Optimizing the mechanical performance of sustainable industrial waste modified concrete using supervised machine learning modeling and feature importance analysis

Abstract

The goal of the current study is to close a significant gap in the literature about the representation of concrete produced from industrial waste. In example, prior research experiences low levels of model explainability, limited generalizability, and limited data, especially because of their inability to capture nonlinear interactions among features. In order to overcome these limitations, authors apply seven supervised machine learning models to a large dataset of 711 observations. The two main new features of the suggested strategy include a thorough, systematic assessment of ensemble learners and SHAP-based model interpretability. With a maximum test accuracy of R² = 0.881, RMSE = 5.65 MPa, MAE = 4.17 MPa, and MAPE = 25.15%, the Gradient Boosting Regressor (GBR) outperformed other cutting-edge models such as CatBoost (R² = 0.879) and Histogram Gradient Boosting (R² = 0.878). Based on the findings from the SHAP analysis, it is clear that the machine learning model shows more sensitivity to the fine aggregate content than to cement and other materials, making it the most influential factor in the global model. Water content and curing age follow. In actual use, the suggested model and its intuitive Python-based graphical user interface (GUI) can quickly and affordably estimate compressive strength without requiring expensive and time-consuming laboratory testing, guaranteeing the successful incorporation of industrial waste products into the concrete mixture.

Supplementary Information

The online version contains supplementary material available at https://doi.org/10.1038/s41598-026-57625-9.

Introduction

Concrete is the most widely used construction material in the world, forming the literal foundation of our civilization^1,2. The rapid pace of progress in the fields of economics and urbanization has led to an explosion of infrastructure development and subsequent decommissioning of tired infrastructure^3–5. A key player is Portland Cement (PC), a key constituent of concrete. Cement production is a parasitic process that consumes high quantities of energy, leading to the emission of large amounts of carbon dioxide (CO₂), causing rapid degradation of our planet^6–9. Hence, it requires a move away from traditional manufacturing and towards alternative, eco-friendly sources of energy, which can ensure a sustainable path for the future^10–13. Failure to do so will result in irreversible damage to posterity, which necessitates an improvement in the current recycling strategies that can help to provide alternatives to certain materials used in manufacturing, resulting in the least footprint on the environment^14,15. Research has shown that the most feasible solution to solve the environmental concerns associated with the construction industry is the utilization of industrial by-products, which can act as supplementary cementitious materials (SCMs)^16–21. The most commonly used materials for concrete, namely silica fume (SF), fly ash (F Ash), and GGBFS, have been successfully used to partially replace the clinker in cement, which can reduce greenhouse gas emissions from traditional concrete by up to 80%^22–26. This solution addresses not just the natural resources used for concrete construction, but it also addresses the waste disposal problems of vast amounts of industrial waste sent to landfills^27–31.

The intricacy and amount of industrial refuse also pose serious logistical problems related to disposal activities, further exacerbated by the increase in environmental mandates and the reduced landfill availability^32–34. The construction industry may be able to improve sustainable resource activities by utilizing these materials as the aggregate within the concrete product^{29,31,35–46}. Even though industrial waste-based concrete offers significant advantages over alternative concrete products, the use of such products is limited by the intricacy of this product^47–50. This product is a heterogeneous composite material that exhibits mechanical properties as a result of an intricate relationship with cement, sand, gravel, and industrial by-products^51,52. The compressive strength of concrete, which is the most vital property to ensure the safety and durability of structures, depends on a variety of nonlinear factors such as the water-binder ratio, particle size distribution of SCMs, percentages of aggregates, and curing conditions^{26,28,29,43,53–58}.

The compressive strength of high-performance concrete materials has always been obtained by conducting experiments. This involves taking concrete materials and casting them in cubical or cylindrical forms, where they are allowed to cure for a specific period, weeks being the standard unit of time, which can usually be 7, 14, or 28 weeks. Afterwards, a compression test is conducted on these materials. Though this methodology is effective, labor-intensive experiments are expensive, time-consuming, and cumbersome to conduct⁵⁹. The logistics of conducting experiments are adversely affected by insufficient space to allow these materials to cure⁶⁰. Though empirical formulas have been put forward to effectively gauge the strength of these materials, these are ineffective when describing non-linear aspects involving concrete materials infused with waste materials. Thus, a high degree of urgency exists for developing advanced models to establish a nexus between input parameters and concrete materials infused with industrial waste.

The paradigm of Artificial Intelligence (AI) and Machine Learning (ML) has enabled a revolutionary approach in solving problems in civil engineering involving regression and classification^61–65. ML is different from natural approaches as it can learn from data and establish patterns and relationships between variables with a high level of accuracy^{25,38,66–68}. In addition, other techniques such as numerical simulations and modeling also garnered significant attention because of the low cost and rapid prediction of materials properties^69–72. Various ML approaches, such as ANN, SVM, and DT, have already been employed for the evaluation of the mechanical properties of different types of concrete materials. Gradient Boosting (GB), Categorical Boosting, Bagging Regressor, and Random Forest (RF) can also be employed for predicting the strength of concrete with recycled aggregate concrete, where the main influencing factor is identified as cement content^73,74. Other approaches, such as Multi-Expression Programming (MEP) and Gene Expression Programming (GEP), have demonstrated the merit of these approaches over conventional regression methods for alkali-activated concrete. It is worth mentioning that though these ML approaches contribute to cost-effective solutions, the latest trend of research has focused on ensemble learning to ensure better predictive efficiency of the model. Ensemble learning approaches combine several weak learners to develop a strong learner, which is able to perform better during generalization. For instance, the application of Gradient Boosting (GBoost), Light Gradient Boosting Machine (LightGBM), and CatBoost has demonstrated their merit against conventional approaches to the prediction of concrete strength. In fact, the development of stacked ML approaches, which combine the efficiency of GBoost, RF, and Extra Trees Regressor (ETR), has demonstrated their merit to the extent of achieving R² values as high as 0.96 for concrete mixes involving concrete. In addition, the combination of DNNs with gradient boosting methods represents a comprehensive tool for the efficient prediction of tabular as well as complex non-linear data^{38,39,56,75,76}. From simple to complex, this process has recently been even more streamlined with the development of Automated Machine Learning frameworks like AutoGluon through the automation of an expert-driven task, hyperparameter tuning. These hybrid approaches improved the R² score from 5 to 10% and considerably reduced MAE, hence giving a strong basis for construction material design^56,77–79. A major bottleneck towards the broad acceptance of AI in the construction sector is related to the “black box” nature of complex models like DNNs. Engineers have to be absolutely clear as to how a model arrives at a particular prediction to ensure structural safety. Recently, researchers started using interpretability tools like SHapley Additive exPlanations (SHAP)⁸⁰.

SHAP values provide a unified measure of feature importance, revealing the global and local contributions of features such as SF, F Ash content, water binder, and curing ages, among others. SHAP, for instance, might give insights into the contributions made towards the models’ ability to make the predicted outcomes, e.g., 20%-25% is contributed by the fly content in the decision-making involved in the modeling of the sustainable mixes of concrete materials^42,81–83. Although progress has been made in addressing the problem using ML, there are still critical gaps in the existing literature regarding the modeling of concrete materials, regardless of the material source, due to small, limited data sets, which lead to overfitting. In addition, there is limited parametric analysis using robust tools such as SHAP and PDP in the analysis of the nonlinear sensitivity associated with industrial waste materials, modifying the concrete mixes. The majority of existing works consider network-based approaches, i.e., MLPNN and SVM, or tree-based approaches, i.e., RF and DT, with few works attempting to offer a comprehensive comparison or a hybrid combination of these quite different paradigms.

Using interpretability models like SHAP, which have advanced beyond the “black-box” approach to prediction, this research has made significant progress in predicting tangible qualities. For instance, Zhang, Pei-Fu, et al.⁸⁴ were able to clarify the non-linear impacts of SCMs in high-performance and ultra-high performance concrete using hybrid ML-SHAP models. Similarly, Wang, Hongwei, et al.³⁸ employed SHAP to examine how the water-to-cement ratio and FA content affected the compressive strength of the concrete. However, there is a significant research gap because many of the studies that are now available employ small data sets and do not take into account the complex relationships between variables in concrete mixed with various types of industrial wastes.

This research aims to fill these gaps by implementing a robust set of traditional and hybrid ML techniques, such as GBoost, Categorical Boosting, Bagging Regressor, Random Forest, Hist Gradient Boosting, Extra Trees Regressor, K-Nearest Neighbor models, on a comprehensive dataset of 711 data points collected from previous published Scopus indexed journals. The highest standards were used when creating this database to ensure that the testing protocols and the laboratory conditions in which the tests were performed were consistent. The study considers nine significant input variables, which represent a wide range of mix components and environmental factors. This study will ensure reliability in the results by performing cross-validation and considering multiple statistical metrics during both the training and testing phases.

While the SHAP approach and ensemble modeling techniques have been used in recent publications to predict concrete strength, this study focuses on analyzing the non-linear mechanical properties of sustainable concrete affected by various types of industrial by-products during its early-age curing period. Specifically, the study uses a unique set of 711 samples to demonstrate that, according to sensitivity tests conducted using the SHAP approach, the primary strength determinant for such waste-based concrete mixes is the fine aggregate content rather than the cement itself.

Research significance

This study is of immense significance in the context of the burning need to achieve sustainable construction through the optimization of concrete modified by industrial waste via a supervised machine learning algorithm. Though the existing body of knowledge had hitherto been restricted to the use of conventional mix or small data sets to make predictions, the work presented in this study utilized a large data of 711 samples, well above the benchmark required to make a quality model, to make accurate high-precision predictions of the compressive strength. Unlike the existing body of knowledge, where the predictive models were considered “black boxes,” this work included SHAP (SHapley Additive exPlanations) to understand the effects of variables such as the amount of fine aggregate and the amount of water, along with other factors, in determining the strength. The study shows that when multiple ensemble learners, such as Gradient Boosting and Random Forest, are compared, they offer higher statistical reliability and accuracy than does existing literature. Moreover, such findings expose critical interactions of materials, such as early curing and the influence of different amounts of industrial waste, not well covered in previous works. This research eventually provides a robust, data-driven framework that enables engineers to optimize the use of industrial by-products without compromising structural integrity, hence moving toward an environmentally friendly and economical approach in modern concrete technology. Table 1 illustrates the ML models used in the previous literature.

Table 1.

ML models adopted in the previous literature.

References	Type of Concrete	Model	Dataset	Year
85	HPC	GBRT	1030	2023
86	HPC	Ensemble and non- ensemble ML approach	471	2021
87	HSC	RF	225	2022
88	HPC	XGB, RF, and SVR	1030	2022
89	HPC	GEP, DT, and Bagging	270	2021
90	HPC	SVM	1761	2018
91	HPC	ANN	1133	2018
92	Pozzolanic concrete	GEP and ANN	310	2022
93	OPC and HPC	ANN and ANFIS	2817	2020
94	HSC	GEP	357	2020
95	UHPC	XGB, LGB, Hybrid XGB-LGB	626	2024
96	UHPC	AB, RF, GB	810	2024
97	RHA-based concrete	LGB, XGB, RF, Hybrid XGB-LGB, Hybrid XGB-RF	1404	2024
98	RHA concrete	CatBoost, GBM, CNN, GRU	1212	2024
81	MPC composites	CNN-LSTM, CNN-GRU, DTR-RFR, GBR-RFR	3108	2023

Methodology

The construction flowchart for the predictive model involves several key steps, shown in Fig. 1. Initially, a dataset of sustainable industrial waste-modified concrete is created. This is followed by the selection of essential features while filtering out unnecessary data. The dataset is then divided into training and testing sets to assess its performance. The fourth stage involves training the different models using the training set, with performance evaluated through a five-fold cross-validation process aimed at optimizing hyperparameters. Once the model is tuned, it is tested on new, unseen data to ensure reliability, and if the accuracy is satisfactory, the model is finalized. The final step utilizes SHAP (SHapley Additive exPlanations) to analyze the influence of each feature on model predictions, identifying key variables and their trends in predicting concrete compressive strength (CS).

Fig. 1.

Methodology flowchart of this study.

Data collection and description

The quantity of data points in a dataset has a major impact on a machine learning model’s dependability. The Compressive Strength (CS) of sustainable industrial waste modified concrete was predicted in this work using a dataset with nine input factors. The input parameter for this study, industrial waste, is regarded as a composite variable that refers to materials with almost equivalent impacts on the mechanical and structural development of the concrete matrix and extremely similar chemical properties. In this regard, industrial waste types such as waste glass powder, waste marble powder, ceramic waste powder, limestone powder, ground granulated blast furnace slag and metakaolin are among the 711 samples acquired for the tests. The materials are organized together to enhance the effectiveness of machine learning algorithms due to their shared additional binding agents or micro-fillers. A rigorous gathering information process was carefully adhered to in data extraction in order to ensure high reproducibility and minimize variations resulting from various testing conditions. Specifically, only those data points that met the following criteria were considered: (i) the concrete specimens were tested under very similar curing and testing conditions; (ii) the compressive strength was evaluated by means of standard tests (e.g., ASTM tests); and (iii) all information regarding mix designs, including the values of all nine input parameters, was available. The meta-data thus created can be employed as a reliable data set for supervised machine learning due to the above-mentioned rigorous data cleaning and filtering procedure, which minimizes the potential for systematic errors. The train-test split method (90% training data and 10% testing data) was used to test the model’s performance metrics for this particular study. It should be mentioned that although the GridSearchCV methodology used a 5-fold cross-validation strategy to optimize the model hyperparameters, the mean standard deviation values were not calculated using this method. Statistical uncertainty analysis is therefore outside the current scope of the study. The Standard Scaler approach, which centers the data around the mean and scales features to unit variance, was used to prepare the data. This procedure improves model performance by guaranteeing constant feature scaling, which is essential for techniques that are sensitive to feature scaling, such as gradient boosting regression. Statistical measures of the experimental results are presented in Table 2; Fig. 2, which include observations on material qualities and mix proportions. CS, cement, water, fine aggregate, coarse aggregate (CA), silica fume (SF), fly ash (F Ash), industrial waste superplasticizer, curing age are important variables in the dataset. The mean, median, mode, standard deviation, maximum, minimum, and skewness are among the statistical indicators displayed for each parameter.

Table 2.

Statistical overview of concrete composition and performance.

Parameters	Max	Min	Mean	Std.	COV (%)	Skewness	Kurtosis
Cement (kg/m3)	604	30	331.73	106.89	32.22	-0.792	0.671
Fine Aggregate (kg/m3)	1032.8	0	502.95	368.60	73.29	-0.094	-1.495
Coarse Aggregate (kg/m3)	2649	305	1145.43	439.79	38.40	0.249	-0.773
Silica Fume (kg/m3)	150	0	14.91	34.98	234.69	2.304	3.855
Fly Ash (kg/m3)	247	0	28.46	53.91	189.43	1.737	1.751
Water (kg/m3)	262	78	165.31	39.24	23.74	0.519	-0.190
Industrial Waste (kg/m3)	509	0	60.93	70.63	115.92	1.410	2.558
Superplasticizer (kg/m3)	24	0	2.13	2.62	123.45	2.744	18.608
Curing Age (days)	365	3	31.53	33.92	107.59	6.147	52.923
Compressive Strength (MPa)	77	1.5	26.51	16.96	63.97	0.490	-0.499

Fig. 2.

Histograms of characteristics of the dataset.

Table 2 shows that in the summary of the concrete mix, the dominant component is coarse aggregate with a mean value of 1145.43 kg/m³, the second in line being fine aggregate with a mean value of 502.95 kg/m³, and the third being cement of 331.73 kg/m³. The final compressive strength has a big variation, with a minimum of 1.5 MPa, a maximum of 77 MPa, with an average of 26.51 MPa. While considering COV, or the relative spread, special additives showed the highest amount of variation. SF of 234.69%, F Ash of 189.43%, and superplasticizer of 123.45%, indicating that the mixes were focusing much on varying these constituents. Water usage was very consistent at 23.74%, as was cement usage of 32.22%. Another major implication from the shape of this data series is the skewness of 6.147 and the minimum value of 3 days; hence, the majority of the mixes were tested at a very early age of curing. Concluding the analysis, the experiments were greatly designed to study the impact of small, variable amounts of additives and the effects of early curing on the final strength of the concrete.

Figure 2 presents a hexagonal binning joint plot of the relation between variables (kg/m³) on the x-axis and compressive strength (MPa) on the y-axis, with individual distributions shown. The green color of the hexagons represents the density of data, with darker hexagons indicating high concentrations of the points. Figure 2(a) also shows that, with an increase in the cement content from 300 kg/m³ to 450 kg/m³, the compressive strength increases to a maximum between 20 MPa and 50 MPa, where the points are closely packed. The compressive strength distribution is often slightly biased to the right (0.490), however the majority of the data falls between 15 and 40 MPa, which is consistent with the average strength value of 26.51 MPa. Figure 2(b) shows many mixtures with either very low or very high fine aggregate content, ranging from 0 to 1000 kg/m³. Importantly, the dark clusters confirm that very low fine aggregate content, near 0 kg/m³, correlates strongly with very low compressive strength, near 5 MPa. Unlike for cement, however, increasing the fine aggregate content does not ensure high strength; the points remain scattered throughout the high-strength range. Figure 2(c) shows that concrete mixes containing low coarse aggregate (500–800 kg/m³) tend to achieve high compressive strength. Conversely, mixes that utilize high coarse aggregate (1400–2000 kg/m³) typically yield low compressive strength. The most important observation shown in Fig. 2(d) is that mixes with minimal or no silica fume exhibit the highest compressive strengths, ranging between 0 to over 70 MPa in mixes containing 0 to 5 kg/m³ of silica fume. Fly ash also shows the same trend as silica fume, which is shown in Fig. 2(e). Figure 2(f) illustrates the inverse relationship between water content and the highest compressive strength, since the highest strengths are only observed with the lower water content of approximately 100 to 150 kg/m³, which are 60 MPa and up to 75 MPa. Like silica fume and fly ash, industrial waste and superplasticizer show similar trends, which are depicted in Figs. 2(g) and (h). At the end, Fig. 2(i) exhibits a few distinct clusters in curing age, with a very dominant peak near 28 days and less dense clusters around 3 days, 7 days, 56 days, and 90 days. The data set’s statistical distribution shows notable variations in mix proportions, especially when it comes to additional cementitious elements like fly ash and silica fume. The goal of this concentration is to improve these materials’ qualities. This distribution also shows that low w/b ratio values and cement percentage result in high compressive strengths; however, other parameters like as curing age, w/b ratio changes, and cement proportion may also have an impact on this relationship. More significantly, the strength curve clearly shows how age affects the material mix’s performance, reaching its maximum at 28 days. However, the complex relationship of several input elements largely determines the material’s mechanical behavior.

Feature description and correlation

Figure 3 shows a correlation matrix with a pairwise linear relationship between several variables within the concrete dataset. Coefficients range from − 1.0, with a dark blue color for strong negative correlation, to + 1.0, with a dark brown color for strong positive correlation. The intensity and shape of color demonstrate the strength and direction of the correlation. The highest correlations are between coarse aggregate and fine aggregate, with r = -0.88, which means that both have a very strong inverse relationship; that is, when coarse aggregate is high, fine aggregate is low, and vice versa. Another strong negative correlation was between coarse aggregate and industrial waste, r = -0.69. Compressive Strength, which is probably the dependent variable in this dataset, has its strongest positive correlation with fly ash at r = 0.65 and superplasticizer at r = 0.46, suggesting both are critical for high strength. On the contrary, compressive strength shares its most intense negative correlation with coarse aggregate at r = -0.66 and silica fume at r = -0.17, with increased amounts of the two components tending to be associated with low compressive strength.

Fig. 3.

Pearson heatmap of different variables.

ML framework

Gradient boosting regressor

An ensemble regression technique is the Gradient Boosting Regressor (GBR). It builds on a basic model by iterating over a randomly selected training set; faster regressions result from smaller training data fractions, its framework shown in Fig. 4. The number of trees and the learning rate are important tuning factors. The core idea of GBR is “boosting,” which improves predictions by using additive training to improve a “weak” learner. By employing an objective function, GBR efficiently reduces overfitting and maximizes computing resources, enhancing performance and lowering error. Figure 5 depicts its development conceptual diagram.

Fig. 4.

Framework of gradient boosting regressor.

Fig. 5.

Conceptual diagram of Gradient Boosting Regressor model development.

The fundamental objective of GBR is to identify a function f(x) that minimizes a given loss function L(y, f(x)) over the training data as T = (x_1,y₁),(x₂,y₂),(x₃,y₃),….,(x_n,y_n), where y_i represents the output sample and x_i the input features. In many cases, the square error function is employed as the loss, defined as:

1

The learning process begins with an initial approximation f_o(x), selected to minimize the loss function over all samples:

2

At each iteration m, the algorithm computes the residuals for each sample, which are equivalent to the negative gradients of the loss function with respect to the current model predictions:

3

In the context of the squared error loss, these residuals simplify to the difference between the actual target and the current prediction:

4

Categorical boosting

A machine learning model called CatBoost makes use of gradient boosting decision trees. It achieves excellent accuracy, generalizability, and lowers the danger of overfitting while successfully addressing categorical features, gradient bias, and prediction shifts.

The experimental dataset can be represented as D = (x_k, y_k)_k=1,2,….n, having assumed x_k = (x¹_k, ⋯, x^m_k ) as a random vector of the m features and y_k ∈ R as a target. The supervised training phase of the model is aimed at training a function F: R^m → R that minimizes the expected loss L defined as L (F): = L (y, F(x)), with L a smooth loss function. The gradient boosting procedure improves the prediction of y by means of a function F^t+1:

5

α is defined as step size and h^t+1 as base predictor, selected from a set of functions H in order to minimize the loss function.

6

Either by means of the Taylor approximation or by using the negative gradient step approach:

7

having defined:

Bagging regressor

An ensemble approach called bootstrap aggregating, often known as bagging, improves predictions by reducing variance through the use of several classification and regression algorithms. A randomly dispersed training set made by sampling instances from the original set with replacement is used to generate individual models. This might lead to events in the training set being repeated or left out. The average of the predictions from several regression models that have been built yields the final forecast, shown in Fig. 6.

Fig. 6.

Schematic representation of Bagging Regression using multiple decision trees.

The regression setting in this study had pairs (X_i, Y_i) (i = 1,2,.…, n), where X_i ∈R^d denoted the d-dimensional predictor variable and the response Y_i ∈ R (regression). The target function of interest was usually E [Y|X = x]. The function estimator, which was obtained by a given base procedure, was

8

where h_n(.) defines the estimator as a function of the data.

Random Forest

A popular approach for classification and regression, Random Forest (RF) builds several classification or regression trees (CART) using bootstrap and aggregation techniques. Regression uses samples of the input variables to fit the output variables. To determine the Sum of Square Error (SSE) between expected and actual values, data is divided into points, and each node’s minimal SSE is chosen. By permuting input values and analyzing the effect on prediction accuracy using out-of-bag samples, variable relevance is determined, shown in Fig. 7.

Fig. 7.

Workflow of Random Forest Regression with 5-fold cross-validation.

Histogram gradient boosting

A new machine learning technique is called histogram-based gradient boosting. In contrast to previous methods, it creates feature histograms during training by bucketing continuous feature values into discrete bins. This approach lowers the model’s memory use and expedites the training phase and its process illustrates in Fig. 8. The scikit-learn library’s algorithm serves as the foundation for the one created here.

Fig. 8.

Process of histogram-based decision tree construction.

Extra trees regressor

ETR is a sophisticated version of the Random Forest (RF) architecture that uses many unpruned regression trees created using a conventional top-down method. The effectiveness of ensemble learning is greatly increased by this paradigm. Unlike the RF model, ETR generates trees using a deterministic approach in which each feature’s split point is selected at random. The workflow diagram of ETR is shown in Fig. 9.

9

Fig. 9.

Workflow of an Extra Trees Regressor ensemble.

Where, Y_ETR denotes the output of ETR model, P represents the tree number in ensemble, T_i is the i-th tree in the ensemble, and X is the input feature vector. By combining mathematical components, the ETR algorithm offers a unique regression technique that strikes a compromise between accuracy and unpredictability.

K-Nearest neighbor

K-Nearest Neighbor is a highly reliable and versatile non-parametric machine learning method that has been used in this study for regression purposes. The use of K-NN regressor, as a regression algorithm, helps to estimate the value of the output variable, which is compressive strength of concrete modified with industrial wastes, by measuring the similarity between feature vectors in the training dataset. Unlike using pre-specified mathematical functions or linear relationships, K-NN finds the ‘k’ number of nearest neighbors using the following Euclidean distance formula:

10

The non-parametric K-Nearest Neighbor (KNN) technique is used in statistics for both regression and classification, shown in Fig. 10. Without assuming anything about unknown functions, it uses data points to estimate functions. In contrast to regression, which averages the k closest values, KNN classifies by voting among the k-nearest neighbors, usually producing a class output. Although it might enhance bias, raising the value of k reduces variance. In order to forecast results based on closeness to training data, the algorithm learns from stored data.

11

Fig. 10.

Classification process of the K-Nearest Neighbors (KNN) algorithm.

Hyperparameter tunning with GridSearchCV

In this study, a dataset was used to train many models in order to determine which model performed best for different machine learning (ML) tasks. Even pointed out, it’s still unclear if the chosen model is the best way to solve the issue at hand, underscoring the necessity of continuous development. The hyperparameters are a crucial component of model effectiveness since choosing the right values for them can greatly improve model performance. Grid search cross-validation (GridSearchCV) was used to determine the optimal model for each machine learning technique. This method provides a more precise evaluation of generalization performance by automatically fitting a new model using the parameters that produce better cross-validation performance over the whole training dataset. The prediction errors are then estimated using cross-validation as follows:

12

here k is the number of subsets, n is the size of the dataset, T is the loss function, and f^−k(i) is the fitted function. Figure 11 illustrates the GridSearchCV approach used in this work for model training and hyperparameter selection.

Fig. 11.

Hyperparameter tuning with k-fold cross-validation (GridSearchCV)⁹⁹.

Model development

As illustrated in Fig. 12, the establishment of the GBR, model consists of three key steps:

Fig. 12.

Optimization process for tuning model hyperparameters and model evaluation¹⁰⁰.

Step 1: Defining the Initial Hyperparameters.

Learning_rate, n_estimators, subsample, criteria, min_samples_split, min_samples_leaf, max_depth, min_weight_fraction_leaf, min_impurity_decrease, max_features, loss, and alpha were the hyperparameters employed in this study. The sample domain for the optimization process is defined by the starting ranges and categories of these hyperparameters. For example, n_estimators are sampled between 50 and 500, but learning_rate is sampled from a logarithmic scale between 1 × 10^− 4 and 1.0. This enables a large configuration space, improving the framework’s quick convergence to ideal parameters and enabling effective sampling.

Step 2: Training.

Instead of employing the initial 80:20 ratio, a repeated 5-fold cross-validation (CV) technique was used to increase the statistical significance of the model assessment. It is an iterative process that splits the data into five sets, trains the model on four partitions, then assesses it on the fifth partition. The cycle is repeated several times by mixing the data at random. By allowing each point in the database to be tested and trained several times on the same sample, this procedure greatly lowers the degree of bias in the outcomes. All rounds of the cross-validation procedure are averaged to determine the values of the metrics used in model assessment.

Step 3: Hyperparameter Optimization.

The tuning process for all machine learning models in this research project was performed using grid search cross-validation method known as GridSearchCV. Grid search cross-validation method exhaustively searched for the optimal parameter combination within a certain portion of the parameter space by performing five folds cross validation process. The ideal configuration is determined by combining the hyperparameters that produce the best average cross-validation performance.

The final model is trained on the whole training set using the optimal hyperparameters when the optimization procedure is finished. Root Mean Square Error (RMSE), Mean Absolute Error (MAE), and the coefficient of determination (R²) are then used to assess the model’s prediction performance on the test set. The outcomes validate the advantages of employing hyperparameter optimization by showcasing the model’s efficacy and capacity for generalization.

Feature importance (Shapley method)

Many machine learning algorithms are highly predictive, yet they frequently fail to provide an explanation for their predictions. In order to improve the explainability of compressive strength (CS) predictions for sustainable industrial waste modified concrete, this work presents SHapley Additive exPlanations (SHAP). SHAP calculates how each input variable—including mix and material design characteristics—contributes to the model’s forecasts. It calculates SHAP values, which show how a feature’s knowledge affects the model’s output for particular predictions. SHAP uses Shapley values derived from game theory to combine local interpretability with efficient credit allocation. Both global and local interpretability are produced by this method; the former shows how feature distributions impact overall model performance, while the latter makes clear how particular characteristics contribute to particular predictions. SHAP’s primary objective is to connect model outputs to specific attributes by evaluating their influence¹⁰⁰. SHAP facilitates systematic interpretation of the model’s performance in forecasting the CS of sustainable industrial waste modified concrete by modeling characteristics as a linear system and representing each occurrence as a vector of feature values. SHAP specifies the explanation model as,

13

where M is the maximum number of features, g (·) is the explanatory model, and ϕ_i ∈ R is the contribution of feature i to the prediction. Along with the efficiency, symmetry, and additivity that are intrinsic to Shapley values, SHAP also maintains three important features: consistency, missingness, and local accuracy. The degree to which the outputs of the explanatory model match those of the original machine learning model is measured by local accuracy. This can be expressed as,

14

Where

Missingness guarantees that features that are absent from a model contribute zero, resulting in Shapley values of zero for such features. According to consistency, a feature’s Shapley value will either rise or stay the same if a model is modified in a way that causes its marginal contribution to either increase or keep the same. Furthermore, tree-based SHAP has advantages like lower computing costs without sacrificing prediction accuracy. In order to improve the interpretability of the model’s predictions, this work calculates Shapley values using tree-based SHAP.

Performance evaluation metrics

To ensure that the performance of our predictive tool is assessed in a fair manner relative to other models that might be used on this scenario, a number of different metrics can be applied: Mean Absolute Error (MAE), Mean Absolute Percentage Error (MAPE), Root Mean Squared Error (RMSE), and the Coefficient of Determination (R²). While previous models tend to have assessed their output in terms of perhaps MAE and MAPE at best, we will examine a number of different factors in this case in order to provide a full assessment. MAE provides a mean level of error and does not take outliers into account as heavily due to the varying properties of concretes. While MAE is used for mean purposes in this case due to practicality. RMSE calculates this result to a standardized level in regards to its original units for better understanding and interpretation, while R² provides a representation of how much of the variances in the actual result within the data can be accounted for by this.

15

16

17

Results and discussions

Model performance evaluation

Table 3 shows the performance metrics of seven trained and tested ML models-GBR, CatB, BR, RF, HGB, ETR, and KNN-on the compressive strength prediction dataset based on four metrics: R², RMSE, MAE, and MAPE. Even while MAPE is a commonly used measure for relative error, it tends to be more dependent on minor absolute values such as those seen in the concrete sample with low compressive strength, where modest changes would result in a comparatively greater percentage error. MAE and RMSE were used as the two main measures for absolute error to ensure robustness with respect to any biasing impact. RMSE highlights severe mistakes, whereas MAE provides a linear measurement of average error regardless of the response variable’s scale.

Table 3.

Outcomes of seven machine learning models.

Model	R2 Train	R2 Test	RMSE Train	RMSE Test	MAE Train	MAE Test	MAPE Train	MAPE Test
GBR	0.903	0.881	5.28	5.65	3.77	4.17	21.30	25.15
CatB	0.926	0.879	4.64	5.70	3.02	3.71	16.17	21.07
BR	0.922	0.856	4.77	6.21	3.01	4.08	17.32	22.05
RF	0.926	0.862	4.64	6.09	3.01	4.03	17.05	22.44
HGB	0.912	0.878	5.04	5.72	3.42	3.99	18.35	21.59
ETR	0.932	0.839	4.45	6.57	2.58	4.19	13.79	22.71
KNN	0.874	0.844	6.05	6.47	4.20	4.61	26.11	31.90

Accordingly, there was fair predictive power, as the R² values for all models were high, ranging from 0.874 to 0.932 for the training dataset and from 0.839 to 0.881 for the test dataset. Among the top models, the best training R² =0.926 and the lowest training MAE = 3.02 and RMSE = 4.64 were derived by the CatB model; however, the GBR marginally outperformed it on the test set by obtaining the highest test R² = 0.881 and the lowest test RMSE = 5.65. Meanwhile, the best fit on the training data was derived by the ETR model in terms of the lowest MAE = 2.58 and RMSE = 4.45; however, this also indicated its largest drop in R² between the training (0.932) and testing phases (0.839); therefore, it may have also demonstrated potential overfitting. In turn, the KNN model performed the worst across all metrics, such as the lowest test R² =0.844, with the highest test RMSE = 6.47 and MAPE (31.90). On the whole, the CatB and GBR models provided the most robust and generalized performance, as evidenced by their high test R² scores of 0.879 and 0.881, respectively, and relatively low test MAE of 3.71 and 4.17, respectively. GBR performs well because it can adjust to the size and dimensions of the data set. GBR offers a good balance between model bias and variance in lower dimensional data sets with fewer physical factors, in contrast to advanced models like CatBoost, XGBoost, and LightGBM that are built for higher dimensional data sets and greater sample numbers. Unlike its boosted variants, GBR’s design simplicity allowed it to learn and represent extremely non-linear interactions without requiring substantial hyper-parameter adjustment. Without overfitting the 711 experimental points, GBR got the highest testing R² of 0.881. Figure 13 shows the comparison of the performance of these metrics of different ML models. An overview of several research is shown in Table 3, arranged according to experimental design, method, data size, and performance measures like R², RMSE, MAE, MAPE, MSE, and CE where appropriate. The analyses only cover research using the entire data set of 711 samples in order to provide an objective comparison. Average performance and cross-validation performance are the two main categories into which the results are separated.

Fig. 13.

Comparison of performance metrics of different ML models: (a) R², (b) RMSE, (c) MAE, (d) MAPE.

Table 4 provides the maximum, minimum, and average error of seven machine learning models on both training and testing datasets. Concerning the average error, which corresponds to MAE, the best performance on training data was given by the ETR model (2.5794), but it performed less optimally on the test set (4.1913); meanwhile, the KNN model gave the largest average errors in both the training set (4.1964) and the test set (4.6106). As for the maximum error, representing the largest single deviation in prediction, the ETR model recorded the highest value on the test set, 34.5000, and the lowest on the train set, 16.9750. This implies a high possibility that this model will generate a very large outlier while predicting data that it has never seen. On the other hand, the HGB model maintained the lowest test max error at 22.4249 among all models, indicating consistency even in its worst-case predictions. The minimum error of all is around zero in the set of training data, and the exact error of 0.0000 was obtained by BR and ETR, while in the test set, the minimum value of the error was realized by the HGB model as 0.0039, meaning that the best predictions of this model were extremely accurate.

Table 4.

Descriptive statistics of variables used in modelling.

Model	Max error (Train)	Max error (Test)	Min error (Train)	Min error (Test)	Avg error (Train)	Avg error (Test)
GBR	20.0256	23.3166	0.0094	0.0065	3.7710	4.1670
CatB	17.8108	28.6484	0.0063	0.1374	3.0216	3.7135
BR	20.0284	31.9550	0.0000	0.0133	3.0148	4.0800
RF	18.8749	29.0811	0.0047	0.0434	3.0104	4.0304
HGB	18.4906	22.4249	0.0127	0.0039	3.4186	3.9899
ETR	16.9750	34.5000	0.0000	0.0000	2.5794	4.1913
KNN	31.4884	24.2300	0.0000	0.0380	4.1964	4.6106

Figure 14 shows a combined scatter plot of observed values versus predicted values for both the training and testing datasets, with a plot of the residuals. To assess the models’ dependability, the outliers found in the residuals plot were quantitatively evaluated based on the prediction errors. For the training data in the range from 0 to 580, the blue dotted predicted values generally follow the trend of the black observed values, although there is clear scatter, with the predictions clumping around the 5-50 MPa range. For the testing data in the range from 580 to 720, the orange-colored dot, predicted values similarly show a reasonable agreement with the observed values, with the test data appearing to contain a larger proportion of high-strength observations (50-70 MPa) where the predictions are generally lower than the observed values. Residuals cluster close to the zero line as most predictions lie close to actual values. A few outlier residuals reach as high as 150 or more, although the majority fall within a range from − 15 to 15, implying that generally the GBR model makes small, balanced errors but will have a few very big misses in its predictions, which are mostly positive. CatB and BR show likely the same trend as the GBR. But the predictions of CatB appear to align slightly better with the higher test observations (above 50 MPa) compared to the GBR model. Besides being similar to GBR and CatB, BR and RF show significant positive outlier residuals. In contrast, the overall testing set prediction of the HGB model seems to catch the higher observations above 60 MPa slightly better than either the BR or RF models. Besides, it has fewer extreme positive outliers up to 125, compared with CatB and ETR, which is consistent with its reported low test max error. The ETR, while achieving the minimum average training error in Table 5, has the largest residual spread on the test set, including the highest positive outlier. Meanwhile, the KNN model seems to have the least variance in both its train and test predictions, leading to the highest concentration of large residuals between − 20 and 20. This agrees with the highest test RMSE and MAE reported for the KNN model. Overall, all the models have an excellent fit for the bulk but share a significant weakness in the accurate prediction of the extremely high values of compressive strength.

Fig. 14.

Combined actual predicted residuals of different models: (a) GBR; (b) CatB; (c) BR; (d) RF; (e) HGB; (f) ETR; (g) KNN.

Table 5.

Comparison with previous studies.

Reference	Concrete types	Input features	Model	Dataset	R 2	RMSE	MAE	MAPE	MSE	CE
101	HPC	Cement, fly ash, blast-furnace slag, water, superplasticizer, age, and coarse and fine aggregate	ANN	1030	0.909	5.030	-	0.109	-	-
			SVM		0.886	5.619	-	0.128	-	-
			Multiple regression		0.611	10.429	-	0.317	-	-
			Multiple Additive Regression Tree		0.911	4.949	-	0.139	-	-
			Bagging Regression Tree		0.890	5.572	-	0.142	-	-
102	UHPC	Silica fume, fly ash, marble powder, cement, aggregates, water, superplasticizers, and curing age	GB	249	0.960	2.210	1.750	-	4.910	0.960
			XGB		0.954	2.650	1.990	-	7.010	0.954
			LGBM		0.952	2.530	1.910	-	6.380	0.952
			CatBoost		0.950	2.660	2.070	-	7.080	0.950
103	HPC	Cement, Blast-furnace slag, Fly ash, Water, Superplasticizer, Coarse aggregate, Fine aggregate, Age.	Decision Tree	1030	0.818	-	-	-	-	-
			Bagging DT		0.879	-	-	-	-	-
			Gradient boosted DT		0.889	-	-	-	-	-
			Random Sub spaced DT		0.868	-	-	-	-	-
104	Concrete	Coarse/fine aggregates, cement, water, additive, age	CatBoost	1030	0.952	4.856	3.205	11.39	-	-
105	HPC	Cement, Blast-furnace slag, Fly ash, Water, Super-plasticizer, Coarse aggregate, Fine aggregate, Age	SVM	1030	-	5.59	3.75	0.12	-	-
			Stacked CART + SVM + LR		-	5.08	3.52	0.12	-	-
106	HPC	Cement, water, sand, gravels, superplasticizer, concrete age, fly ash and granulated blast furnace slag	Modified Random Forest Ensemble	1030	0.923	4.600	3.230	-	-	-
107	FRC	-	CNN	1030	0.970	3.98	2.68	-	-	-
			CNN-AP		0.970	4.09	2.92	-	-	-
			CNN-MP		0.960	4.18	2.89	-	-	-
108	Quaternary blend concrete	Cement, GGBS, Fly Ash, water, superplasticizer, coarse aggregate, and fine aggregate, age.	XGB	1030	0.947	3.764	2.437	0.085	-	-
			LGBM		0.947	3.764	2.439	0.086	-	-
			CBT		0.951	0.951	2.246	0.077	-	-
			GBR		0.944	0.944	2.425	0.084	-	-
109	HPC	Cement, Blast-furnace slag, Fly ash, Water, Super-plasticizer, Coarse aggregate, Fine aggregate, Age	ANN	1030	0.909	5.57	4.18	-	-	-
			Bagged ANN		0.928	4.87	3.60	-	-	-
			Gradient boosted ANN		0.927	5.24	4.09	-	-	-
			Wavelength bagged boosted ANN		0.940	4.54	3.30	-	-	-
			Wavelength gradient boosted ANN		0.953	5.75	4.83	-	-	-
110	Conventional concrete	-	ANN	1030	0.890	5.900	-	-	-	-
			SVM		0.830	7.500	-	-	-	-
			Random Forest		0.900	5.600	-	-	-	-
111	Foamed Concrete	Water/Cement Sand/Cement, Age	Gaussian Process Regression	1030	0.884	5.702	4.058	-	-	-
			Gaussian Process Regression		0.888	5.597	3.913	-	-	-
			GPR using Exponential Kernel		0.888	5.600	3.924	-	-	-
			GPR using Square Exponential Kernel		0.878	5.849	4.242	-	-	-
			GPR using Rational Quadratic Kernel		0.880	5.793	4.182	-	-	-
			Levenberg–Marquardt ANN		0.890	5.447	4.274	-	-	-

Figure 15 shows the performance of various algorithms is assessed for the best fit on training and test data using R² values during the analysis of predictive model performance. Data points are visually presented, while the orange circles show training data, and black circles correspond to the test data. The red solid line represents a perfect prediction (Predicted=Observed), and black dashed lines represent a 20% error band. For GBR, the R² values were 0.903 for training and 0.881 for test data, showing nearly a perfect fitting, as most of the points lie within this prediction error band. Even if the GBR achieved a high R² score of 0.881, the ensemble models are crucial for engineering applications since they provide a reliable, data-driven alternative for initial lab tests to produce sustainable concrete. The models’ engineering application includes the quick assessment of high-strength concrete’s feasibility utilizing industrial waste combinations up to 77 MPa, ensuring structural safety during the assessment procedure. CatBoost model (CatB) outperformed GBR with R² values of 0.926 for training and 0.879 for test data. BR, with R² values of 0.922 for training and 0.856 for test data, showed high-quality performance but suffered from a greater loss in generalization ability than CatB and GBR. Also, RF performed similarly with R² values of 0.926 and 0.862, respectively, showing high-quality outputs but larger errors than CatB and GBR. The model HGB had an R² value of 0.912 for the training set and an R² value of 0.878 for the test data. Thus, this model generalized better compared to GBR. Lastly, ETS had the maximum R² value for the training set: 0.932. However, it suffered from a huge decline in its generalization, as evidenced from the low R² of 0.839 for the test dataset, indicating overfitting with a decrement of 0.093. The KNN model performed the poorest of all models, with R² values of 0.874 for the training data and 0.844 for the test data, hence confirming low-performance capability across the tested dataset. There were extremely few outliers in the most reliable model (GBR), with 92% of predictions having errors under ± 20%. These outliers were primarily from high-compressive strength mixtures (more than 50 MPa) with errors of more than 15 MPa. The model’s moderate inaccuracy in predicting high-strength concrete mixes was shown by the presence of outliers in this range.

Fig. 15.

Correlation between predicted and observed compressive strength values of ML models.

Model explainability and interpretability

Although the initial statistical analysis using distribution characteristics is a crucial beginning point that enables the investigator to gain a fundamental understanding of linear correlations and density distribution, it is insufficient for modeling the complex nonlinear relations typical of sustainable concrete compositions. The SHapley Additive exPlanations (SHAP) approach, which helps go beyond straightforward correlations between variables toward a comprehensive high-dimensional viewpoint of machine learning algorithms, is used to overcome this issue. SHAP values represent a consistent measure of relevance for each characteristic and its marginal contribution to all potential sets of features, such as fine aggregate or curing age. This is the main difference between SHAP values and linear correlations.

Figure 16 shows that the average impact of the different parameters in the model can be observed in the form of a bar graph in the SHAP feature importance bar chart. The most significant impact was observed in the fine aggregate with an enormous positive impact value of 10.68. The amounts of fine aggregates determine the skeletal structure and ITZ qualities since the water-to-cement ratio is linked to the porosity and chemical bond strength of the cement paste matrix. In order to achieve the ideal packing density, which ensures a reduction in the volume of cement paste required to fill the voids within the aggregate framework and avoids stress concentration, fine aggregates are crucial to the micro-filling process. Furthermore, the model’s enhanced sensitivity to the fine aggregates’ concentration may be explained by their effect on the overall surface area involved in the interaction with the cement matrix, which in turn affects the material’s resistance to cracking. Followed by the water with 2.74, and curing age with 2.72, which have approximately the same significance. After that come the medium variables that have significant effects, such as the 1.37 impact by the coarse aggregate, 0.97 by the superplasticizer, and 0.91 by the industrial waste. The cement with an impact value of 0.72, fly ash with 0.52, and silica fume with 0.18, which indicates that the less significant importance differences between supplementary cementitious materials and aggregate variables exist. The model’s reported global sensitivity to the percentage of industrial waste components and other SCM components is a direct result of their distinct physicochemical characteristics. Finely ground industrial waste materials, such as waste glass powder, waste marble powder, ceramic waste powder, limestone powder, ground granulated blast furnace slag and metakaolin, have two effects when added to the sustainable concrete matrix: physical particle packing and late-stage pozzolanic reactions. On the one hand, the small particles of glass and marble powder serve as tiny fillers, ensuring the best possible particle packing and minimizing the cementitious matrix’s porosity from the beginning. However, industrial wastes that include reactive aluminosilicates like metakaolin and glass fines contribute to pozzolanic reactions whereby the unstable Ca(OH)₂ crystals that have been produced through the hydration of Portland cement are consumed. Therefore, the latter are transformed to very stable gel forms of calcium silicate hydrates (C-S-H) and calcium aluminosilicate hydrates (C-A-S-H).

Fig. 16.

Bar plot of SHAP analysis of several different variables.

In addition, in the SHAP values graph in Fig. 17, the fine aggregate was observed as having the most significant impact, with its positive relation with SHAP Values (+ 5, + 15), but its small quantities reduce the output to -15. On the contrary, the large quantities in the dataset reduce the output with its negative SHAP values (-2, -4). However, in the curing age variable, it shows an entirely different pattern where its small quantities reduce the output significantly with − 7, but the longer age improves the output with its positive impact. Nevertheless, variables such as the coarse aggregate, superplasticizer, and industrial waste merely result in a moderate impact with an approximate value nears-zero, but the cement, fly ash and silica fume were shown with minimal impact in between SHAP values (-1, + 1). The importance of the packing density and the hydration of the concrete altered by the industrial waste is shown by the SHAP analysis results. Important factors connected to the internal skeleton structure that affect its load-bearing capacity are the volume fraction and the distribution of the fine aggregate particles. C-S-H gel forms as a result of improved hydration mechanisms brought on by aging. As a result, the transition zone between the cement paste and the industrial waste particles becomes denser as individuals age. The low importance of cement’s contribution as compared to fine aggregate suggests that, in these environmentally friendly combinations, optimizing the ratio of aggregate to binder is crucial for compressive strength.

Fig. 17.

Summary plot of SHAP analysis of several different variables.

Figure 18, the SHAP heatmap that depicts the feature importance varied concrete mix components across 140 total instances, ranging from 0 to 140 on the x-axis. Colors range from red which indicates positive SHAP values for an increase in the model output up to + 15.83, to blue which indicates negative SHAP values for a decrease down to -15.83. The most important feature is fine aggregate, it has a large negative contribution, deep blue, for instances up to 50 and then turns into a positive leading contributor, deep red, across the instances from 50 to 140. Water is highly volatile, starting with a positive contribution before (0–8), middle (40–46), and at the end (120–140) and giving way to a negative one around instance 20–35. Curing age has a strong negative, blue contribution between instances 50 and 75, while superplasticizer, industrial waste, and fly ash have a local spike for positive, red contributions in the instances that fall between 120 and 140. The cement, silica fume, and coarse aggregate are more neutral, white or pale in color for most of the instances, indicating that in the overall scheme of things for this model’s predictions, these features bear less importance than others such as fine aggregate and water. The change in density in the Interfacial Transition Zone (ITZ) is supported by earlier research. Experimental work indicates that although the substitutes for glass powder exhibit little or no hydraulic activity at first, long curing periods result in considerable positive SHAP contribution as micro voids present in the ITZ get filled with hydration products. The addition of marble powder improves the hydration process initially by providing additional nucleation sites and improving the matrix. The mixes of metakaolin undergo quick pozzolanic reaction during curing, resulting in filling of pores with an alumina gel network.

Fig. 18.

Heatmap of SHAP analysis of several different variables.

From Fig. 19, cement has a stable effect with SHAP ranging from − 2 to + 2, but outlier values reach + 6. Besides, fine aggregate has a highly non-linear effect, being negative below 400 kg/m³, SHAP − 18, and positive above 500 kg/m³, SHAP + 5 to + 20. The effect of coarse aggregate occurs at 1800 kg/m³, SHAP + 15. The absence of silica fume, at 0 kg/m³, results in a SHAP value being negative, while low quantities, 20–40 kg/m³, are slightly positive. Higher silica fume content, 80–90 kg/m³, may result in detrimental effects, SHAP as low as -5. Fly ash has a downward trend, showing low content, SHAP values ranging from 0 to -1 at 0 kg/m³, but increased, positively, at + 6, at higher quantities, 150–200 kg/m³. Water content has an optimum value between 125 and 175 kg/m³, SHAP + 13, but reducing beyond 180 kg/m³. Industrial waste has a non-linear effect, being zero below 150 kg/m³, but varying between SHAP values of -4 and + 8 at 150–250 kg/m³. Superplasticizer initially has a negative effect at 0 kg/m³, but a peak effect at + 5, around 5–7 kg/m³, tapering off. The curing age has a log function, SHAP values increasing logarithmically from − 9, at values less than 28 days, but reaching + 14 at 360 days, thus validating its critical importance in improving model performance at increased simulation times.

Fig. 19.

Scattered plots of SHAP analysis of different variables.

Figure 20 depicts that the interaction of cement content (100–600 kg/m³) and industrial waste (0–500 kg/m³) on compressive strength, ranging from approximately 24 to 32 MPa. The plateau region is noticed for cement content of 400 kg/m³ and above, while cement content below 400 kg/m³ is associated with large variability and strength reduction. Moreover, fine aggregate content of 0–1000 kg/m³ and coarse aggregate content of 0–2500 kg/m³ have a wide range of strength from 15 to 45 MPa, where strengths are noticed at particular aggregate proportions. The role of silica fume content of 0–140 kg/m³ and fly ash content of 0–250 kg/m³ has a range of strength from 27 to 33 MPa, where highest strengths are noticed at higher cement contents of silica fume. The interaction of content of mixing water of 75–250 kg/m³ and superplasticizers of 0–25 kg/m³ indicates that highest strengths of compressive strength are noticed at reduced moisture content, where increased moisture is significantly decreasing the strengths of compressive strength. Finally, an important role of curing age is noticed, where increase of curing age brings progressive increase of strengths up to 37.5 MPa with optimum cement content. Although PDP analysis provides helpful information about how each feature affects compressive strength, it is important to understand that the PDP plot only shows the average response and does not provide an uncertainty bound. Furthermore, because PDP analysis relies on feature independence, it must be cautious when analyzing linked inputs; in these situations, PDP models end up taking into account probable values of data points in the joint space. Because of this, SHAP analysis was conducted in this case study in addition to the PDP analysis.

Fig. 20.

PDP analysis for the top five features.

Graphical user interface

This section outlines a significant advancement meant to meet the practical demands of engineers and designers for the effective implementation of hybrid boosting models. A Python web application has been created that features a model with optimized hyperparameters, accessible through a Graphical User Interface (GUI) tool. This GUI is designed to predict and classify the compressive strength (CS) of concrete using the best available model for both classification and prediction, as shown in Fig. 21. Users can input values for example Cement (30–604 kg/m³), fine aggregate (0–1032.8 kg/m³), coarse aggregate (305–2649 kg/m³), silica fume (0–150 kg/m³), fly ash (0–247 kg/m³), water (78–262 kg/m³), industrial waste (0–509 kg/m³), superplasticizer (0–24 kg/m³), and curing age (3-365 days), into a streamlined layout, with the CS value and its classification displayed dynamically after input. The GBR, the GUI engine, has outstanding generalization capabilities (Test R² = 0.881). A distinct validation approach that used an external dataset not used in the initial training process was used to assess the generated GUI-based web application’s reliability. The absence of fresh laboratory test results allowed for this. A range of peer-reviewed mixtures were employed in the external validation procedure to confirm the model’s generality. The compressive strength of the exterior mixtures was predicted using the GBR, and the results showed low variation. The database contains the web application’s code. By contrasting the GUI’s output with the held-out 10% test set, its validity is confirmed.

Fig. 21.

GUI screenshot for classifying and predicting the concrete’s compressive strength.

The study promotes cooperation and the field’s acceptance of AI-driven solutions. By automating the evaluation of concrete strength, this instrument reduces reliance on conventional, resource-intensive experimental testing. As a result, it might improve material selection, streamline quality control procedures, and improve decision-making in building projects.

Conclusion

Improving the mechanical properties of sustainable concrete with the use of industrial waste materials is an important aspect in promoting sustainable construction. This is the first research to play a crucial role in using supervised machine learning to enable the estimation of the compressive strength of sustainable concrete. With the identification of key factors affecting sustainability, this research will be important in allowing reductions in costs in experiments and promoting sustainable materials for use in the construction sector.

• The study assessed seven machine learning algorithms, Gradient Boosting Regressor (GBR), CatBoost (CatB), and Random Forest (RF), to illustrate that ensemble methods can be highly effective predictive tools for the strength of concrete.

• Though the ETR and CatB models performed wonderfully on the training process, the Gradient Boosting Regressor (GBR) proved itself to be one of the best on the testing dataset with a high value of the R² statistic of 0.881.

• From SHAP (SHapley Additive exPlanations) value analysis, it is identified that the most dominating parameter is fine aggregate, followed by water content and curing age.

• It was shown that strength reaches a plateau after 400 kg/m³ of cement, while the interaction between industrial waste and other materials has a substantially contributing effect on the final mechanical strength.

• The GUI that has been developed serves as a highly intuitive interface which effectively bridges the gap between machine learning models and their implementation in practical engineering for optimizing the strength of concrete.

Sustainable Waste Integration: The study was able to identify a proper way to integrate industrial waste, silica fume, and fly ash into concrete to ensure it remains strong while being environmentally sustainable. In this case, a simple explanation would be that this research involved the use of intelligent computer algorithms that can determine how strong green concrete will be without the need to construct and then destroy the concrete in a laboratory setting each time. The study has demonstrated the effectiveness of using computer science approaches to evaluate the mechanical properties of sustainable concrete. The study has provided an accurate technique to predict the compressive strength of sustainable concrete without necessarily carrying out elaborate experiments, thanks to the application of machine learning ensembles. With this research, construction individuals can save time and money while protecting the environment at the same time by incorporating waste industry material.

Even though machine learning algorithms’ prediction skills have shown great success, there are a few limitations that must be acknowledged in order to appropriately evaluate the results. First off, despite the sizeable dataset (711 samples), there could be issues extrapolating the results to any mixes outside of the input parameter range under consideration, particularly those with odd water-to-cement ratios or curing conditions. The accuracy of the predictions provided in different geographic contexts may also be affected by the inherent variability of the chemical and physical properties of the waste utilized, which frequently depends on the region of collection and manufacturing operations. Lastly, the algorithms may have trouble predicting durations longer than the actual values since the majority of the data samples are concentrated at the normal ages for curing (e.g., 28 days).

Limitations and future works

The created ensemble learning models for the compressive strength of industrial waste-modified concrete have a high prediction reliability, it is important to highlight some limitations.

• The initial limitation is that a different train-test split structure is used to assess the predictions of model performance. Fold-by-fold reporting of standard deviations and probabilistic bounding of variance were not carried out due to the higher computational expense of performing uncertainty quantification (UQ) processes for different ensemble learning algorithms. In our upcoming research, we will address the problem of conducting comprehensive UQ, which will demand particular algorithmic design and expertise.

• Secondly, despite having a robust database with a complete collection of 711 data, there exists a natural heterogeneity within the database. This is because the concrete mixes involve different mixes of alternative supplementary cementitious materials (SCMs) and industrial wastes that might limit the model generalization within the local area due to varying chemical and physical properties.

• Thirdly, the scope of prediction in this study is limited to early age and mechanical properties only. Durability properties such as chloride permeability, shrinkage, sulfate attack resistance, and carbonation have not been considered owing to unavailability of data in the literature matrix.

Supplementary Information

Below is the link to the electronic supplementary material.

Author contributions

Mohamed AbdelMongy: Conceptualization, Methodology, Investigation, Formal analysis, Writing - original draft, Writing - review & editing. Md. Alhaz Uddin: Conceptualization, Methodology, Data curation, Validation, Supervision, Writing - original draft, Writing - review & editing. Md. Habibur Rahman Sobuz: Formal analysis, Data curation, Validation, Data curation, Writing - original draft, Writing - review & editing. Samir Mazumdar: Formal analysis, Data curation, Validation, Writing - original draft, Writing - review & editing. Mohammad Azad: Methodology, Validation, Investigation, Writing - original draft, Writing - review & editing. Md. Kawsarul Islam Kabbo: Formal analysis, Validation, Writing - original draft, Writing - review & editing. Mohammed Jameel: Methodology, Validation, Investigation, Writing - original draft, Writing - review & editing. Sani Aliyu Abubakar: Investigation, Formal analysis, Writing - original draft, Writing - review & editing.

Funding

This work was funded by the Deanship of Graduate Studies and Scientific Research at Jouf University under grant No. (DGSSR-2025-02-01516).

Data availability

The dataset developed and analyzed during the current study is provided as a supplementary file with the manuscript.

Competing interests

The authors declare no competing interests.

Ethics approval

This work is original and has not been published elsewhere, nor is it currently under consideration for publication elsewhere.