Hybrid Machine learning-based modeling to predict and optimize the compressive strength of electric arc furnace slag-modified concrete Authors: Md Alhaz Uddin, Md Habibur Rahman Sobuz, Md Kawsarul Islam Kabbo, Ibrahim Y Hakeem, Ratan Lal, Ali Masria, Sani Aliyu Abubakar Published: 2026 ABSTRACT AbstractThe growing worldwide population has increased the use of electric arc furnaces (EAF), resulting in a surge of EAF slag and a huge environmental concern. EAF slag’s complex physical qualities have a considerable impact on concrete’s mechanical performance, mainly its compressive strength (CS). This study introduces a novel framework for forecasting the CS of EAF slag concrete that uses advanced machine learning models such as Extreme Gradient Boosting (XGB), AdaBoost (ADB), Random Forest (RF), Hybrid XGB-RF, and Hybrid XGB-ADB. A full dataset of 730 samples was meticulously created, containing essential input parameters such as binders, aggregates, and other necessary variables, with CS as the desired outcome. Based on the findings, the XGB model showed highest accuracy, with an test R2 of 0.951, MAPE of 1.128, and RMSE of 1.393 MPa, indicating its potential for dependable performance forecasting. In addition, the hybrid XGB–RF model also demonstrated strong predictive accuracy, achieving an R2 value of 0.947 during the testing phase. Furthermore, explainability tools such as SHapley Additive ExPlanations (SHAP) and partial dependence curves (PDCs) identified the curing period and content of cement as the most influential factors to predict the CS of EAF slag concrete. The methodologies and outcomes of this study will help to reduce reliance on resource-intensive experimental methods, pave the way for efficient, precise, and ecologically conscientious concrete design.Keywords: Arc furnace slag concrete, Compressive strength prediction, Hybrid machine learningSubject terms: Engineering, Materials science, Mathematics and computing Introduction In contemporary research, there has been a marked increase across the globe emphasis placed on the combination of waste materials into concrete and mortar1–4. Research has demonstrated that incorporating extra ingredients into cement-based concrete improves its properties and significantly aids in environmental conservation5–9. The manufacturing of steel, which is a significant industrial activity, results in the production of a by-product known as slag10,11. This by-product has been successfully exploited in the building industry to reduce its impact on the environment11–13. Besides this, construction and demolition waste has also gained significant attention in the development of modified sustainable concrete, with notable progress in Asian, European and Middle East (such as Najran city) regions14–17. Blast furnace slag (BFS), electric arc furnace oxidizing slag (EAF), and ladle furnace slag (LFS) represent typical forms of slag that are frequently utilized in the construction industry18–22. Some of these slag types are derived from the procedures that are used to manufacture steel. Researchers from all over the world have investigated the possible applications of electric arc furnace (EAF) slag in civil engineering and construction23–25. This is due to the enormous quantities of EAF slag that are produced, which are estimated to constitute between 10 and 15% of the total weight of steel production26. Rondi, et al.27 studied the utilization of EAF slag as a complete substitute for aggregates in concrete and discovered that higher specific density and water absorption than concrete using natural aggregates. However, the slag showed reduced volumetric expansion and an acceptable quantity of contaminant leaching. Sezer and Gülderen28 found that concrete incorporating EAF slag as coarse aggregate exhibited reduced water penetration depth during pressure testing; however, its use as fine aggregate resulted in higher penetration depths compared to conventional limestone aggregate mixtures. González-Ortega, et al.29 found that water absorption within concrete containing EAF slag was lower at 48 h compared to conventional concrete, with absorption reducing as limestone coarse particles were replaced by EAF slag. Santamaría, et al.30 observed that concrete incorporating EAF slag exhibited greater resistance to water penetration under pressure compared to combinations employing natural aggregates. Their findings also demonstrated that steel rebars embedded in conventional concrete had higher electrochemical corrosion rates than those in EAF slag-based concrete, according to ASTM C-876. The existence of free CaO in EAF slag, which contributes to Ca-ion leaching, was thought to be favorable for steel reinforcing preservation over time. de Lima31 conducted carbonation depth tests under accelerated and laboratory circumstances and discovered that replacing typical fine aggregates with EAF slag lowered carbonation depth. However, replacing coarse particles with EAF slag resulted in deeper penetrations than in the reference concrete. Ortega-López, et al.32 investigated water absorption via capillarity in EAF slag-based concrete with fiber reinforcement and without fiber reinforcement that had been moist-cured for 90 days and found substantial differences. ARRIBAS GARCIA33 investigated steel bar corrosion in EAF slag concrete exposed to a sea environment for a year and discovered that these specimens had similar corrosion behavior to those made with limestone aggregates. In concrete research, machine learning (ML) has become established as a potent statistical technique for the prediction of critical attributes such as strength, durability, and workability, resulting in increased attention34–37. Without the need for human involvement, ML models can discover trends and make intelligent predictions based on historical data38–41. These algorithms excel in analyzing nonlinear and multidimensional correlations between multiple parameters, resulting in very accurate forecasting42–44. Compared to traditional experimental methods, ML-driven approaches improve the precision of concrete property predictions by reducing errors45–47. Several machine learning (ML) methods, including Extreme Gradient Boosting (XGB)48, LightGBM (LGB)49, Gradient Boosting Machines (GBM)50, random forest (RF)51, Decision Trees (DT)52, Artificial Neural Networks (ANN)53, AdaBoost54, and K-Nearest Neighbors (KNN)55, have been used in studies. In addition to standalone algorithms, researchers have created hybrid models that integrate the best features of different methodologies to improve forecast accuracy56–58. Recent studies from Najran city using date palm fibers in high-performance concrete also highlighted the potential of regional waste materials and the need for ML-based models to capture their strength behavior59. Other techniques, such as Support Vector Machines (SVM)60, Convolutional Neural Networks (CNNs)61,62, and multi-model based approaches, have been investigated for their ability to simulate the nonlinear and complex interactions that influence tangible attributes63,64. These advances highlight the critical role of machine learning (ML) in increasing the predictive accuracy and operational efficiency of predictive models for concrete strength. Güçlüer, et al.65 found that Decision Tree (DT) models performed well, with strong correlation (R² = 0.86) and low mean absolute error (MAE = 2.59), especially for samples cured for 28 days. Ensemble models such as GS-XGB obtained over 99% accuracy in 28-day CS predictions. These models are further supplemented by user-friendly software platforms that aim to simplify material design and testing operations66–68. The Categorical Gradient Boosting (CatBoost) approach performed well with an R2 of 0.966 and an RMSE of 3.06 MPa, indicating its potential to simplify assessments and reduce experimental costs69. With an R2 of 0.9356 and an RMSE of 5.5604 MPa, the GB_PSO model optimized for recycled concrete outperformed individual methods. Furthermore, evaluations stressed the importance of cement content in maximizing strength70,71. In another study, Multiple Linear Regression (MLR) proved highly effective in predicting the compressive strength of concrete with Electric Arc Furnace (EAF) slag, with a minimum Mean Absolute Percentage Error (MAPE) of 2.20% and a minimum Mean Squared Error (MSE) of 46.95, demonstrating MLR’s robustness in such applications72. Furthermore, the Grey Model (GM) emerged as a powerful tool for prediction of compressive strength, achieving remarkable results with a minimum MAPE of 0.9% and MSE of 8, and outperforming ANN, particularly with limited data—GM successfully predicted strength with as few as four data points, demonstrating superior performance for scenarios with constrained datasets73. These findings show the potential of ML-based models for improving concrete strength prediction and providing efficient, reliable, and data-adaptive solutions to the construction sector. Although numerous studies have successfully applied machine learning techniques to predict the CS of conventional concrete, limited attention has been given to concrete incorporating EAF slag. Existing investigations on EAFS-based concrete are often constrained by small datasets, or a limited number of input variables. Moreover, few studies have integrated advanced interpretability tools to systematically explain the influence of individual mixture components on strength development. To address these gaps, this study develops a large and structured dataset comprising 730 samples collected from published literature, incorporating eight critical input parameters. The novelty of this research lies in multiple primary aspects. First, it establishes one of the most comprehensive datasets specifically focused on EAFS-modified concrete for CS prediction. Second, it performs a systematic comparative assessment of standalone ensemble models (ADB, RF, and XGB) alongside hybrid stacking-based models (XGB-ADB and XGB-RF), which enables a robust evaluation of predictive performance. Third, it integrates parametric modeling techniques, including SHAP analysis and partial dependence plots (PDPs), to quantify feature importance and interpret nonlinear interactions among mixture constituents. Through this comprehensive framework, the study advances predictive modeling for EAFS-based concrete and provides interpretable, data-driven insights to support the design of high-performance and environmentally sustainable construction materials. Research methodology The most useful machine learning model for forecasting the CS of EAFS-based concrete was determined by testing five alternative approaches, including two hybrid models. The following machine learning techniques were applied in this work: (1) ADB, (2) XGB, (3) RF, (4) XGB-ADB, and (5) XGB-RF. The dataset was divided such that 80% was used for training, 10% for testing and remaining 10% for validation. Here, the validation set was employed to further verify model robustness and consistency on unseen data, thereby reducing the risk of overfitting. All analyses were conducted using the Anaconda Navigator distribution, where an integrated Jupyter Notebook environment (version 7.2.2) was employed for workflow execution. The programming language used was Python (version 3.12.7). Furthermore, widely recognized and well-documented Python libraries were utilized for different stages of the study. Data preprocessing and numerical computations were performed using NumPy and Pandas, while visualization tasks were carried out using Matplotlib and Seaborn. ML model development, training, and evaluation were implemented using Scikit-learn, along with model specified standard libraries. In addition, SHAP library was employed for model interpretability analysis. Figure 1 depicts the overall workflow of the ML implementation, with detailed step-by-step representations of the process. Overall flow diagram of the work. Dataset development and data preprocessing The dataset used in this study consists of 730 data points compiled from previously published peer-reviewed literature. Furthermore, the statistical attributes of all the input and output features have been provided in Table 1. The collected studies were carefully screened to ensure consistency in reported units, testing standards, and curing conditions. All CS values were converted to MPa, and mixture proportions were standardized to the same unit to maintain dimensional uniformity across the dataset. Furthermore, several preprocessing steps were performed prior to model development. The dataset was first examined for missing, duplicate, or incomplete entries, and inconsistent samples were removed to preserve data integrity. Only experimental results obtained under comparable laboratory conditions were retained. These conditions included standard curing regimes and conventional CS testing procedures. This filtering process minimized variability arising from incompatible experimental protocols and ensured that the dataset represented similar mechanical and environmental conditions. Table 1Statistical analysis of the input and output features.FeaturesUnitMeanMin.Max.St. Dev.Co. of Var (%)OPCkg/m3311.32108.3540101.0832.47Arc Furnace Slagkg/m359.920359.483.06138.61SCMskg/m344.340174.759.47134.13Fine Aggregatekg/m3774.89594992.684.3410.88Coarse Aggregatekg/m3987.53801114574.237.52Water contentkg/m3179.52126.622822.1112.32Superplasticizerskg/m35.86032.26.44109.79Curing PeriodDays52.20136572.98139.80CSMPa38.354.7882.617.3845.31 Statistical analysis of the input and output features. Outlier detection was then conducted to reduce the influence of extreme or physically unrealistic observations. Statistical screening based on distribution analysis and interquartile range criteria was applied to identify abnormal data points. Following data cleaning, min-max normalization was applied to all numerical input variables. The processed dataset was then randomly divided into 80%-20%. 80% of the data was allocated for model training. The remaining 20% was reserved for testing (10%) and validation (10%) to evaluate model generalization performance on unseen data. Extreme gradient boosting (XGB) XGB represents an advanced machine learning algorithm renowned for its exceptional capacity to generate precise forecasts. The main basis of its boosting method is to train iteratively learners progressively, with each learner emphasizing on rectifying the mistakes made by its predecessor. During this procedure, each succeeding model is fine-tuned to the preceding one’s residual errors, and these iterative adjustments are combined to improve the model’s overall forecast accuracy. This step-by-step correction greatly strengthens the projections74. Furthermore, XGB uses effective resource management for both CPU and memory, resulting in faster computation and shorter processing times. Figure 2(a) depicts a simplified illustration of the mathematical formulation underpinning this prediction process, as defined in Eq. 1. Training sequence for (a) XGB, (b) ADB and (c) RF. Where the XGB model is denoted by , and the structurally learners engaged in the procedure are recognized by . Adaptive boosting (ADB) As seen in Fig. 2(b), ADB Regression is a sophisticated ensemble learning technique intended to improve regression job accuracy. This technique combines multiple weak regression models into a cohesive ensemble, creating a more reliable and robust prediction system75. Unlike conventional methods, ADB Regression adapts dynamically to diverse data patterns. During each cycle, the algorithm adjusts the weights of erroneously classified examples, compelling following models to focus on improving predictions for these difficult contexts. The final ensemble aggregates individual models using weighted contributions, reflecting their performance. This adaptive process fosters continual improvement, enabling ADB Regression to handle complex datasets and variable interactions effectively. Its capacity to yield precise and dependable predictions makes it particularly suited for scenarios involving challenging data structures and intricate patterns76,77. Random forest (RF) RF is an integrated-models learning approach that amplifies model performance by mixing numerous decision trees, each trained on a distinct subgroup of the original data via “bagging” or bootstrap aggregation. Bagging entails sampling with replacement, in which certain samples appear several times while others are omitted, hence reducing model variance by averaging predictions from these independent trees, as explained by Breiman78, shown in Fig. 2(c). RF computes splits using a randomly selected subset of features at each node, reducing computational load and increasing performance, making pruning unnecessary. The computational complexity of RF is proportional to Eq. (2): Where T represents the number of trees, M is the number of features per split, and N is the size of the training sample78. In the absence of a separate test set, RF employs “out-of-bag” (OOB) samples—roughly 33% of the training set that is excluded during each resampling—as impartial indications of accuracy. Votes from each OOB sample are counted across trees to determine the final class. Hybrid machine learning modeling This paper describes a fusion approach that combines machine learning techniques to improve predicted accuracy using ensemble modeling. Two advanced variants, XGB-ADB and XGB-RF, were created using stacking techniques. The XGB-ADB model incorporates the strengths of two gradient-boosting techniques, XGB and ADB, to increase accuracy by addressing complex data patterns together. During the first stage, each base learner was independently trained using the same training dataset and hyperparameters. After training, the predicted compressive strength values produced by the base learners were extracted and used as new input features for the second stage of the stacking process. In the final stage, a meta-learner (Linear regression model) was trained on these prediction outputs to generate the final compressive strength estimate. This approach allowed the hybrid models to combine the individual strengths of each algorithm while reducing the limitations associated with standalone models. Hyperparameter optimization In this study, hyperparameter tuning was performed to improve model accuracy and increase generalization ability. Each ML algorithm was configured within a predefined search space, and combinations of influential parameters were systematically evaluated to identify the most effective configuration. A grid-based search strategy was implemented to explore different parameter settings in a structured manner. This approach enabled the evaluation of multiple ML models under consistent conditions and ensured a fair comparison among algorithms. The selection of optimal hyperparameters was based on minimizing prediction error while maintaining consistency between training and validation performance. The optimization process also contributed to improved computational stability and balanced model complexity. The finalized hyperparameter settings for all implemented models are summarized in Table 2. Table 2Optimized hyperparameter for all ML models used in this study.ModelHyperparameterRangeOptimizationXGBn_estimators100–500100learning_rate0.01–0.30.05max_depth3–103gamma0–50reg_alpha0–10.1reg_lambda0–51RFn_estimators50–500300max_depth5–2010min_samples_split2–103min_samples_leaf1–53max_features‘auto’, ‘sqrt’, ‘log2’‘sqrt’ADBn_estimators50–500150learning_rate0.01–1.00.1loss‘linear’, ‘square’, ‘exponential’‘square’ Optimized hyperparameter for all ML models used in this study. SHapley additive explanations (SHAP) SHAP values have acquired widespread recognition in recent years as an effective way for evaluating machine learning model predictions79. This technique is well-known for its capacity to explain complex black-box models. SHAP applies Shapley values, which originated in collaborative game theory, to machine learning80. In machine learning, SHAP refers to features as “players” and the model’s output as “reward.” Each feature is assigned a score indicating its relevance to the total forecast80. This methodology not only finds the most important traits but also explains how these factors influence model outcomes81. Visualization techniques, such as summary plots and feature importance graphs, expand their utility by providing explicit insights into model behavior and decision-making processes82. Partial dependence plots (PDPs) With all other parameters held constant, PDPs are frequently employed to investigate the link between certain features and machine learning model predictions. These graphs can demonstrate if a feature’s relationship to the model’s output is linear, nonlinear, or influenced by interactions with other variables83. PDPs improve model transparency by isolating the effect of specific parameters, making them an essential tool for data-driven decision-making. They let users decide if the relationship between an input feature and the model response is simple or complex. This method computes the average predicted output for specific feature values over the whole dataset, maintaining consistency while isolating the feature’s influence. The approach can be extended to investigate the effects of two traits at the same time. PDPs are powerful interpretive tools that provide precise information about how individual or combination characteristics influence model predictions. Evaluation metrics for prediction performance Three different statistical techniques were used at three phases to evaluate the algorithms’ prediction accuracy: training, validation, and testing. Root mean square error (RMSE), mean absolute error (MAE), and mean absolute percentage error (MAPE) are these factors that determine statistical performance. The degree to which the forecast and the experimental result match is also assessed using the coefficient of determination (R2). The evaluation metrics were selected because they provide complementary and comprehensive insights into regression model performance. RMSE penalizes larger prediction errors more severely, which makes it effective for identifying models sensitive to high-magnitude deviations in CS. MAE offers a direct and interpretable measure of average absolute error and MAPE expresses prediction error in percentage terms, enabling relative error comparison across different strength ranges and improving interpretability for engineering applications. Furthermore, R2 evaluates the proportion of variance in CS explained by the model, thereby indicating goodness-of-fit and predictive consistency84,85. Equations (3–5) explain the calculation procedure of these performance metrics to evaluate the prediction accuracy. In this study, denotes the observed values, signifies the predicted values, i represents the average, and m corresponds to the total data points. Results and discussion Analysis of dataset The contour plots of each concrete input variable’s density against compressive strength are shown in Fig. 3. Figure 3(a) highlights that the association between cement and strength has a dense cluster of data points (300–390 kg/m³). Figures 3(b) and (c) display a substantial cluster of data points for SCMs and slag, respectively, between 0 and 15 and 0 to 25. Figures 3(d) and (e) indicate a cluster of data points between 750 and 800 kg/m³ and 930 and 990 kg/m³ for fine and coarse aggregate, respectively. A clear correlation between the amount of water and the compressive strength of concrete can be shown in Fig. 3(f). In Fig. 3(g), the strength values associated with superplasticizer (SP) amounts vary from 0 to 2 kg/m³. Relationship between individual concrete ingredients and compressive strength. Correlation between Input parameters The coefficient typically varies between − 1 and 1. A value greater than 0 indicates a positive correlation between the two variables. Whereas a negative number signifies a negative correlation between them86. The Pearson correlation is a statistical measure used to assess the strength and direction of a linear relationship between two variables75. It is fundamental to acknowledge that Pearson correlation serves as a useful tool for evaluating linear relationships. The Pearson correlation, while a valuable tool for analyzing linear interactions, may fail to reveal nonlinear correlations. The correlation matrix depicted in Fig. 4 demonstrates relationships between input and output variables, with R-values ranging from − 0.51 to 0.46. The heatmap indicates a positive correlation between the four input attributes and the output parameter, compressive strength, and a negative correlation between the four input variables and strength. The highest positive correlation (R = 0.45) is observed between the superplasticizer and strength, suggesting a high degree of consistency with the strength measurement. Cement also shows a consistent correlation with strength, with R value of 0.42. As expected, water has a strong negative relationship with strength as the strength reduces with the addition of water. It should be noted that arc furnace slag has a positive linear correlation with strength, indicating the incorporation of slag within a certain range have a beneficial effect on the strength. Although the Pearson Correlation analysis provides useful information regarding the linear relationships among input variables and compressive strength, it cannot capture nonlinear interactions or complex multivariable dependencies. Therefore, additional interpretability tools such as SHAP and PDP were employed to better understand the nonlinear behavior of the dataset. Heat map illustrating the correlation between input and output variables. Error distribution of ML models Figure 5 presents the error distribution of all developed models across training, testing, and validation stages. Overall, the predicted values show strong agreement with the experimental CS results for all models, although variations in error magnitude are observed. As can be seen in Fig. 5(a), the XGB model demonstrates the most consistent performance. During training, it yields an average error of − 0.000065 MPa, with minimum and maximum errors of − 24.83 MPa and 8.94 MPa, respectively. In the testing phase, the average, minimum, and maximum errors are 0.68 MPa, − 8.82 MPa, and 16.30 MPa. Similar behavior is observed in validation, with corresponding values of − 0.37 MPa, − 10.73 MPa, and 11.47 MPa, indicating stable generalization. Figure 5(b) showed that, the ADB model produces predictions that closely follow the experimental values but exhibits slightly greater variability. For training, the maximum, minimum, and mean errors are 15.60 MPa, − 17.40 MPa, and − 1.31 MPa. Testing errors range from − 16.59 MPa to 22.04 MPa, with an average of − 0.99 MPa, while validation errors vary between − 15.66 MPa and 17.87 MPa, with a mean of − 1.53 MPa. Furthermore, the RF model (in Fig. 5c) also shows strong alignment between predicted and observed values. In the training phase, the maximum, minimum, and average errors are 14.15 MPa, − 25.05 MPa, and 0.09 MPa. During testing, errors range from − 13.57 MPa to 11.32 MPa, with a mean of 0.77 MPa. Validation results indicate an average error of − 0.24 MPa, with minimum and maximum deviations of − 12.16 MPa and 14.05 MPa. Error distribution of (a) XGB, (b) ADB, (c) RF, (d) XGB-ADB and (e) XGB-RF model for training, testing, and validation stage. For the hybrid XGB-ADB model, the training phase records maximum, minimum, and average errors of 17.43 MPa, − 23.08 MPa, and − 0.06 MPa, respectively. The testing and validation stages maintain deviations within acceptable ranges, confirming improved stability compared to standalone ADB. Similarly, the hybrid XGB-RF model exhibits balanced error behavior, with training errors of 13.33 MPa (maximum), − 25.33 MPa (minimum), and − 0.14 MPa (average), while testing and validation errors remain within a comparable and controlled range. The error distributions of hybrid models are illustrated in Fig. 5(d) and (e). Overall, the error distribution analysis indicates that XGB and its hybrid variants demonstrate more stable and lower average deviations, whereas ADB shows comparatively higher variability. Predictive accuracy of ML models Figure 6 presents the scatter plots of observed versus predicted compressive strength values for all developed models. The plots demonstrate the degree of agreement between experimental measurements and model predictions. As can be seen, the XGB model shows strong predictive capability, with R2 values of 0.993 for training, 0.951 for testing, and 0.945 for validation. These high coefficients confirm the robustness and generalization ability of the model across unseen data. The ADB model yields comparatively lower R2 values. Its performance remains acceptable, with R2 values of 0.856 (training), 0.836 (testing), and 0.816 (validation). However, the reduced coefficients indicate weaker predictive strength relative to XGB. The RF model improves prediction accuracy compared to ADB and performs competitively with XGB. It achieves R2 values of 0.981 for training, 0.926 for testing, and similarly high consistency during validation. This demonstrates its suitability for compressive strength estimation. Linear correlation between the experimental and predicted values for the (a) XGB, (b) ADB, (c) RF, (d) XGB-ADB and (e) XGB-RF model. The hybrid XGB-ADB model further enhances predictive reliability, with R2 values of 0.988 (training), 0.926 (testing), and 0.947 (validation). These values surpass those of standalone ADB and RF models and approach the performance of XGB. Similarly, the hybrid XGB-RF model delivers strong and stable performance, achieving R2 values of 0.991 for training, 0.947 for testing, and 0.942 for validation. Its predictive accuracy remains comparable to the individual XGB model and slightly superior to the standalone ADB and RF models. Overall, the scatter analysis confirms that XGB and its hybrid variants provide the most reliable and consistent predictions, while ADB demonstrates comparatively lower predictive strength. Comparing and validating the performance of ML models The Taylor diagram is an effective tool for visually contrasting the performance of several models with experimental data. It computes the correlation coefficient (R) and standard deviation to assess the extent to which the predicted values correspond with the experimental data. A perfect match between predicted and experimental values in terms of relationship and variability is indicated by an R-value of 1. Figures 7 (a) and (b) display Taylor diagrams for five models applied to training and test datasets, each demonstrating a correlation coefficient exceeding 0.93 and a standard deviation between 13 and 18. For training data, all models, including XGB, RF, XGB-ADB, XGB-RF demonstrate the same and the highest correlation coefficients i.e., greater than 0.99 and standard deviations in the range of 16 to 18. Conversely, the ADB model exhibits some deviation from other models, confirming that its predictions are generally consistent based on experimental data but display certain variability. For this model, the correlation coefficient was around 0.93, with a standard deviation of 14. For test data, three models, including XGB, XGB-ADB, and XGB-RF attained the same and the highest correlation coefficients i.e., greater than 0.97 and degree of variation in the range of 17 to 18. The RF model has a correlation coefficient of 0.96. and a standard deviation of 16. The ADB model differs slightly from the other models. However, its predictions closely match the experimental results but show slightly higher variability. For this model, the correlation coefficient was around 0.93, with a standard deviation of 13 to 14. Taylor plot for train and test data of all models. Figure 8 displays four radar maps that evaluate the performance of five models: XGB, ADB, RF, XGB-ADB, and XGB-RF. The charts use four different metrics: R2, RMSE, MAE, and MAPE. The XGB model in Fig. 8(a) has an R2 of 0.993 for training, 0.951 for testing, and 0.945 for validation. This was the highest R2 value of all the models. Whereas, the ADB model has the lowest R2 values, such as 0.856, 0.836, and 0.816 for the training, testing, and validation stages. By contrast, the other three models, including RF, XGB-ADB, and XGB-RF have equivalent R2 values (0.981–0.991 for training, 0.926–0.947 for testing, and 0.931–0.947) like the XGB model. Figure 8(b) depicts the RMSE of various models. During the training phase, the RMSE values for the XGB, XGB-ADB, and XGB-RF models are quite similar (1.3932–1.8929). Conversely, the RF, XGB-ADB, and XGB-RF models exhibit comparable RMSE values (4.0010–4.7445) during both the testing and validation phases. The ADB model exhibits higher error metrics, with a training RMSE of 6.5801, a testing RMSE of 7.0950, and a validation RMSE of 7.1872. In contrast, XGB exhibits the lowest RMSE, with values of 1.3932 for training, 3.8707 for testing, and 3.9350 for validation, indicating superior accuracy and reliability relative to other models. The test set’s models for MAE are ranked as follows: XGB, XGB-RF, XGB-ADB, RF, and ADB as shown in Fig. 8(c). Thus, as the most promising predictive model, the XGB model performs the best, producing optimal results across all metrics. The mean difference in percentage between the expected and experimental values was determined utilizing an alternative method, MAPE, as illustrated in Fig. 8(d). The XGB model exhibits the lowest error values, with a training MAPE of 1.1286, a testing MAPE of 9.3101, and a validation MAPE of 9.5179. Whereas, the ADB model has the highest MAPE values of 20.3658, 23.0521, and 21.5627 for the training, testing, and validation phases, respectively, which indicated less reliability in predicting the data compared to other models. In addition, Table 3 provides a detailed summary of the model’s performance metrics, comparing the training, testing, and validation stages. Performance evaluation of used models. Table 3Performance of ML algorithms.Models R 2 RMSE (MPa)MAE (MPa)MAPE (%)TrainTestValidationTrainTestValidationTrainTestValidationTrainTestValidationXGB0.9930.9510.9451.393.873.930.362.662.551.129.319.51ADB0.8560.8360.8166.587.097.185.505.495.9320.3623.0521.56RF0.9810.9260.9312.354.744.381.433.532.934.5812.7311.54XGB-ADB0.9880.9260.9471.644.004.010.812.902.542.5210.349.68XGB-RF0.9910.9470.9421.894.004.231.192.792.873.659.9210.38 Performance of ML algorithms. Table 4 presents a comparative evaluation of the proposed models against previously published ML approaches for CS prediction. The comparison is based on commonly reported performance indicators to ensure consistency and transparency. This analysis enables validation of the present modeling framework within the broader research context and highlights its relative predictive capability. The results demonstrate that the developed XGB and hybrid models achieve competitive, and in several cases superior, performance compared to existing ML-based studies, confirming the robustness and effectiveness of the proposed methodology. Table 4Comparing and validating the performance of ML algorithms.MaterialsML techniquesFindingsRef.GGBFSDT, RF, SVM, KNN, ANNThe DT model achieved the highest accuracy with R2 = 0.96.87Steelmaking slagSVR, ANN, XGB, and GPR.XGB model outperformed other algorithms with an R2 of 0.93.88Fly ash and slagRF, XGB and SVMXGB showed better performance with R2, RMSE and MAE are 0.95, 0.05 and 0.15, respectively.89Fly ash and BFSADB, XGB, CatB, and SVMAmong all four models, CatB exhibited superior accuracy with an R2 of 0.92 and an RMSE of 0.03.90Fly ashLinear regression, SVM, DT, RF, XGB, ADB.XGB exhibited highest performance with a coefficient of determination (R2) of 0.95.91EAF slagXGB, ADB, RF, XGB-ADB and XGB-RFXGB (R2 = 0.951) and XGB-RF (R2 = 0.947) showed balanced and superior performance comparing R2 and RMSE value.This study Comparing and validating the performance of ML algorithms. *GGBFS = Ground granulated blast furnace slag, GPR = Gaussian Process Regression, CatB = Categorial boosting, SVM = Support vector machine, BFS = Blast furnace slag, KNN = K-Nearest Neighbors, DT = Decision tree. SHAP analysis Figure 9(a) illustrates the mean absolute SHAP values, with curing duration having the biggest effect (mean SHAP = + 9.52). This illustrates that it has the biggest effect on CS. Cement content comes next, with a significant effect (mean SHAP = + 5.08). Superplasticizer dose and water content also have a significant effect, showing how they work together to make the combination more workable and densify the matrix. Fine aggregate and arc furnace slag have a moderate effect, whereas coarse aggregate and supplemental cementitious materials have a smaller effect in the dataset that was investigated. (a) Mean SHAP value, (b) Impact on model output, and (c) SHAP heatmap for the input features. Figure 9(b) shows the SHAP feature distribution plot, which shows how each variable affects model predictions in a certain direction. Longer curing times and more cement are mostly linked to positive SHAP values, which backs up the idea that they improve compressive strength. When the amount of superplasticizer increases, the effective water demand goes down. This makes the cementitious matrix denser and stronger. SCMs, on the other hand, have a weaker and more spread-out effect, which means that they don’t seem to be related to strength improvement in this model. Figure 9(c) shows the SHAP heatmap, which shows how changes in features across various samples affect each other. The curing length and cement content have the biggest positive and negative effects, which means that changes in these parameters have a big effect on the model’s outputs. The SHAP intensity shows that superplasticizer, water content, and fine aggregate have a moderate effect, whereas coarse aggregate, and SCMs have a little effect. In addition, the SHAP summary plot indicates that EAFS content exerts a measurable influence on CS estimation, with both positive and negative contributions depending on its dosage level. Higher EAFS proportions generally shift SHAP values toward enhanced strength prediction within an optimal range, whereas excessive replacement levels tend to reduce predicted strength. Figure 10 illustrates the potential relationship between SHAP values and various features. Arc furnace slag concentration up to 130 kg/m3 had a moderate influence on concrete strength as seen in Fig. 10(a). There is a moderate correlation between coarse aggregate and fine aggregate up to 930 kg/m3 to 1080 kg/m3 as seen in Fig. 10(b). In comparison, cement content and curing period exhibited a strong correlation with water content in strength. Fine aggregate having 700 kg/m3 to 880 kg/m3 has a moderate correlation with superplasticizers with a decreasing trend. From Fig. 10(f), there is no significant correlation between SCMs and superplasticizers. On the other hand, the superplasticizer content also has a strong correlation and impact with OPC, as shown in Fig. 10(g). The outcomes can be a valuable solution for the development of sustainable concrete in developing and developed countries, such as Najran city. Interaction plot for all input parameters used in this study. PDP Analysis Figure 11 demonstrates how each input variable influences arc furnace slag concrete’s compressive strength. Both a decision tree and a multi-layer perceptron were used to perform the partial dependency analysis. As expected, the concrete strength increases with a higher curing period, cement content, and superplasticizers as shown in Fig. 11(a), positively impacting the compressive strength. The compressive strength steadily increased from 33 to 48 MPa, a 45% rise, as the cement content was raised from 200 to 500 kg/m3. Additionally, the use of arc furnace slag led to an overall 12% improvement in compressive strength, increasing from 37.2 to 41.8 MPa at a proportion of 160 kg/m3 (Fig. 11(d)). The increase in compressive strength at moderate EAF slag contents can be attributed to the rough surface texture, angular shape, and higher hardness of slag particles, which improve interfacial bonding and internal friction within the concrete matrix. However, excessive EAF slag replacement may reduce strength because of higher porosity, and greater water absorption. Partial dependency of all parameters. In addition, the substitution of SCMs by an overall 1.4% decrease in compressive strength up to 90 kg/m3, then a slow rise of strength as depicted in Fig. 11(e). According to Fig. 11(f), the maximum compressive strength was obtained when the coarse aggregate density was in a range of 850 to 940 kg/m3. In contrast, there was a general tendency of decreasing compressive strength with the addition of fine aggregate and water content. Practical applicability of the developed models The developed ML models can be directly applied as a decision-support tool for the design and optimization of EAF slag-based concrete mixtures. Instead of relying solely on repeated laboratory trial batches, practitioners can input the proposed mixture proportions, including cement, EAF slag, SCMs, aggregates, water, superplasticizer dosage, and curing duration, into the trained model to rapidly estimate the expected CS. From a practical perspective, the developed framework can support engineers during both the preliminary design and optimization stages. A conceptual workflow for field implementation should involve four steps. First, the user defines the available material properties and target CS. Second, the selected mixture proportions are entered into the trained model. Third, the model predicts the expected CS and highlights the relative importance of each variable. Finally, the user can refine the mixture proportions until the predicted performance satisfies the project requirements. This approach can provide a practical pathway for integrating sustainable waste materials into concrete production while maintaining mechanical performance and reducing dependence on extensive experimental testing. Conclusions This study developed and evaluated multiple machine learning models to predict the compressive strength of EAF slag–based concrete using an extensive dataset compiled from published literature. The most important findings of this study are as follows: All five machine learning models showed excellent performance and great generalization abilities, making them extremely dependable for forecasting the CS of EAF slag concrete. The XGB model has the lowest average error values across all stages, with the highest prediction accuracy (R2 values of 0.951 for testing, and 0.945 for validation), indicating its robustness and reliability. Hybrid models (XGB-ADB and XGB-RF) outperformed single ADB and RF models, with R² values up to 0.947. These new models lowered prediction errors dramatically, demonstrating a gain in accuracy over previous methods. All five machine learning models showed excellent performance and great generalization abilities, making them extremely dependable for forecasting the CS of EAF slag concrete. The XGB model has the lowest average error values across all stages, with the highest prediction accuracy (R2 values of 0.951 for testing, and 0.945 for validation), indicating its robustness and reliability. Hybrid models (XGB-ADB and XGB-RF) outperformed single ADB and RF models, with R² values up to 0.947. These new models lowered prediction errors dramatically, demonstrating a gain in accuracy over previous methods. The hybrid models achieved predictive accuracy comparable to the standalone XGB model while clearly outperforming the individual ADB and RF algorithms. This improvement reflects enhanced stability, stronger generalization capability, and reduced susceptibility to overfitting, confirming the effectiveness of hybridization in strengthening model robustness. The hybrid models achieved predictive accuracy comparable to the standalone XGB model while clearly outperforming the individual ADB and RF algorithms. This improvement reflects enhanced stability, stronger generalization capability, and reduced susceptibility to overfitting, confirming the effectiveness of hybridization in strengthening model robustness. SHAP investigation demonstrates that arc furnace slag concentration up to 130 kg/m3 has a minor impact on strength, while fine aggregate (700–880 kg/m3) exhibits a decreasing trend with superplasticizer usage. SHAP also identifies a substantial link between superplasticizers and OPC, underscoring crucial feature interactions. SHAP investigation demonstrates that arc furnace slag concentration up to 130 kg/m3 has a minor impact on strength, while fine aggregate (700–880 kg/m3) exhibits a decreasing trend with superplasticizer usage. SHAP also identifies a substantial link between superplasticizers and OPC, underscoring crucial feature interactions. PDP analysis indicates that CS improves with curing time, cement content (45% rise from 33 to 48 MPa at 200–500 kg/m3), and superplasticizers. Arc furnace slag boosts strength by 12% at 160 kg/m3, but SCMs initially decrease strength by 1.4% up to 90 kg/3. Other parameters also showed noticeable correlation with the CS. PDP analysis indicates that CS improves with curing time, cement content (45% rise from 33 to 48 MPa at 200–500 kg/m3), and superplasticizers. Arc furnace slag boosts strength by 12% at 160 kg/m3, but SCMs initially decrease strength by 1.4% up to 90 kg/3. Other parameters also showed noticeable correlation with the CS. Limitations and future studies Although this study reveals some significant outcomes, several limitations should be acknowledged. The dataset was compiled from published experimental studies, which may contain variability in material properties, testing standards, curing conditions, and mix design procedures. Although efforts were made to ensure consistency, uncontrolled heterogeneity may influence model generalization beyond the considered parameter ranges. In addition, the models were trained on predefined input variables, and other potentially influential factors such as chemical composition variations of EAF slag, and long-term durability indicators were not incorporated. Future research should focus on expanding the dataset with real time experimental results. In addition, considering chemical composition based parameters, and validating the models through independent laboratory testing would greatly enhance the generalizability. The development of deployable graphical user interface (GUI) tools and real-time optimization frameworks for practical field applications would further enhance the applicability of the proposed modeling approach. Acknowledgements This work was funded by the Deanship of Graduate Studies and Scientific Research at Jouf University under grant No. (DGSSR-2025-02-01499). Author contributions Md. Alhaz Uddin : Conceptualization, Methodology, Formal analysis, Investigation, Writing- original draft, Writing - review & editing. Md. Habibur Rahman Sobuz : Conceptualization, Methodology, Validation, Supervision, Writing - original draft, Writing - review & editing. Md. Kawsarul Islam Kabbo : Methodology, Formal analysis, Data curation, Writing - original draft, Writing - review & editing. Ibrahim Y. Hakeem : Investigation, Formal analysis, Data curation, Writing - original draft, Writing - review & editing. Ratan Lal : Formal analysis, Data curation, Validation, Writing - original draft, Writing - review & editing . Ali Masria : Formal analysis, Data curation, Writing - original draft, Writing - review & editing. Sani Aliyu Abubakar : Formal analysis, Data curation, Writing - review & editing. Funding There are no specific funding for this research work. Data availability The datasets generated and/or analyzed during the current study are not publicly available due to maintain control over unsupervised usage that may lead to unintentional duplication of research efforts or reduced novelty in future studies but the dataset are available from the corresponding author Dr. Md. Habibur Rahman Sobuz (email: [habib@becm.kuet.ac.bd]) on reasonable request. Declarations Competing interests The authors declare no competing interests. Data access statement The processed dataset used in this study has been provided as supplementary material. Ethics statement This work is original and has not been published elsewhere, nor is it currently under consideration for publication elsewhere. Footnotes Publisher’s note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations. Contributor Information Md. Habibur Rahman Sobuz, Email: habib@becm.kuet.ac.bd. Ibrahim Y. Hakeem, Email: iyhakeem@nu.edu.sa Sani Aliyu Abubakar, Email: saliyu@kiu.ac.ug. References Associated Data Data Availability Statement The datasets generated and/or analyzed during the current study are not publicly available due to maintain control over unsupervised usage that may lead to unintentional duplication of research efforts or reduced novelty in future studies but the dataset are available from the corresponding author Dr. Md. Habibur Rahman Sobuz (email: [habib@becm.kuet.ac.bd]) on reasonable request.