Experimental and machine learning-based analysis of red mud influence on recycled aggregate concrete properties

Abstract

This research presents a comprehensive experimental study on the compressive strength and durability performance of environmentally sustainable concrete incorporating red mud (RM) and recycled concrete aggregate (RCA). Throughout all mix designs, the quantities of fine aggregate, silica fume, basalt fiber and water/binder ratios were kept constant, while the cement and coarse aggregate contents were systematically replaced with red mud (0–15%) and recycled concrete aggregates (0–100%) respectively. The experimental results revealed that increasing red mud content led to a consistent increase in compressive strength up to 10% replacement of cement with RM and then showed the decline, with the increment of 5.46% for 0% RA, 12.77% for 50% RA, 18.45% for 75% RA and 19.14% for 100% RA corresponding to the 10% red mud replacement for 28 days curing. Similarly, replacing NCA with RCA caused strength reductions of 22.52% for 50% RA, 38.96% for RA 75% and 57.07% for 100% at 10% replacement of cement with red mud and 28 days of curing, respectively. A similar trend is seen in acid resistance of red mud-based RAC and this reduction reaches 65.1% for 0% RM, 61.7% for 5% RM, 59.44% for 10% RM and 60.62% for 15% RM when recycled aggregates replacement reached from 0 to 100% respectively. However, acid resistance of concrete increased from 0 to 10% inclusion of red mud and then decreased at 15% replacement. In the second phase of the study, the potential of machine learning (ML) algorithms to predict the compressive strength of concrete mixes containing RM and RCA was evaluated. A total of four ensemble-based ML models including random forest (RF), KNN, AdaBoost and XGBoost were trained and tested using 251 datasets for RM and RCA mixes, respectively. Among the models, XGBoost exhibited the highest accuracy and reliability in predicting compressive strength. The study also included an analysis of the influence of individual input parameters on the model’s predictive outcomes using SHAP analysis and found the curing age is the most influential parameter while RCA, SF and NCA are the following one.

Introduction

Global waste is greatly influenced by the construction sector which accounts for 30% of all waste and produces more than 33% of all carbon emissions¹. Because of this, one must take steps to deal with the harmful impacts of construction waste on both the environment and the economy^2–4. Managing waste effectively is crucial to reaching sustainable development by lowering reliance on landfills. A strategy for this is to use these waste building materials again in civil engineering projects. This allows the industry to use renewable materials that lessen the reliance on traditional energy-intensive processes. Waste generated during construction or demolition such as recycled concrete is valuable for civil infrastructure and aids the environment. Due to how durable, flexible and economical it is, as well as being easily available for most construction projects, concrete is the most used building material globally^5–8. Using red mud, another type of industrial waste, with recycled aggregates is a more environmentally friendly approach to managing problems caused by mining and construction/demolition waste⁹. The Bayer process, used in making alumina, leaves behind a red mud by-product that is very alkaline and its production is rising worldwide; while much of it is landfilled, this can cause environmental risks¹⁰. Over the past two decades, studies have examined how red mud can be used for different purposes in cementitious mixtures and the results suggest it can improve many concrete features^11,12.

Some studies,^13–15, indicate that red mud can contribute favorably to concrete, especially when with supplementary cementitious materials or recycled aggregates. Red mud improves the durability and strength of concrete when used in correct amounts, but because it contains high alkaline levels and heavy metals, it must be treated or stabilized first¹⁶. Red mud also helps cut down the use of natural resources, since it has a specific activity and a very small particle size which make it suitable to replace a part of the cement in concrete¹⁷. Yet, more than the recommended amount of red mud can reduce the ease of working with the concrete mixture because it is so fine which can, in turn, make it drier and need more water¹⁸. So, when the amount of red mud does not exceed 30% and chemicals and recycled aggregates were also used, this helps maintain and even increase the strength and quality of high-strength concrete^19,20. In addition, research indicates that red mud in concrete can impact on how chemically durable concrete is, primarily in severe environments, because red mud has high alkalinity and many reactive compounds²¹. Because red mud is reactive, it can create products like ettringite or different complex hydration phases which may cause the concrete to crack under the effects of moisture²². However, some approaches can help overcome these problems, like combining red mud and materials like fly ash, metakaolin or silica fume to make the mix more stable and durable²³. Other solutions are using the right concrete proportions, placing importance on proper curing and using red mud at reduced levels (usually below 20%) to manage harmful compounds²⁴. It is demonstrated by research that blending red mud (up to 15%) into cement in recycled aggregate concrete, together with other supplementary materials, improved strength and made the concrete less susceptible to chemical attacks²⁵.

In addition, scientists say that negative effects from using red mud could also happen with ordinary concrete if alkalis and reactive aluminates are included in the mixture²⁶. In addition, recycling concrete is now widely used to give new life to waste produced during the demolition and restoration of old concrete structures^27–29. Construction and Demolition Waste (C&DW) and the precast concrete industry supply much of the recycled concrete aggregates and contribute greatly to the circular economy in the construction field^30,31. Using red mud, instead of part of the cement, recycled aggregates can make concrete manufacturing more environmentally friendly by decreasing cement usage and making use of leftover materials³². This way of use helps solve problems with red mud wastes and gives a greener option for conventional concrete binder^23,33. L. Kan et al. study¹¹ revealed that including up to 15% red mud in place of cement led to stronger and more durable recycled aggregate concrete because of filler and pozzolanic effects. In addition, it is noted that red mud and RCA, when used jointly, caused the concrete to be less easy to mold because of the angular shape and porosity present in the aggregates and fine red mud; nevertheless, it contributed to higher strength development through time³⁴. Keeping the red mud content at the right level prevented a major drop in concrete’s strength and minor changes to the mix could take care of the slump issue without harming the structure³⁵. Incorporating red mud and recycled concrete aggregates (RCA) in concrete can affect several properties, for example, density, compressive strength, elastic modulus, shrinkage and water absorption³⁶.

The severity of these effects is reliant on the chemical and physical traits of red mud (like fineness, alkalinity and iron), the connected mortar and the porosity of the aggregate²⁸. When red mud is added to concrete with RCA, to ensure the strength isn’t reduced by alkalis, its ratio in the mixture needs to be carefully controlled³⁷. The water–cement ratio in RCA’s original concrete is very important. Lower water to cement ratio in the original concrete can lead to lower concrete strength, mostly if red mud is present, since it has both finer particles and demands more water³⁸. Additionally, mechanically processed RCA with less adhered mortar and rounded edges improved the workability of red mud concrete, which is otherwise sensitive to changes in water content³⁹. Their tests indicated that concrete made with red mud and RCA can achieve comparable compressive and tensile strengths to that of conventional concrete, provided red mud content is kept below 15%. However, they also noted an increase in drying shrinkage and a reduction in abrasion resistance by about 10–15% due to the combined effects of RCA porosity and the reactive nature of red mud. Moreover, studies found that while carbonation resistance and water absorption remained within acceptable limits, the mechanical performance particularly bond stress and early strength was slightly lower in RM-RAC mixes compared to mixes with natural aggregates and Portland cement alone⁴⁰. Estimation of properties of red mud-based concrete (RMC) normally involves a huge laboratory testing that can be costly, time-consuming, and labor-intensive⁴¹. Also, the feasibility of conducting comprehensive experimental studies can be obstructed by the existence of practical difficulties; e.g., specific curing and storage conditions⁴².

Several parameters affect the strength of concrete such as the curing time, mix design and type of aggregates used which take much time and effort to make a thorough analysis^43,44. In view of these difficulties, recent research works have recommended some of the approaches to determine compressive strength (CS) and most of them are based on empirical formulae used to enhance computation speed and eliminate site-specific drawbacks^45,46. Nonetheless, there is a tendency that such empirical formulas do not always succeed in taking into account the nonlinearity of relations between different characteristics of concrete and the obtained strength⁴⁷. Because of this, there is an increasing demand of more rigorous and precise predictive models that can more adequately correlate the input variables to the compressive strength of RMC. In this respect, Artificial Intelligence (AI) methods have recently attracted a lot of attention because they can rapidly learn data, represent complex behaviors, and make precise predictions on the basis of various types of inputs^48,49. Most diverse studies have been based on predicting the mechanical response of different types of concrete with the assistance of diverse machine learning (ML) algorithms^50–54. For instance, ensemble deep neural networks accurately predict the compressive strength of fly ash concrete, achieving superior model precision with low variance⁵⁵. In a similar vein, Random Forest (RF) algorithm outperformed other models, such as the M5 rule-based tree, M5 prime tree, and chi-square automatic interaction detector, in predicting the mechanical properties of roller-compacted concrete pavement⁵⁶. Meanwhile, some researcher introduced an artificial neural network (ANN) model for predicting the compressive strength of recycled aggregate concrete⁵⁷.

Similar study used a multi-objective metaheuristic optimization method to calculate the compressive strength of recycled aggregate concrete (RAC) with the use of pozzolanic materials⁵⁸. A relative contrasted various prediction method such as linear regression, nonlinear regression, quadratic model and artificial neural networks (ANNs) to predict the compressive strength of concrete prepared using ground granulated blast furnace slag⁵⁹. In a similar manner, the impact of different mix proportions on compressive strength of Ultra-High Performance Fiber Reinforced Concrete (UHPFRC) using a mixture of prediction models including linear regression, pure quadratic, M5P-tree, and ANN to determine the most effective model⁶⁰. In another study, they formulated four predictive models, namely, linear, nonlinear, multi-logistic regression, and M5P-tree, to determine the strength of high-strength concrete (HSC) in the presence of metakaolin⁶¹. Moreover, different indexing approaches to determining the strength properties of polymer-modified concrete and showed that the most precise approaches were nonlinear regression, ANN, and the M5P-tree model⁶². Similarly⁶³, applied four prediction models ANN, full quadratic (FQ), multi-logistic regression (MLR) and M5P-tree to forecast compressive strength of traditional concrete with steel slag aggregates. These authors in a later publication forecasted the strength of ordinary concrete reinforced with carbon nanotubes based on five models: ANN, M5P-tree, nonlinear regression (NLR), MLR and linear regression (LR)²⁹. To be more exact, enhancing of predictive efficiency of the already existing artificial neural network (ANN) model by integrating it with two optimization techniques Particle Swarm Optimization (PSO) and the Imperialist Competitive Algorithm (ICA).

The compressive strength of fly ash concrete was predicted by using this combined methodology⁶². In a different related study⁶⁴, examined a data frame of 236 records of fly ash concrete samples obtained in various research sources. They applied four alternative prediction models (Full Quadratic [FQ], Nonlinear Regression [NLR], Multilinear Regression [MLR], Artificial Neural Networks [ANN]) that they took into consideration the impact of the varying specimen size and shape to the compressive strength. Similarly, 455 concrete mix designs that used fly ash in varying amounts and ground granulated blast furnace slag⁶⁵. They were attempting to estimate the compressive strength of the NSC as well as the HSC by the models such as LR, ANN and the Pure Quadratic methods. Meanwhile⁶⁶, have realized good predictive capability (R = 0.98) in cases where the Random Forest (RF) algorithm was used to predict the strength of carbon nanotube-reinforced cement composites. The last and seemingly the most significant method of studying the properties of the red mud-based materials was the model and prediction of the important properties using machine learning techniques. Another study forecasted with high accuracy (R² = around 0.94–0.96) the compressive strength of red mud fly ash concrete as the road base material (ML-based, i.e., neural networks and regression) cross-validated on a diversified mix design⁶⁷.

Another author estimated compressive strength, porosity, and water absorption of red mud GGBS sustainable bricks using multiple linear regression models with high R² up to 0.98, RMSE between 0.223 and 0.465, and MAPE of 3.49%–5.09%⁶⁸. The ANN model to optimize the strength of modified red mud mixtures intended for infrastructure used, achieving R ≈ 0.995, MAE = 0.119, RMSE = 0.149, and average relative error of 5.58%⁶⁹. Similarly ⁷⁰, utilized MLP, VR and XGBoost in the study to predict the workability and compressive strength of Class F and Class C fly ash-based geopolymers based on 156 datasets. XGBoost provided the most optimal results (workability: R² = 0.96; strength: R² = 0.89). CaO, added water and fine aggregate were significant workability factors whereas fly ash and added water were significant strength factors. Class F attained 79.96 MPa with improved workability and Class C contained 61.45 MPa with poor workability because of high CaO content. A backpropagation (BP) neural network in predicting unconfined compressive strength and optimized the mix design of red mud-based road base materials is utilized. The model fit using 114 data points was highly accurate (R² = 0.995, RMSE = 0.149, MAE = 0.119) with more than 74.56 percent of predictions falling within 10 percent error. The best network topology was 6-10-1 with train ML algorithm. The optimal mix design (0.02% polymer agent, 1.96% fly ash, 4.78% cement, 33.93% water) produced the 7-day strength of 2.987 MPa with the cost of 17.099 CNY/ton, which demonstrated the possibility of ANN in the design of red mud-based road materials⁶⁹.

The paper aims to fills a research gap by attempting to used red with RCA and then to predict the compressive strength (CS) of recycled aggregates concrete with red mud through implementing standard machine learning models which includes KNN, RF, XGBoost and AdaBoost. It analyzed 251 data samples taken through experimental work and past research and took into account ten input features which helped it to achieve significantly higher prediction results. The performance of the models was assessed by comparing the predicted results and actual data through various statistical measures in training and testing processes. Moreover, SHAP analysis were used to explain the effect of each input feature on the predicted CS. Such interpretability tools enable sensible optimization of the mix design, helping the researchers to reach the desired strength values using a minimal number of experimental runs which will eventually result in saving time, effort, expenditure, and materials. In short, while previous studies explored red mud or RCA separately, limited research has integrated experimental analysis with ML-based prediction and feature importance explanation (SHAP). This study addresses that gap.

Methodology

The research is based on a combined experimental and machine learning methodology. The study aims to assess the effect of red mud on recycled aggregate concrete (RAC). The experimental part implies the preparation of concrete mixes with red mud as well as RCA contents varying and testing of their compressive strength and acid resistance characteristics following standardized test methods. Subsequently obtained data is used to train machine learning models, including Random Forest, XGBoost, AdaBoost and KNN, to compressive strength prediction and feature importance analysis based on SHAP values. This combined approach guarantees empirical certification and information based on data to enable the formation of sustainable concrete materials.

Materials

Coarse aggregates, fine aggregates, cement, and water with superplasticizer are the components which are utilized to make RMC. The coarse aggregates consisted of natural coarse aggregate (NCA) and recycled concrete aggregate (RCA) whereas query sand were used as the fine aggregate. At different percentage intervals, red mud is added to the concrete at partial replacement in place of cement to improve the quality of concrete and red mud was procured from Khushab, Pakistan. The ACI codes are followed in the Mix design. The type of cement used was the ordinary portland cement of grade 53 and the properties of cement compared to the red mud are indicated in Table 1.

Table 1.

Physiochemical properties of cement and red mud.

	Particles size (µm)		Setting Time (min)				Soundness (%)		Consistency (%)		Fineness (m2/kg)
			Initial		Final
Physical properties	OPC	≤ 50	30		600		0.8		30		225
	Red mud	≤ 90	–		–		–		–		457
Chemical properties	LOI (%)		Al2O3	Fe2O3	SiO2	Na2O	CaO	SO3	P2O5	K2O	MgO	TiO2
	OPC	2.77	5.7	4.3	28.3	0.17	57.8	1.95	–	0.51	1.27	–
	Red mud	12.1	63.49	5.28	11.78	1.617	3.04	–	–	0.652	1.44	1.601

In the study, the natural coarse aggregates are crushed basalt dolomite of small particles measured less than 12.5 mm to suit the dimensions of samples to prevent possible segregation in the mixtures. Also, particles of NCA bigger than 12.5 mm were resistant through the sieve of 12.5 mm. RCAs were obtained as a result of the manual crushing of old tested concrete and the material of which is the crushed basalt dolomite. These concrete samples averagely had compressive strength of 35 MPa. The steps in which the RCA production process went through included; crushing of the concrete samples using a heavy hammer to small sizes up to 150 mm, grinding these stones to the required size of particles in an electric crushing machine at the concrete laboratory, separating fine granules smaller than 2 mm by placing the sample into a sieve with openings 2 mm, and also excluding the particles size larger than 12.5 mm using 12.5 mm sieves, washing the RCAs well with water by immersing them for 24 h and lastly drying them up. Figure 1 reveals the NCA and RCA grading curves. It was revealed from the grading curves that NCA and RCA were acceptable within limits recommended⁷¹.

Fig. 1.

Gradaton curve of NCA and RCA.

Some of the experiments were conducted on coarse aggregates (NCA and RCA) to determine their properties and the obtained values were shown in Table 2. The absorptions of NCA were less as compared to RCA since old cement was present on the surface of the particles. The biggest coarse aggregate (NCA and RCA) needed was 12 mm, which was selected so that it would suit the small size samples that were to evaluate.

Table 2.

Physical properties of fine and coarse aggregates.

Property	Fine aggregates	Coarse aggregates	Recycled aggregates
Abrasion resistance (%)	–	19.37	33.59
Fineness modulus	2.42	3.19	3.87
Water absorption (%)	3.91	3.66	4.58
Specific gravity	2.74	2.96	2.19
Bulk density (kg/m3)	1574	1643	1451

Sample preparation

A total of 384 RMC samples have been cast and these mixes were based on the standard specifications ACI 318–19⁷² for a target mean compressive strength of 60 MPa of the controlled concrete and these mixes are represented by using mix design table as given in Table 3. The weight ratio of red mud to cement was replaced to 0, 5, 10 and 15%^73,74. The NCA was replaced by 50, 75 and 100% RCA respectively^75,76. The RA were done after being crushed and fully immersed in normal water over a period of 24 h to keep the water-binder ratio consistent during the process of casting. Moreover, the water saturated RCA was maintained at the sack in the dry air till saturated surface dry condition was reached. The cement and the RM were adequately mixed with the dry components of the concrete for 2 min. The dried mix of concrete was slowly added to water and the pan mixer of 40 litter capacity that rotates at 20 to 24 RPM was left to mix the concrete in 5 min, consequently, the homogeneous fresh concrete was ready. The fresh concrete workability was preserved by applying sika superplasticizer and the value of slump and density was obtained. Immediate casting of concrete was done in the 100 × 200 mm cylinder and this concrete has been compacted with hammer manually. The samples of hardened concrete piece were taken out of the molds in 24 h, and since then have been stored in the curing tank during 7, 14, 28, 56 and 90 days.

Table 3.

Mix design of red mud concrete.

Group ID’s	Cement (kg/m3)	Red mud (kg/m3)	Silica-fume (kg/m3)	Basalt fiber (%)	FA (kg/m3)	NCA (kg/m3)	RCA (kg/m3)	W/B ratio (kg/m3)	Super-plasticizer (kg/m3)
RM0-RA0	508.75	0	41.25	10.2	632	920	0	160	2.54
RM0-RA50	508.75	0	41.25	10.2	632	460	460	160	3.82
RM0-RA75	508.75	0	41.25	10.2	632	230	690	160	5.1
RM0-RA100	508.75	0	41.25	10.2	632	0	920	160	6.34
RM5-RA0	481.25	27.5	41.25	10.2	632	920	0	160	2.54
RM5-RA50	481.25	27.5	41.25	10.2	632	460	460	160	3.82
RM5-RA75	481.25	27.5	41.25	10.2	632	230	690	160	5.1
RM5-RA100	481.25	27.5	41.25	10.2	632	0	920	160	6.34
RM10-RA0	453.75	55	41.25	10.2	632	920	0	160	2.54
RM10-RA50	453.75	55	41.25	10.2	632	460	460	160	3.82
RM10-RA75	453.75	55	41.25	10.2	632	230	690	160	5.1
RM10-RA100	453.75	55	41.25	10.2	632	0	920	160	6.34
RM15-RA0	426.25	82.5	41.25	10.2	632	920	0	160	2.54
RM15-RA50	426.25	82.5	41.25	10.2	632	460	460	160	3.82
RM15-RA75	426.25	82.5	41.25	10.2	632	230	690	160	5.1
RM15-RA100	426.25	82.5	41.25	10.2	632	0	920	160	6.34

Experimental work

Compressive strength

The compressive strength of the cylindrical concrete specimen was tested in the current research in line with the recommendations set out in ASTM C39/C39M⁷⁷, Standard Test Method for Compressive Strength of Cylindrical Concrete Specimens, cited in ACI 318 and utilizing nominal diameter and height of 100 mm and 200 mm respectively. A semi-automatic compression testing machine (ACTM) having the greatest loading capacity of 3000 kN was used in the testing process. All the concrete samples were set up properly to have consistency in loading the concrete. Each mix group was ultimately tested in three specimens and average compressive strength was computed and reported to characterize the performance of the group.

Acid resistance test

The performance of red mud-based recycled aggregate concrete was tested with the provisions of the ASTM C267⁷⁸. Samples were prepared as concrete cylinders (100 mm x, 200 mm), cured in hardened water conditions by prior immersion for 3 days, which were then placed in 5% sulfuric acid (H₂SO₄) solution at room temperature in order to provide an aggressive environment. The solution of the acid was recharged every week to concentrate. Compressive strength was taken after 28, 56 and 90 days exposure to acid in accordance with 3000 kN semi-automatic compression testing. Three specimens were tested per mix at a time and their average strength was recorded.

Machine learning algorithms

Extreme gradient boosting (XGB) algorithm

The XGBoost (XGB) is an improved version of the frequent Gradient Boosting Machine (GBM) method proposed by Chen and Guestrin⁷⁹. It is a supervised machine learning algorithm that has gained complexion in current ML practices owing to high resilience, accuracy, and computing cost-effectiveness^80,81. Its ease in solving complex problems using less learning time has contributed to its reliability as opposed to other learning algorithms. XGBoost can be directly applied to the boosting principle, which means that in one way or another a series of weak learners is added one after another to aggregate a strong predictive model^82,83. Every new learner, through an iterative process of refinement, is capable of fixing the mistakes of the previous ones, and the final model comes out as so accurate⁸⁴. To allow accurate prediction, the algorithm employs second-order Taylor series expansion of loss function, that is it considers both first and second derivatives of the loss function to adjust the weights of the model.^85,86. It calculates the negative derivative of the loss of the actual model in every iteration and updates the new model in lines with it. The new model is subsequently appended to the ensemble and the procedure repeated until some predetermined accuracy is attained⁸⁷. Besides, the XGBoost loss function has terms to prevent overfitting using regularization to improve generalization⁸⁸. The common expression of the prediction model used in XGBoost is given in Eq. 1.

1

In above expression the f_k(x_i) represents the learning function at the p_th stage, while and denote the predicted values at phases p, and p−1 respectively, with x_i serving as the input feature. The XGBoost model provides a structured approach to evaluate the performance and effectiveness of the selected model using Eq. 2.

2

Here, denotes the loss function (LF), n represents the total number of observations, and σ stands for the regularization component, which is derived using the following Eq. 3:

3

In this context represents the vector of scores assigned to the leaves, γ denotes the minimum loss threshold required to proceed with further leaf splitting, and λ refers to the regularization parameter used to control model complexity.

Adaptive boosting (AdaBoost) algorithm

AdaBoost (Adaptive Boosting) is an effective bagging algorithm that developed by Freund and Schapire (1996) that transforms a series of weak classifiers (usually decision stumps) into a strong classifier by paying more attention on the training examples that are difficult to learn. In the beginning equal weights assigned to each training instance by the Eq. 4:

4

At each iteration t, a weak learner is trained to minimize the weighted classification error (Eq. 5):

5

The importance (weight) of the weak learner in the final ensemble is calculated using Eq. 6 as:

6

The sample weights are then updated to emphasize misclassified examples (Eq. 7):

7

After updating, the weights are normalized by utilizing the Eq. 8:

8

The final strong classifier is the weighted majority vote of all weak learners (Eq. 9):

9

AdaBoost adapts by focusing on the errors of previous models, making it highly effective for classification tasks, although it may be sensitive to noisy data or outliers.

Random forest

The Randon Forest (RF) classifier, first proposed by Breiman, works based on bagging mechanism, comprising two main steps: bootstrap sampling and aggregation. It is an ensemble learner and it uses the bagging method to build several decision trees⁸². Wherein these trees are trained in parallel, random subsets of the data thereby offering efficient and robust learning^89,90. RF has a few hyperparameters controlling the depth, size, and degree of randomness that would be introduced into the model⁹¹. Bootstrap sampling uses the original dataset to create several sub-samples of the dataset, as shown in Fig. 3 which graphically describes the operation of the RF algorithm⁸². To increase the accuracy of prediction and to decrease the overfitting possibility, every tree in the forest takes into account a random fraction of features. In the process of tree building, the best feature is chosen at every node so as to split the data and this plays a great part in the capability of the model to generalize well^66,92. Doing so, the average of all trees is taken as the final output of the RF regression model⁹³.

Fig. 3.

Acid resistance concrete with RM and RCA levels at different curing days.

K-nearest neighbor (KNN) algorithm

K-Nearest neighbors (KNN) algorithm uses the intuitive but powerful idea that similar data objects are likely to be close to each other in the feature space⁹⁴. KNN uses k closest most instances in the data set and uses the values to predict the outcome of a new sample, being that of compressive strength. The similarity is the extent to which a degree of the similarity can typically be considered based on distance metrics, e.g., Euclidean distance, Manhattan distance or a Minkowski distance⁵². In this study, the Minkowski distance was applied, expressed as stated on Eq. 10:

10

where p being the power parameter. This measure of distance involved the variables that included: the type and percentage of graphene oxide (GO), type and percent of superplasticizer, method of dispersion and age at curing of concrete. The prediction value will be calculated by averaging the six nearest neighbor values of compressive strength, and in the end, the final predicted compressive strength is calculated.

SHapley Additive exPlainations (SHAP)

SHAP was first introduced by Lundberg and Lee in 2017, is an explanation method that uses the basics of game theory. It gives each feature a quantitative rating, which is the extent of its contribution to a prediction made by a machine learning model^95,96. The input features in this case are regarded to be the participants of the cooperative game, and the overall reward or payout is regarded to be the model prediction. The primary role of SHAP is to assess the effect of every feature on the output of model prediction^84,93,97. In addition, SHAP attaches an importance value to each feature considering individual predictions, this can help in providing the behavior insight at fine grain level. From a mathematical perspective, SHAP models the output of the model as a combination of input features multiplied by weights, and it adds up these for each feature to compute the contributes each of them with each other using the Eq. 11⁹⁵.

11

The Shapley values (ϕ_i) in this case are the coefficients of the linear expression of the model and is a simplified expression of the original input vector and g is the explanation model. In any Shapley value, it is possible to refer to a single contribution of the predictor variable to the overall model output of the machine learning⁹⁸. The values are calculated with the help of the Eq. 12 as it was first introduced by Lundberg and Lee⁹⁵.

12

In this notation, F is the entire input feature set, whereas S is a subset of F and is the set of all of the features except the feature at index i. Within the framework of the current paper, the SHAP tool was used to explain the predictive outcomes of the hybrid XGBoost model.

Result and discussions

Compressive strength

The compression tests were performed on 100 × 300 mm cylindrical samples at 7, 14, 28, 56 and 90 days for accelerated curing. The testing equipment adhered to ASTM C 39 standards to perform failure tests on the specimens. The strength records displayed in Fig. 2 revealed identical patterns among all tested specimens while RM levels increased the strength but RA presence reduced it. The use of Recycled Concrete Aggregates (RCA) in concrete production has gained attention as a sustainable alternative to natural aggregates, but it significantly influences the properties of the resulting Recycled Aggregate Concrete (RAC). RCA typically contains adhered mortar from old concrete, which makes it more porous and less dense compared to natural aggregates. This leads to increased water absorption and microcracks, affecting both the fresh and hardened properties of RAC. One of the most critical areas impacted by RCA is the Interfacial Transition Zone (ITZ), which is the region between the aggregate particles and the cement paste. In RAC, the ITZ becomes more complex and heterogeneous because it includes the interface between the new cement paste and the old adhered mortar, as well as between the new paste and the original natural aggregate. Studies have shown that this region is generally weaker and more porous in RAC than in conventional concrete, which negatively affects the mechanical performance of the^28,99,100. A similar behavior is observed in this research also and the strength reduction of 32.4, 50.3 and 65.7% is noted for 0, 50, 75 and 100% RA inclusion respectively. The similar finding are also described by^101,102. However, these reductions in strength can be minimized through appropriate mix design strategies, such as lowering the water-cement ratio, pre-saturating the RCA to limit further water absorption during mixing and incorporating supplementary cementitious materials. Increased compressive strength emerged throughout all test series when the use of RM at 10% cement replacement reached acceptable levels. Various researchers including^103–105 confirmed that 10% represents the best replacement ratio for cement. Tang et al.¹² studied how RM functioned as a substitute for FA in concrete among their research findings noting that zeolite among other mineral compounds from RM assists concrete internal curing which results in enhanced compressive strength at later stages¹⁰⁶. At RM replacement levels above 10% compressive strength dropped according to findings reported by^104,105. The research conducted by Ortega and his colleagues¹⁰⁷ demonstrated that mortar strength weakened after exceeding 10% RM usage in mortars. In fact, it was reported that RM has a low pozzolanic reaction compared to that of cement, restricting high contents of RM in the construction industry as cement replacement^108,109. However, using RM at low level of cement replacement in concrete mixes could be effective in terms of compressive strength¹⁰⁶, as the high fineness of RM particles and its filler effect can densify the matrix of concrete¹⁰⁹. Mixes containing 10% of RM across different groups had a compressive strength of 69.5, 50.9, 40.1 and 28.2 MPa at 28 days when RA increased from 0, 50, 75 and 100% respectively. When this mix of RM10 was compared across the groups of 50, 75 and 100% of RCA, the strength reduction was found 14.8, 25.6 and 37.5% as shown in Fig. 2 The main reason for decreased calcium content in the mixture results from higher replacement levels of RM. The change in material composition weakens phase transformations which leads to decreased hydration gel product creation and weakens structural properties²². The particle size of RM exceeds the characteristic size of OPC which leads to additional properties. The material strength reduces as RM usage rises beyond certain quantities due to its increased porosity that creates lower matrix density. Therefore, the combination of RM10-RA75 can be used in this combination for structural concrete as high strength concrete.

Fig. 2.

Variation of compressive strength with RM and RCA levels at different curing days.

Acid resistance test

Acid resistance of red mud concrete was tested through compressive strength of the specimens after 28, 56 and 90 days in 5% sulfuric acid solution and the results were compared with those samples which was cured in tap water for 3 days and their results are shown in Fig. 3. The acid resistance of concrete decreases when it contains recycled aggregates because RAs have increased porosity and higher water absorption which makes them more susceptible to aggressive chemical attacks along with residual mortar in the mixture¹¹⁰. The use of higher recycled aggregate levels in concrete results in increased mass loss under sulfuric and nitric acid exposure according to¹⁰¹. Changes in mass loss reach over 10% in concrete made with 100% recycled aggregates compared to natural aggregate concrete with a loss of about 2%^111,112. Similar results are observed in present study, the acid resistance of concrete decreases as the percentage of recycled aggregates increases which can be seen via strength reduction within the groups. The maximum strength reduction is recorded for RM0-RA100 which is as low as 4.9 MPa. A 60–65% strength reduction is observed for each group when recycled aggregates rises from 0–100% for 28 days of acid exposure. The performance of concrete can be enhanced by implementing supplementary cementitious materials (SCMs) like fly ash and silica fume and metakaolin since they refine the pore structure and lower permeability^113,114. Red mud (bauxite clay) I utilized in this research for this purpose and an opposite trend to recycled aggregates is observed across the groups, strength is increased due to the addition of red mud. The RM10-RA0 mix shows the lowest reduction in compressive strength exhibited the highest durability against sulfuric acid attack at all curing ages when compared with controlled sample e.g. 28 days compressive strength of RM10-RA0 was 22.98 MPa which is 14.33% higher than control samples having compressive strength of 20.1 MPa.

Acid resistance increased 6.25% with 5% and 6.7% with 10% inclusion of red mud but 3.52% decreased for 15% inclusion when no recycled aggregates are incorporated respectively. The concrete’s sulfuric acid resistance depends on mineral additive characteristics particularly particle size shape and texture because these features produce mixed positive and negative outcomes. Research indicates that red mud properly applied leads to improvements in sulfate resistance and chloride permeability and long-term strength. The material achieves these enhancements because it works as a filler while generating extra hydration processes according to^115,116. The durability of concrete mixtures significantly improves by adding red mud or bauxite clay because of its small particles and pozzolanic reaction characteristics and distinctive chemical elements. Red mud inserted into void spaces of concrete mixtures enables microstructural refinement that decreases porosity and enhances chloride ion resistance which results in improved durability²³. Red mud usage outside recommended levels creates adverse effects on concrete properties by causing workability issues and time delays while heavy metal extraction increases¹¹⁷. Using large quantities of red mud sometimes leads to reduced sulfuric acid resistance because its porous structure produces more voids in the mix which compromises durability according to¹¹⁸.

Machine learning

Dataset description and analysis

A wide-ranging data set with 269 samples was utilized for the current study where 84 were from experimental testing, while the rest was obtained from previous researchs^119–123. Many approaches relying on machine learning have been able to predict the strength of concrete using sample sizes just like or in some cases smaller than what was used here. Zhao et al.¹²⁴ reported in their review that many researchers use data with as little as 67 samples or as many as 226 and over 10 input features for their experiments. According to machine learning experts, using less than 10 pieces of data for each feature should be avoided¹²⁵.

Nine (9) input features are used from a dataset of 269 samples, providing a ratio of 1:30 which is much higher than what is typically suggested. Red mud (RM), silica fume (SF), basalt fibers (BF), ordinary Portland cement (OPC), natural and recycled aggregates, superplasticizer (SP), as well as water, were all part of the concrete mix used. For the present study, the input parameters were OPC, RM, SF, BF, FA, NCA, RCA, W/B ratio, curing period in days, and SP dosage. The input and output variables are provided in Table 4 along with their respective statistical metrics. The summary addresses important descriptive statistics such as minimum, maximum, SD, kurtosis, and skewness for every parameter. Low SD in RM, SP, BF, SF, compressive strength (CS), and water indicates that the values are mostly clustered around the mean. On the other hand, the higher SDs for Cement, NCA, RCA, and FA point to a higher degree of variation. The skewness of a distribution measures how its values are asymmetrically distributed about the mean and may take values of positive, negative, zero, or undefinable¹²⁶. A positive skew usually means data cluster at the left of the distribution, for example, as displayed in Fig. 4 for RM, CS and FA, while a negative skew implies higher data density on the right, such as in the case of cement NCA. Kurtosis, which is usually between − 10 and + 10, determines if a particular set of data has fatter or thinner tails than a normal distribution¹²⁷. It provides the view of vertical shape of the probability distribution. Table 4, the increased kurtosis value in case of the BF is attributed to the lack of available data.

Table 4.

Statistical measures of the RMC data sets.

Variables (kg/m3)	Type	Minimum	Maximum	Mean	Standard deviation	Kurtosis	Skewness
Cement	Input	306.52	508.75	426.58	50.25	− 0.62	0.08
RM	Input	0	131.40	38.51	32.66	− 0.87	0.34
SF	Input	0	85.40	24.73	27.46	− 0.63	0.64
BF	Input	0	10.20	3.25	4.76	− 1.40	0.78
FA	Input	589	772.00	658.22	47.31	0.78	1.09
NCA	Input	0	1189.70	873.49	387.57	− 0.05	− 1.20
RCA	Input	0	920	190.16	311.14	0.17	1.30
Water	Input	160	240	186.26	26.45	− 1.28	0.56
SP	Input	0	7.15	2.87	2.43	− 1.12	0.36
CS (MPa)	Output	17.17	110	45.44	17.33	1.68	1.15

Fig. 4.

Distribution graph of input parameters.

Pairplot charts in the Fig. 5 give a visual explanation of the association between compressive strength in ready-mix concrete and various materials that are mixed in. There is an abundance of CS data taken between 7 and 28 days, showing how common it is to test concrete in these early stages. Cement is usually most concentrated, at around 400 to 500 kg/m³. Since different studies report varying amounts of RM, the connection between RM and CS is not easy to understand. Also, the water distribution exhibits a similar complexity, ranging from 160 to 240 kg/m³ and with clear double peaks. However, most of the data falls below 7.15 kg/m³ for the superplasticizer content. In the case of natural coarse aggregate (NCA) and recycled coarse aggregate (RCA), a density of 1000–1280 kg/m³ is typically seen to be the range in density.

Fig. 5.

Pairplot graph of input parameters.

Any number higher than 0 on the correlation coefficient represents a positive link between two items and any value lower than 0 means the link is a negative one. Stronger relationships are shown by high absolute values. Figure 6 demonstrates that the main correlation between each input feature and compressive strength is within − 0.42 and 0.52. Figure for these correlations were constructed using the Seaborn library in Python. Looking at the heatmap, see that four input variables are connected positively to compressive strength. Altering age corresponds the most strongly to CS, indicated by the high and consistent amount of data (R = 0.52). Similarly, the correlation relationships between the silica fume (SF) and natural coarse aggregate (NCA) and compressive strength (CS) indicate a significant correlation, that is, the correlation coefficients (R values) between the fumes and that of natural coarse aggregate are 0.40 and 0.39 respectively. Also, there exists a slight correlation between decreased strength and the percentage of the recycled coarse aggregate (RCA) content within the concrete mass. Fine aggregate also declines, matching the growth trend for the other variables. Cement and SP share a linear connection with CS, demonstrating that increasing their amounts tends to improve strength until maximum improvement is achieved. Results from recycled material (RM) suggest that a small increase in strength is achievable with the correct RM addition.

Fig. 6.

Correlation heatmap of variables.

Quantitative performance assessment of models

This work used four main indicators to assess the efficiency of the created soft computing models. Frequently, the trustworthiness of ML algorithms is found by looking at different kinds of statistical error measures. RMSE, R², MAE and MAPE are the main indicators that are applied. A short summary of the formulas for each metric can be found in the following Table 5. Due to these tools, one can easily see how well the planned ML models function and learn important things about their prediction abilities^96,128,129.

Table 5.

Evaluation criteria for regression model accuracy.

Evaluation metris	Abbreviation	Equation	Standard value
R-square	R2		0–1
Root mean square error	RMSE		0 is best
Mean absolute error	MAE		0 is best
Mean absolute Percentage error	MAPE		1 is best

The main goal was to observe the impact of design parameters on compressive strength by exploring how machine learning works, which motivated the comparison of four models. Both simple algorithms, like linear regression and advanced ones such as extreme gradient boosting, are included. “K-Nearest Neighbors (KNN), Random Forest (RF), XGBoost and AdaBoost” are considered modern techniques that effectively handle the complicated behavior found in concrete strength . To forecast results, KNN often chooses familiar (nearby) designs from the input data and has been found to be highly accurate by recent work from Duan¹³⁰. Also, RF, XGB and AdaBoost are tree-based models that exhibit excellent ability to predict compressive strength¹²⁴. Sorting design parameters into branches on a tree helps models reduce variability and allows the structure to form logical rules that show strength accurately. For their effectiveness, RF and XGB are now standard in research projects¹³¹. Figure 7 illustrates how machine learning algorithms are used to predict the compressive strength of Red Mud concrete (RMC) in this study. The main phases in the study are data collection, data validation, picking a suitable model and checking the model’s performance. There were ten input features and one output variable in my study. Out of the data, 80% used for training and the remaining 20% for testing. The study used four different algorithms to perform its predictions. Four ways to evaluate performance were used: R-squared, RMSE, MAE and MAPE. The predicted results for compressive strength were checked against the true strength values. SHAP analysis was also used to find out how input features affected the results and show the impact from visuals.

Fig. 7.

Flowchart of machine learning implementation for concrete strength prediction.

XGBoost algorithm prediction performance

Figure 8a shows how the results from the Extreme Gradient Boosting (XGB) model are compared to the measured strength of Red Mud Concrete (RMC). The graph demonstrates how well the XGB algorithm performed by looking at its actual predictions versus the experimental values. The consistent relationship between the model’s predictions and measurements reveals that XGB’s predictions are reliable, with only minor differences in realizing compressive strength. In addition, Fig. 8a clearly displays how well the XGB model estimates compared to the real experiment results. During training, the results show that errors are small, typically less than ± 5 MPa. However, once the framework is tested, larger deviations occur in the calculation, up to around 9 MPa. While the model performs well on both data sets, testing shows more uncertainty which is what tends to happen when applying a model to data that wasn’t involved in training it. Figure 8b shows the scatter diagram of XGB predicted compressive strength and its true value including both training and testing phase. The plot shows the agreement between predicted and reference points and their associated ± 25% error bands centered at the best-fitting regression line. In an ideal scenario, a perfect prediction model would result in a regression line with a slope of 1, indicating a one-to-one relationship between predicted and observed values. In this figure, the scatter is indicating the model prediction versus the data that has been predicted for both training and validation data in which ± 25% error boundaries are captured which indicates the model is robust. The coefficient of determination (R²) values also prove the accuracy of the model with an outstanding R² of 0.998 for the training set and a good R² of 0.985 for the testing set.

Fig. 8.

(a) XGB model cs relative error training vs testing; (b) XGB model cs correlation predicted vs actual.

Here, it can be seen high R² values, which indicates that the XGB can model the underlying pattern in the data well, that signifies the good predictive power and the ability to generalize to unseen data. Compared to others, XGB is superior because it results in more accurate and similar predictions.

K-nearest neighbor algorithm prediction performance

In Fig. 9a the compressive strength results measured in RMC are compared with the predictions using the KNN model. The outcomes show that the difference between the observed and forecast numbers is moderate, yet both training and testing maximum errors go over 20 MPa, much higher than those noted in XGB and RF. It can be seen that 0.904 was achieved for the R² value during training and the same sample achieved 0.879 during testing in Fig. 9b. Even so, these values are relatively accurate, but the KNN model performs feebly compared to the separate models (XGB and RF). Even so, all R² values are nearly the same which means the data is dependable and accurate. Moreover, the small gap between how the traditional models perform during training and testing implies overfitting is not an issue here.

Fig. 9.

(a) KNN model CS relative error: training vs testing; (b) KNN model CS correlation predicted vs actual.

Randon forest algorithm prediction performance

Comparison of actual and predicted compressive strengths for Red Mud Concrete is given in Fig. 10a. Prediction errors related to both training and testing are also pointed out in this method. Most of the errors with these models are in a 10 MPa range of the correct results. Importantly, most of the errors are not very large, which proves the strength and accuracy of the proposed model. Further investigation indicates that about half the prediction errors are smaller than 2.5 MPa and a further 20% are between 2.5 and 5 MPa. Approximately 10% of our predictions have a difference of up to 10 MPa. The results show that the model works well both on data it saw earlier and on unseen cases. While there are little differences, they stay within the normal range for engineering specifications. Its performance is proven by a high coefficient of determination (R²) which yields a value of 0.992 over training and 0.971 over testing. Accordingly, the model reveals its ability to recognize difficult patterns in the data without relying solely on learning from the training set. Figure 10b highlights the actual and predicted values for compressive strength in a scatter graph. Both the training and testing data points are closely grouped near the 1:1 line and within 25% of the ideal ratio. The fact that most of the data lie close to the ideal line and the low error of about 10 MPa suggest that the predictions and observations are very close. All of the findings conclude that the RF model in the model is able to accurately predict concrete compressive strength for a variety of conditions.

Fig. 10.

(a) RF model CS relative error: training vs testing; (b) RF model CS correlation predicted vs actual.

Adaptive boosting algorithm prediction performance

The actual compressive strength values of RMC are compared with those estimated by the ADB model in Fig. 11a. Even though the ADB model tracks the mean trend of the experimental data, the errors between its results and the measured values are higher than those from other models such as XGB and RF. More specifically, the errors at the greatest points in both training and testing are larger than 15 MPa, showing that ADB does not accurately predict the values in these cases. Scatter plot data in Fig. 11b show how well predicted and actual compressive strength values are distributed for the training and testing groups. The R² coefficient for training was 0.894 and for testing was 0.901. The results suggest that the ADB model is reasonably good at predicting, but less so when compared to XGB, KNN and RF models. Even with a relatively low overall accuracy, the remainders of the R² values between training and testing are close which suggests overfitting is not an issue for the ADB model. Despite some lack of precision, the model is able to work reasonably well on different kinds of data. However, compared to using XGB or other ensemble learning models, ADB is not as good when it comes to this application’s predictions.

Fig. 11.

(a) AdaBoost model CS relative error: training vs testing; (b) AdaBoost model CS correlation predicted vs actual.

Performance evaluation of machine learning models: In this study, a number of statistical techniques were applied in order to determine the effectiveness of the AI models. These techniques comprised coefficient of determination (R²), root mean square error (RMSE), mean absolute error (MAE) and mean absolute percentage error (MAPE), and the outcomes are listed in Table 6 and graphically presented in Fig. 12. All the metrics give informative information on the model predictive performance where it can be seen as accurate and where the prediction does not represent the actual values. Different datasets and machine learning models were used to cross-validate the reliability of predictions. This method was used to have a detailed assessment of the predictive abilities of the models and how they were consistent when new and unknown inputs were encountered. Figure 12 provides a summary of the R² values of different machine learning models. All in all, the majority of models worked well in predicting the outcome with their values of R² nearing the number 1 meaning they were very accurate. Experimentally, data suggest that a model with a high R² can make more accurate predictions¹³². Particularly, the R² of the models were as follows: XGB attained R² of 0.985, RF 0.971, KNN 0.879 and ADB 0.901 in testing. Nevertheless, the KNN model had a low R² values than the others. Although R² is a great indicator of accuracy, it must be mentioned that the complexity of the model can influence this indicator. As an example, the high R² of XGB was attributed to its capability to fit more complex trends, but it frequently cost more in complexity.

Table 6.

Evaluation metrices of models.

Model	R2		RSME		MAE		MAPE
	Test	Train	Test	Train	Test	Train	Test	Train
KNN	0.904	0.879	5.217	6.348	3.85	4.779	9.77	11.359
RF	0.992	0.971	1.508	3.124	1.012	2.436	2.502	5.193
XGB	0.998	0.985	0.723	2.237	0.397	1.676	1.094	3.96
ADB	0.894	0.901	5.468	5.721	4.123	4.402	10.964	10.377

Fig. 12.

R², RME, MAE and MAPE of different ML models.

There is a trade-off between predictive accuracy of the model and the computational resources needed to train and make inferences. This can be taken into account during the process of choosing the model that can be used in practice, when a more simplistic model, like RF (R2 = 0.971) can be chosen due to its speedy processing, even though it has a slightly lower accuracy. In addition to R², RMSE was also computed for all models, and the results confirmed that XGB achieved the lowest RMSE of 2.237 MPa, outperforming RF slightly with an RMSE of 3.124 MPa, while performing far better than both KNN at 6.348 MPa and ADB at 5.721 MPa. The RMSE indicates that XGB’s predictions were closer to the actual values, but this performance came at the expense of higher training time and complexity. It’s important to note that while XGB had a lower RMSE, its increased complexity may not always justify its use in real-time applications, especially when faster, less complex models such as RF (RMSE = 3.124 MPa) are nearly as effective. The trade-off between the predictive power and the computational efficiency of these models needs to be carefully considered based on the context in which they will be deployed.

Further evaluation through MAE and MAPE also supported the superior performance of XGB and RF. For MAE, XGB and RF both had values of 1.676 MPa and 2.436 MPa, respectively, indicating high reliability. However, KNN and ADB had higher MAE values of 4.779 MPa and 4.402 MPa, respectively, suggesting they were less effective at providing accurate predictions. The high MAPE observed in ADB (10.377%) and KNN (11.359%) indicates that these models were less reliable in predicting the compressive strength of RMC. This limitation is important when making decisions about model selection based on both accuracy and complexity. The MAPE values for XGB (3.96%) and RF (5.193%) were much lower, showing their superior ability to generalize well to new data. Thus, XGB and RF are recommended for applications where accuracy and consistency are paramount, though practical considerations such as model complexity and computational costs should also factor into the final decision.

Model transparency and feature importance using SHAP values

To learn how much each input parameter affects the prediction of CS in RMC, the research applied SHAP methodology. To interpret how important each feature is in the prediction process, the study applied SHAP to XGB and examined average SHAP values linked to each feature. As shown in Fig. 13, the SHAP values are sorted from highest to lowest, highlighting the variables that played the biggest part in affecting the model’s output. The SHAP analysis points out that curing days are by far the top contributor to predicting concrete compressive strength, with a very significant mean SHAP value of + 7.38. This is important as the time taken to cure isn’t the only factor, it also depends on how much the cement in the concrete has hydrated. More time to cure means cement particles become fully hydrated, causing the creation of C-S-H which then fills holes and holds different components together. Less curing than required means the concrete does not absorb enough water, leading to fragile areas and an increase in its openness. The model shows that good curing is important because, even with ideal other materials, strength can’t be reached without proper curing. For this reason, curing days determine how well and how long the concrete’s internal structure matures. Recycled Concrete Aggregates (RCA) hold second high value in importance, with a SHAP value of + 3.46 to highlight the rise in using sustainable materials for concrete. RCAs can enhance or decrease the way concrete compressive strength is developed. When graded and compared to natural aggregates, they become nearly the same. Even so, using RCAs may result in old mortar, visible cracks or more water soaked into the wall, weakening the masonry. The fact that RCA has a large SHAP value suggests that the data contained mixes of differing RCA quality and caused predictions to be largely affected. With a SHAP value of + 3.14, the ultrafine particles and strong pozzolanic reactivity of silica fume are well known. With calcium hydroxide, a hydration by-product, it transforms to further C-S-H, improving the structure and lessening pore size. Because of this, concrete is denser, stronger and more durable which is the reason for its strong influence in the model. Natural Coarse Aggregate (NCA) adds a high value (+2.92) because it provides the structural support within the mix. A stronger and more evenly graded aggregate enables the mix to be strong, resistant to weight and better able to resist cracks.

Fig. 13.

Mean SHAP values indicating variable importance in ML model.

With a SHAP value of + 2.63, basalt fibers work to enhance how crack avoidance and energy absorption in concrete composites. Despite not raising compressive strength, they assist in load distribution, prevent some cracking, control shrinkage during early curing and so preserve concrete’s structural strength. Workability and how densely a mixture can be packed are influenced by fine aggregates (+ 2.15). Sand graded correctly is able to fit between the larger grains, creating a dense structure in the concrete with fewer small problems. Red mud has a moderate SHAP value of + 1.03 and is an output of aluminum production sometimes used as an extra cementitious material. While pozzolanic activity is lower than silica fume, its ability to take part in binding phases and fill pores means it plays a part in building strength. Its chemical compositions differ widely which makes its contributions unreliable. Cement, in fact, turns out to have the smallest impact (+ 0.71) out of all the features considered. This does not make cement less important, in this database, its proportions in samples were probably the same or its influence was supported by SCMs such as silica fume and red mud. When making modern concrete, designers use materials that are based on their abilities rather than simply the cement content. The model thinks that the results mainly depend on curing, SCMs and aggregate quality when their effects are much stronger than those of cement. The model clearly shows that high strength in concrete comes from a combination of time, materials and additives working together.

Figure 14 represents the importance of the individual input variables which is expressed by a separate distribution of colored points on the chart. The x-axis in this visualization provides the SHAP values which show the effect of each input feature on the result. A higher concentration and clearer colors of the points allow us to see how each feature contributes to the model’s outcome and its direction. The summary plot reflects important findings about which input features have the strongest impact on the machine learning model’s ability to predict concrete compressive strength. Each line shows how much the data changes and dots represent samples that are colored by the amount of feature. All things considered, curing days have the greatest impact. Stronger predicted strengths are associated with increased treatment duration (measured as red) which agrees with established concepts in civil engineering. Blue cells which represent shorter curing times correspond to SHAP values that reflect weaker concrete. Next in significance is recycled concrete aggregates, since including a lot of recycled material generally makes a structure less strong. The combined SHAP graph for silica-fume and basalt fiber reveals that the impact on strength is not constant for all added quantities, making it both helpful and harmful in various applications. Natural coarse aggregate helps by improving load distribution when its value is high, but natural fine aggregate usually leads to buildups in value that help little or even harm the strength when interlocking or workability are issues. Compared to other factors, red mud and cement show a moderate effect, with cement having a slightly better impact than first believed. This analysis reveals the main elements important for predicting compressive strength and guides improvements in the mix design.

Fig. 14.

Visualization of feature influence and value distribution using SHAP.

Conclusions

Artificial Intelligence (AI) is gradually being recognized by construction industries for helping to estimate the mechanical parameters of concrete, especially when used together with experimental work. The study is designed to check if different types of machine learning models can successfully predict compressive strength from data obtained from experiments and literature. The purpose is to see how various raw materials influence the compressive strength of RMC. This analysis includes two boosting strategies called XGBoost and AdaBoost, the RF method used in bagging and K-Nearest Neighbors (KNN) which is an independent learning algorithm not based on ensemble strategies. Ten input features are examined to examine how they affect the proposed predictions. According to the results, the study reaches the below conclusions:

• Replacement of natural aggregates with recycled aggregates affects the compressive strength of concrete and the detrimental effect became severe with each replacement level. Red mud performed its part to overcome this negative impact up to some extent and the optimum dosage of RM is 10% be the weight of cement.

• A similar trend is observed for acid resistance also the acid resistance of concrete reduces as soon as the percentage of RA increases and this reduction reaches 60–65% when 100% RA is incorporated as compared to NAC. However, concrete containing RM showed better performance than natural concrete up to 10% of RM replaced and then followed the same trend when replacement level reached 15%.

• The relationship was found to be positive between both input and output variables, curing age, natural coarse aggregates (NCA) and silica fume, but negative with fly ash (FA) and recycled coarse aggregates (RCA).

• The highest value of estimating the compressive strength of red mud concrete (RMC) accuracy was found with XGBoost (XGB) and Random Forest (RF). Experiments showed that XGB and RF are more precise than AdaBoost and the KNN model gives good but not extraordinary results.

• Among the different models, the XGB model did the best, reaching an R-squared value of 0.998 on training data and predicted compressive strength with a high level of accuracy, at 0.985 on testing data. Alternatively, the AdaBoost (ADB) model did not work as well, results being R-squared values of 0.894 and 0.901 on the training and testing sets. In the same way, R-squared for RF (0.992 for training and 0.971 for testing) and KNN (0.904 for training and 0.879 for testing) were much higher than the random choice.

• Both the training and test datasets showed good results because the models had low RMSE, MAPE and MAE. They indicate that the models are good at learning and can accurately foresee new outcomes.

• It was found by SHAP analysis that curing age, recycled coarse aggregates (RCA) and silica fume are the top three factors that positively affect the predicted strength of the concrete. curing age was highlighted to be most significant, with an important rating of 7.38. But red mud and cement were not closely tied to the changes in compressive strength.

10% red mud and 75% RCA can be viewed as a viable sustainable substitute to the conventional concrete and improve its power and environmental friendliness. The method will not only minimize waste but it will also decrease the carbon footprint associated with reducing the use of virgin materials. However, it is necessary to be aware of the limitations in the study. This could be improved by addressing chemical and physical material properties during the process of building the predictive model. As a result, models might be more correct and well-understood. Things such as how materials react chemically, changes in temperature and their resistance to corrosion can influence the strength of concrete when under observation. Even so, because data was limited, this study did not closely examine these influences. For that reason, studies should pursue increasing the range of data by incorporating a variety of inputs and carefully analyzing combined models that join machine learning approaches and deep learning approaches. It is also important to evaluate the possible use of ML models to predict key properties of RMC such as strength in flexure and tension, how long they last, porosity, how easily they can be used, shrinkage, expenses and environmental effects (carbon footprint).

Author contributions

IH: Conceptualization, Methodology, Investigation, Formal Analysis, Validation, Writing—Original draft. TA: Conceptualization, Methodology, Investigation, Writing—Original draft. MZQ: Conceptualization, Methodology, Supervision, Writing—Original draft. NBK: Methodology, Supervision, Software, Validation, Writing—Original draft. AM: Conceptualization, Methodology, Supervision, Writing—Original draft. GUA: Conceptualization, Methodology, Supervision, Writing—Original draft. AA: Investigation, Supervision, Formal Analysis, Validation, Writing—Original draft.

Data availability

The data that support the findings of this study are available on request from the corresponding author. The data are not publicly available due to privacy or ethical restrictions.

Declarations

Competing interests

The authors declare no competing interests.