. This study demonstrates the potential of data-driven models to unite empirical testing and advanced analytics and hence make materials engineering decision-making smarter. Also, the information on AdaBoost’s weighting mechanism might provide a measurement of those mix constituents that contribute more significantly to durability, therefore, providing a practical tool of guidance for mix design optimization. The good predictability by AdaBoost (with training R2 = 0.9282 and testing R2 = 0.8922) also demonstrates that the method is appropriate for predicting complex concrete behaviour. Figure 7 depicts actual vs. predicted values of the RCPT values for the AdaBoost model for the training and the testing datasets. Most of the points are located on a 45-degree diagonal line, showing good agreement between predicted and actual values. The high precision and low error of the proposed model are evident from the compact spread outlines of the model points about the line for the two data sets. The small perturbations on the test point are intuitive to note and are an indication of good generalization. A general trend was observed between TP and the chloride penetration model for SCM-based concrete, indicating that the model could be regarded as reliable and predictive for the chloride penetration behaviour of SCM-based concrete involved in the use of AdaBoost.

Fig. 7
Fig. 7
Full size image

Training vs testing analysis of RCPT values using the AdaBoost model.

Figure 8 presents the Residual vs. Predicted value plot for RCPT values prediction using the AdaBoost algorithm. The residuals for training (blue) and testing (red) samples are uniformly scattered around the zero line with no apparent systematic bias. That means the model is a good representation of the data, and we have managed not to overfit. The residual distribution also verifies the effectiveness and generalizability of AdaBoost for predicting chloride ion penetration in SCM-based concrete mixes. Figure 9 shows the Error vs. Frequency curve of the AdaBoost model for RCPT values prediction. The residuals vary around zero with no clear tendency, indicating random and free of systematic error. There are some bumps, particularly in the test data, but within reasonable variance. Description also supports the stability of the model and its generalization for different data samples of SCM-based concrete analysis.

Fig. 8
Fig. 8
Full size image

Residual vs predicted analysis of RCPT values using the AdaBoost model.

Fig. 9
Fig. 9
Full size image

Error vs frequency analysis of RCPT values using the AdaBoost model.

Performance analysis of African vultures optimization algorithm

The performance results of the AVOA model are shown in Table 9. In training, the model achieved an RMSE of 37.1070, representing an average error between the model’s predicted value and the actual housing price value; the lower the value, the better. The MAE, also a prediction accuracy index, was 29.2043, which also indicated the prediction accuracy of the model for the experimental data tendency. The resulting standard deviation of 293.6740 reveals the spread of the errors, which, for this application, lies well within an acceptable realm given the variability of concrete mechanical behaviour under different mix designs. Especially, R2 achieved the high value 0.9854, which indicates the model explained 98.5% of the variance in RCPT during training, and it indicates the model is fit very well. The AVOA model successfully generalized to the testing set as well, showing that the model was not simply memorizing the training dataset. The testing RMSE was 79.2836, although being larger than the training error, it still represents a good predictive accuracy for unseen data. The testing MAE was 61.3408, which illustrates the ability of the model to inherit the low errors in prediction in a realistic problem. The average of the testing prediction results was 281.7324, indicating that the model predictions among the different testing samples were consistent and stable. The R2 testing value was 0.9400, which again indicates that the model was able to account for 94% of the variance in testing RCPT scores. This high testing R2 indicates that the model generalizes well not only for the training data but also in predicting unseen data, an important aspect for real-world applications. The significance of employing AVOA in these processes cannot be underestimated. The durability of concrete, including its resistance to chloride ion penetration, plays an important role in the service life of reinforced concrete structures, especially in marine or deicing salt environments. Traditional experimental methods to investigate this behaviour can be laborious, costly, and are frequently plagued with inconsistency from material variation and human error. Through the implementation of an optimization-based soft computing model such as AVOA, accurate prediction tools can be developed that minimize the need for trial-and-error testing, evaluate hundreds of mix design iterations and facilitate data-driven decision making in the selection and proportioning of materials. Furthermore, the fact that the AVOA had been employed in this research work enhanced the literature, reporting the effectiveness of nature-inspired algorithms in geomechanically applications.

Table 9 Statistical outcomes of RCPT values using AVOA.
Full size table

Where; C: Coulombs

The AVOA was an effective method to develop the models for the RCPT results for SF, FA, GGBS and MF incorporated concrete. Achieving excellent performance in both training and testing sets, AVOA demonstrated its efficiency in modelling the complex inter-relationships between multi-input variables and accurately forecasting the chloride ion penetration. The model’s measured performance indicators are low RMSE, low MAE and high R2 values, suggesting its stability and reliability. Figure 10. Predicted RCPT values for training and testing datasets using the AOVA model. For both training and testing datasets, residuals are mostly centered around the zero line, which implies balanced error distribution and very low systematic bias. The residual spread is roughly the same over the predicted range, showing homoscedasticity and good model fit. Although test residuals have slightly larger spreads, within acceptable limits, and therefore, the reliability and robustness of the AVOA model to capture chloride resistance are confirmed in SCM-based concrete (Refer to Fig. 11).

Fig. 10
Fig. 10
Full size image

Predicted RCPT values for training and testing datasets using the AOVA model.

Fig. 11
Fig. 11
Full size image

Residual vs predicted analysis of RCPT values using AVOA model.

Figure 12 shows the error vs. Frequency graph for predicting RCPT with AVOA. The plot displays how the prediction error changes over individual samples, along with samples’ training errors (blue) and testing errors (red) with sample indices. Training error remains flat and of a low magnitude everywhere for the dataset, which implies a fair consistency in the model. While test errors (red) are a bit more variable in the first part, they flatten out, indicating strong generalization of the model. This is a confirmation that AVOA is efficient in predicting RCPT for SCM-based concrete mixtures.

Fig. 12
Fig. 12
Full size image

Error vs frequency analysis of RCPT values using AVOA model.

Performance analysis of the categorical gradient boosting model

The statistical outcomes of RCPT prediction using the CatBoost model are presented in Table 10. The model demonstrated excellent performance during training, with an RMSE of 12.8736, MAE of 10.1969, an SD of 305.7093, and a high coefficient of determination (R2) of 0.9982. These results suggest that during training, the predictive performance of our model is almost perfect and accounts for more than 99% of the variance of the RCPT estimates. The testing results also validated the high robustness and generalization ability of the CatBoost model. In test data, the performance of the CatBoost model produced an RMSE value of 61.4101, an MAE value of 49.0464, 0.9640 R2 and a SD of 288.9549. Although there was a bit more errors during this phase as compared to the training phase, overall performance stayed high, and the model was still describing around 96.4% of the variance in the test data. This indicated that the prediction capacity of the CatBoost model for new unseen concrete mixes was still strong for RCPT outcomes.

Table 10 Statistical outcomes of RCPT values using Catboost.
Full size table

Where; C: Coulombs

Figure 13 shows for the CatBoost algorithm the Actual vs. Predicted values for RCPT values with the results of training (blue) and testing (red) datasets. The data points are closely positioned along the 45-degree line of the reference line, which means an almost perfect fit between values predicted by the model and the actual values. This model performs very well with very little bias, particularly for the training set, and retains good predictive performance on test data. The Residual vs. Predicted values for the prediction of RCPT values is depicted in Fig. 14. The residuals of training (blue) and testing (red) data are all around the zero line, meaning small bias and symmetric error distribution. The training residuals present very low spread, indicating a good fit to the model, while test residuals present a little more dispersion but remain within reasonable limits. This trend substantiates the outstanding prediction of chloride penetration resistances in SCM-based concretes with an excellent accuracy, robustness, and generalization ability of CatBoost. Figure 15 shows error vs. frequency for RCPT values using CatBoost. It was marked that the test errors (red line) are very unstable as many spikes and falls, showing instabilities on the training frequencies. In contrast to the validation errors (blue line), the training errors (blue line) are low and constant over all frequency ranges. The test errors level off and band closely to the training errors as the frequency grows (above an index of roughly 40). This indicates that the model extrapolates better for higher frequencies beyond the initial unstable phase.

Fig. 13
Fig. 13
Full size image

Training vs testing analysis of RCPT using Catboost.

Fig. 14
Fig. 14
Full size image

Residual vs predicted analysis of RCPT using CATBoost.

Fig. 15
Fig. 15
Full size image

Error vs Frequency analysis of RCPT values using the CatBoost model.

LightGBM Regressor analysis

The statistical outcomes RCPT values using LightGBM Regressor are interpreted in Table 11. The predictive performance of the model was strong, with training statistics of RMSE = 31.5692, MAE = 25.3089, SD = 304.4705, and R2 = 0.9894. These findings suggest the model accounted for 98.94% of the variation in the training data, a very strong fit. The RMSE and MAE were high, which means that it could reduce both large and small errors at the same time.

Table 11 Statistical outcomes of RCPT values using the LightGBM Regressor.
Full size table

Where; C: Coulombs

In addition, the LGBMR model had strong generalization ability in unseen data testing. The testing results of RMSE were 63.2178, MAE 49.3846, and standard deviation 306.5213, and R2 of 0.9618. These results demonstrate that the model had strong prediction yet in a separate set, as more than 96% of the variation in RCPT was also explained. The slight increment of RMSE and MAE in the test phase with respect to the train is typical of a realistic data modelling process and tells that the model is not overtrained, but it is well-generalized. The most distinctive feature is that the standard deviation in testing was nearly equal to that of training, indicating the performance stability of the model between various datasets. The value of LGBMR lies in its speed, scalability, and high prediction accuracy. The durability of concrete is a multiple-parameter phenomenon which is dependent on a combination of mix design factors, curing conditions and interaction among different components. These nonlinear interactions are often not well captured by traditional statistical approaches. Figure 16 depicts the actual Vs predicted values of RCPT by LightGBM. It was stated that a strong linear relationship between the true and predicted values of both training (blue) and testing (red) sets, with the points lying mostly on the dashed diagonal line. This demonstrates that the discovered patterns of the data are effectively learned by the LightGBM model. There is some spread in the data, indicating that there may be some slight offset between actual and predicted values, but a good fit of the data to the model indicates good performance and generalization on test data.

Fig. 16
Fig. 16
Full size image

Training vs testing analysis of RCPT values using the LightGBM model.

Figure 17 represents the residuals as a function of the predicted values for RCPT, and for LightGBM, we can appreciate the residuals in the training (blue) and testing (red) sets spread mostly haphazardly around the zero line. This indicates that the errors of the model are mostly not biased and cannot find systematic patterns through the range of the predictions, as expected. Nevertheless, some exceptions (with larger positive and negative residuals) are present, and they are located mainly in the test set. In general, the points on the residual plot are not “hugging” the horizontal line, suggesting a decent fit with approximately no major heteroscedasticity. Figure 18 depicts the error versus the frequency for the RCPT values with LightGBM model. It was observed that the test errors (red line) exhibit a strong variation, at lower frequencies, with some very large peaks. If altered, the training errors (blue line) are also headed to change, but generally do not spread across the frequency spectrum as do the test errors. This indicates that the model generalizes less well to unseen data, particularly at lower frequencies.

Fig. 17
Fig. 17
Full size image

Residual vs predicted analysis of RCPT values using the LightGBM model.

Fig. 18
Fig. 18
Full size image

Error vs Frequency analysis of RCPT using the LightGBM model.

SHAP analysis and feature importance interpretation

Figure 19 depicts SHAP analysis demonstrating the global feature importance and directional impact of each variable on chloride permeability prediction in GPC mixes. It was observed that GGBS and FA exhibit the most significant influence on RCPT reduction due to improving the concrete’s resistance to chloride ingress and enhanced durability. MF also contributes positively, but its effect is comparatively moderate. Conversely, SF contributes to the significant decrease in RCPT values through its high pozzolanic reactivity and ultrafine particle size, which refine the pore structure and densify the interfacial transition zone (ITZ)52. Additionally, Fig. 20 shows the relative contribution of input variables to RCPT prediction based on SHAP analysis. It was revealed that GGBS demonstrates the highest contribution (58.7%), indicating its dominant role in enhancing chloride resistance. Furthermore, FA and MF contribute 17.2% and 14.3%, respectively, suggesting their moderate influence on RCPT result outcomes. SF exhibits the least contribution (9.8%) in densifying the microstructure and improving long-term durability.

Fig. 19
Fig. 19
Full size image

SHAP summary plot for feature importance.

Fig. 20
Fig. 20
Full size image

Contribution of Mean SHAP values of input variables.

Comparison of experimental and predicted RCPT values

The comparison of experimental and predicted RCPT values based on ASTM C1202 classification is depicted in Fig. 21. The result reveals that the CatBoost model demonstrated the highest prediction accuracy among the four ML algorithms. The distribution of predicted samples across permeability categories — High (> 4000), Moderate (2000–4000), Low (1000–2000), Very Low (100–1000), and Negligible (< 100) that closely aligns with the experimental trends. Notably, the CatBoost model effectively captured the dominant “Low” and “Very Low” permeability ranges, which represent mixes with enhanced durability and reduced chloride ion penetrability. The minimal deviation between experimental and predicted proportions in these categories indicates CatBoost’s superior ability to generalise complex nonlinear interactions between binder components and microstructural behaviour influencing RCPT outcomes. In contrast, AdaBoost and AVOA showed slightly higher variation in the Moderate range and High categories, implying limited sensitivity to subtle microstructural effects. The LGBMR model also performed well but exhibited marginal overprediction in the “Very Low” class (Fig. 21).

Fig. 21
Fig. 21
Full size image

Comparison of experimental and predicted RCPT values.

Computational efficiency and practical scalability

The comparison shows that CatBoost achieved the best trade-off between computational efficiency and predictive accuracy for RCPT values estimation. While AVOA required the longest training time due to iterative optimization, LightGBM provided faster training but slightly lower interpretability. CatBoost’s moderate training time, combined with its GPU compatibility and robust handling of nonlinear relationships, makes it ideal for engineering-scale durability prediction tasks and scalable integration into real-time material design frameworks. Table 12 indicates comparative computational performance and scalability of ML Models.

Table 12 Comparative computational performance and scalability of ML models.
Full size table

Comparison between experimental and computational durability assessment

Traditional experimental durability tests such as RCPT, sorptivity, and chloride diffusion are reliable but require extensive time, resources, and laboratory control, which often limit their practical applicability for large-scale mix evaluation. These tests may also exhibit variability due to differences in curing conditions and operator handling. In contrast, ML-based approach developed in this study offers a rapid and cost-efficient alternative, capable of accurately predicting durability indices like RCPT based on input material properties. Once adequately trained, ML models such as CatBoost, AdaBoost, AVOA, and LGBMR can perform instantaneous predictions and support mix design optimization without repetitive experimental trials. Thus, integrating ML with experimental data provides a hybrid, intelligent durability assessment framework that enhances predictive capability, reduces experimental dependency, and facilitates sustainable material design in modern construction practice. Table 13 describe the comparison between the experimental and ML-based durability assessment approach.

Table 13 Comparison between experimental and ML-based durability assessment approach.
Full size table

Integration of ML models into real-world design workflows

The encouraging predictive performance of the developed ML models, particularly CatBoost and LightGBM Regressor, demonstrates their potential to be embedded into practical concrete mix design workflows. In real-world engineering practice, these models can serve as decision-support tools that assist engineers in optimizing binder proportions, fibre content, and curing conditions to achieve target durability levels, such as those classified under ASTM C1202. By inputting key material parameters, it can instantly predicted RCPT values and corresponding durability classifications, reducing reliance on laborious and time-consuming laboratory testing.

Integration into mix design software or web-based predictive dashboards would allow continuous updating of models with new experimental data, enhancing reliability and adaptability across different project conditions. Moreover, coupling these predictive tools with optimization algorithms can automate the selection of cost-effective and eco-friendly mix proportions that meet desired durability thresholds.

Conclusions

This study successfully demonstrates the effectiveness of soft computing techniques in predicting chloride permeability of SCM-based concrete. The following important conclusions were derived.

  1. (1)

    The application of advanced soft computing techniques— AdaBoost, CatBoost, AVOA, and LightGBM of comparison research for estimating RCPT values, enabling efficient and accurate assessment of chloride resistance in SCM-based geopolymer concrete.

  2. (2)

    Among the evaluated models, CatBoost demonstrated superior performance, achieving near-perfect training accuracy (R² = 0.9982) and robust predictive capability on the testing dataset (R² = 0.9640), indicating its suitability for durability assessment of GPC.

  3. (3)

    SHAP interpretability analysis revealed that GGBS and silica fume were the most influential factors in reducing chloride permeability, followed by fly ash and microfibre dosage. This emphasizes the synergistic effect of SCMs and microfibres in improving the durability performance of GPC.

  4. (4)

    ASTM C1202 classifications confirmed that the predicted RCPT values accurately reflected practical permeability ranges (very low to moderate), validating the model’s engineering applicability.

  5. (5)

    The computational approach outperformed traditional experimental methods in efficiency, requiring only material input data for prediction while eliminating repetitive RCPT trials, thereby saving significant time and resources in durability assessment.

  6. (6)

    The study promotes sustainable construction by minimizing cement consumption and utilizing industrial by-products (FA, GGBS, and SF), thereby reducing the environmental footprint, conserving natural resources, and enhancing the durability and performance of GPC.

Limitations and future perspective

In this study, all geopolymer concrete mixes were cured at 80 °C for 24 h, which is a standard practice to accelerate the geopolymerization process and achieve early strength development. However, such elevated temperature curing conditions are not always feasible under field or on-site construction environments, particularly for large-scale structural elements and pavement applications. The reliance on heat curing may limit the direct applicability of the developed mix designs and predictive models to real-world ambient conditions.

Moreover, the variability in the physicochemical properties of supplementary cementitious materials (SCMs) can influence model generalization. For example, Class F (low calcium) and Class C (high calcium) fly ash differ significantly in CaO content, glass phase composition, and pozzolanic reactivity. Class C fly ash exhibits self-cementitious properties and promotes early strength, while Class F primarily contributes to long-term strength and permeability reduction through secondary hydration reactions. These compositional differences modify pore structure development, chloride ion transport, and overall durability, thereby affecting the model’s internal relationships between mix design parameters and RCPT outcomes. To improve model transferability, future work should include a wider dataset encompassing different classes and sources of SCMs, followed by external validation using independent experimental datasets from other laboratories or published studies. Such validation will ensure the robustness and applicability of the developed predictive models across diverse material compositions and environmental conditions.

Model accuracy also relies on the quality and quantity of experimental data; limited datasets may affect generalization. Future work should focus on developing larger and more diverse datasets encompassing varying binder compositions, fibre dosages, and curing regimes. This will further strengthen the practical relevance of the developed CatBoost-based RCPT prediction model for real-world concrete durability design. Besides RCPT, other tests such as sorptivity, water absorption, and chloride diffusion coefficient should be predicted to develop multi-output models for holistic durability assessment. Future work may also integrate ML-based durability prediction with life cycle assessment (LCA) or embodied carbon analysis to promote sustainable geopolymer concrete design.