High-Strength Self-Compacting Concrete Production Incorporating Supplementary Cementitious Materials: Experimental Evaluations and Machine Learning Modelling

Abstract

This study investigates mechanical properties, durability performance, non-destructive testing (NDT) characteristics, environmental impact evaluation, and advanced machine learning (ML) modelling techniques employed in the analysis of high-strength self-compacting concrete (HSSCC) incorporating varying supplementary cementitious materials (SCMs) to develop sustainable building construction. The findings from the fresh characteristics test indicate that mixes’ optimal flowability and passing qualities can be achieved using different concentrations of marble powder (MP) alongside a consistent amount of silica fume (SF) and fly ash (FA). Moreover, the incorporation of 10% MP along with 10% FA and 20% SF in HSSCC significantly improved the compressive strength by 14.68%, while the splitting tensile strength increased by 15.59% compared to the reference mix at 56 days. While random forest (RF), gradient boosting (GB), and their ensemble models exhibit strong coefficient correlation (R²) values, the GB model demonstrates more precision, indicating reliable predicted outcomes of the mechanical properties. Following subsequent testing, it has been demonstrated that incorporating SCMs improves the NDT properties of HSSCC and enhances its durability. The finer MP, SF, and FA particles enhanced microstructural performance by minimizing voids and cracks while improving the C–H–S bond. As noticed by its lower CO₂-eq per MPa for SCMs, the HSSCC mix with up to 15% MP inclusion increased mechanical strength while reducing the environmental footprint, making it an eco-friendly concrete alternative.

1 Introduction

Concrete is one of the most frequently utilized materials for construction purposes in every nation because of its exceptional structural stability and durability. A highly dispersed aggregate, cement, and water mixture makes concrete. The aggregate may impact efficiency and durability of concrete, but the strength of concrete may also be constrained. Reusing waste materials is an effort the engineering community makes to preserve natural resources (Yong and Lee, 2022; Raghunath et al., 2019). In-depth research on the root cause as well as the consequences of concrete degradation is necessary before placing concrete in various sites. There is a substantial durability issue in places like marine environments, underground, etc., due to assault from carbonation, sulphate, acid, and chloride. Self-compacting concrete (SCC) is a clear solution to the aforementioned issue.

There have been numerous studies investigating various partial cement alternatives for SCC (Khan et al., 2023; Rahman Sobuz et al., 2023; Sadek et al., 2016; Hasan et al., 2022). If the investigation illustrates that using marble powder (MP) as an inorganic addition adhering to SCC is acceptable and effective, that may also be a preferred disposal approach. The potential applications of MP, including enhancing soil, the approach to hydrogenation, asphalt placement, pottery, ceramics, thermoset epoxy composites, and polymeric manufacturing methods, have all been investigated in recent years (Aruntaş et al., 2010). Numerous studies recommend the incorporation of MP as a substitute for cement in cementitious materials (Aditto et al., 2023; Aliabdo et al., 2014; Elyamany et al., 2014). Meera et al. (2020) investigated the use of marble powder in SCC revealing positive results. According to the study performed on SCC, Lija and Minu (2016) concluded that replacing cement with a mixture of 15% marble waste and 15–30% silica fume strengthened the flexural strength and stiffness. Additionally, the durability and the chloride resistance of ternary combined SCC comprising large-scale fly ash and silica fume were experimentally investigated by Wongkeo et al. (2014) and they reported that the integration of fly ash in SCC considerably increased the durability qualities. Utilizing fly ash ranging from 10 to 50%, Dhiyaneshwaran et al. (2013) conducted a study presenting improved durability. El-Chabib and Syed (2013) also investigated the impact of a quaternary mix, including additional cementitious elements. The findings suggested that concrete having 10% silica fume (SF) provided enhanced compressive strength. The durability characteristics associated with normal strength self-compacting concrete and their impact on RC construction were examined experimentally by Sideris and Anagnostopoulos (2013). The outcomes were contrasted with concrete that had been vibrated traditionally. Researchers Vasusmitha and Rao (2013) looked at how various mineral additives affected the compression and permeation of HSSCC. According to the findings, HSSCC had less chloride permeability onto concrete, a greater bond and strength against tension, less oxygen permeability, and less plastic settling.

Moreover, sorptivity, porosity, and chloride ion permeability tests were used by Kanellopoulos et al. (2012) to examine the durability characteristics of SCC and traditional concrete. Tennich et al. (2017) evaluated SCC’s durability, utilizing tile and marble wastes that sulphates had attacked. Leung et al. (2016) reported using SF and FA in combination for SCC. Higher resistance was observed when the researchers evaluated the implications of SF and FA on the sorptivity characteristics of SCC. Using proportions of 0–30% of the entire amount of cement volume, Sardinha et al. (2016) replaced the MP. They claimed that MP replacement in cement produced an approximate compressive strength. Likewise, Ahmad et al. (2023) replaced cement with marble waste which resulted in a 20% increase in compressive strength and durability. The MP and the marble tile fragments were added to the SCC by Tennich et al. (2015) as filler. According to their findings, adding MP enhanced the compressive and tensile strength of the material. Jalal et al. (2015) examined the impact of fly ash, silica fume, and nano-silica on the hardened characteristics of a superior-performance SCC. The influence of marble powder (MP) on the strength of compression of SCC was examined by Uysal and Tanyildizi (2012). According to Jain et al. (2013), combining data from ultrasonic pulse velocity testing and Schmidt hammer tests provided a more precise way to calculate the compressive strength of concrete than depending just on the outcomes of distinct experiments. Belouadah et al. (2021) further explored the potential for non-destructive methods to evaluate the physico-mechanical characteristics of concrete using MP and obtained positive results for up to 10% MP replacement. Furthermore, using a non-destructive test and marble sand powder ranging from 5 to 25%, Velci Shridevi et al. (2023) assessed the strength qualities of SCC and discovered increased strength after 28 days. According to Breysse (2012), it is possible to determine the compressive strength of concrete using either the Schmidt hammer or ultrasonic pulse velocity assessments individually or together.

In the construction industry, artificial intelligence and machine learning are advanced techniques to quantify the fresh and mechanical properties of SCC using formulated mixture ratios, focusing on various strength predictions (Ahmad et al., 2021; Datta et al., 2024; Habibur Rahman Sobuz et al., 2024; Xie et al., 2020). ML approaches can accurately predict mechanical characteristics, enabling tasks like classification, correlation, regression, clustering, and enhancing the mechanical as well as fresh properties of concrete (Silva et al., 2020; Xu et al., 2019). Azimi-Pour et al. (2020) employed a support vector machine to evaluate the fresh and mechanical properties of SCC, incorporating fly ash as a potential substitute for cement. Asadi Shamsabadi et al. (2022) reported that ML models using XGBoost and ANN informational models accurately predicted the compressive strength of WMP concrete with a high degree of reliability (R² > 0.97). The utilization of multiple machine learning models to forecast the compressive strength of SCC with the inclusion of marble powder and rice husk ash was reported by Mahmood et al. (2023). Efficient SCC mixture compositions and compressive strengths were part of the extensive dataset used to train the ANN model by Acikgenc Ulas (2023).

There have been relatively very few comprehensive studies addressing the combined impact of SF, FA, and MP SCMs in the production of HSSCC. However, the literature lacks comprehensive information regarding the combined effects of SCMs, such as FA, MP, and SF, on durability, non-destructive testing (NDTs), and machine learning (ML) modelling. Therefore, this study introduces several novel contributions to the field of HSSCC by exploring the synergistic effects of incorporating 0–20% MP alongside 20% SF and 10% FA as SCMs. This investigation provides a comprehensive analysis that not only evaluates the mechanical properties and durability of HSSCC but also explores ML modelling to predict these properties accurately compared to existing studies. The unique approach combines experimental evaluations with ML techniques viz. random forest (RF) and gradient boosting (GB) models, offering a novel perspective on the prediction of concrete properties, which has not been extensively covered in previous works. Moreover, the study brings to light the environmental benefits of utilizing SCMs by demonstrating a significant reduction in CO₂ emissions without compromising the mechanical efficiency of the concrete. A combined investigation into the microstructural characteristics and the interaction between destructive and non-destructive testing results, this study finds the opportunity to extend valuable insights into the optimization of HSSCC mixtures for sustainable construction practices. The rigorously designed experimental program including microstructural properties and a predictive ML model represents a pivotal step towards the use of SCMs in concrete research, particularly in achieving a balance between enhanced mechanical properties, durability, and environmental sustainability.

2 Materials

Locally produced Ordinary Portland Cement (OPC) of type-I, which followed BDS EN 197–1 (BDS.EN:197–1), was utilized for this study. The coarse aggregate originated as angular stone from a nearby source. The coarse aggregate used had an optimum size of 19 mm. Natural river sand passed through a 4.75 mm sieve and was employed as a fine aggregate (Astm, 2018). Silica fume was purchased from a local supplier in Dhaka, Bangladesh. Fly ash was obtained for free from a reputable cement manufacturer in Bangladesh. Marble powder was also collected free of cost from an industrial firm which is situated in Narayanganj, Bangladesh. A superplasticizer based on polycarboxylic ether was employed to achieve the requisite workability. Fig. 1 illustrates the gradation curve for marble powder (MP) as well as coarse and fine aggregate with lower and upper limits following ASTM C33 (ASTM-C33., 2003). The physical characteristics of materials provide several benefits, including strength, workability, durability, and chemical resistance. The physical features of the substances are presented in Table 1. The chemical reactions between the components influence the overall effectiveness of HSSCC, whereas fly ash lowers permeability, silica fume strengthens density, OPC hydrates, and marble powder improves strength. Table 2 illustrates the findings of an X-ray fluorescence (XRF) analysis that evaluated the chemical properties of OPC, MP, SF, and FA.

Fig.1

Gradation curve of marble powder, fine and coarse aggregate

Table 1 Characteristics of the substances of physical composition

Table 2 Chemical compositions of cement (OPC), fly ash (FA), marble powder (MP), and silica fume (SF)

3 Experimental Program

3.1 Mix Design and Concrete Mixing

To investigate the characteristics of SCC in line with the standards established by the EFNARC, several concrete mixtures were created at a constant water-to-binder ratio of 0.36 with OPC, MP, FA, and SF at varied proportioning ratios (EFNARC, 2005). The mix proportions for every SCC batch are displayed in Table 3, whereas ‘S’ implies silica fume, ‘F’ implies fly ash, and ‘M’ implies marble powder content. To partially replace cement, MP, FA, and SF were utilized. The replacement rates of 0, 5, 10, 15, and 20% of MP in place of OPC enable an assessment of the impacts on concrete strength, durability, and other characteristics. Experimental conditions are kept consistent and uniform by maintaining consistent proportions of 10% FA and 20% SF, thereby isolating the implications of MP substitution. Also, in real-world scenarios, the availability and applicability of alternative materials for concrete production might change, which is reflected in the variation of MP proportions depending on waste materials. Through the consideration of these variables, the experimental design provides a thorough investigation of the influence of MP substitution on concrete qualities while preserving uniformity and practical applicability. The control mix called S20F0M0 contains 20% SF, 0% FA, and 0% MP.

Table 3 Mix proportion for all SCC mixtures

3.2 Preparing and Curing Samples

The compressive, splitting tensile strength, and water permeability evaluations were performed on cylindrical molds measuring 100 × 200 mm. Concrete discs 100 mm in diameter and 50 mm thick were saw-cut from 100 × 200 mm cylinders as the samples for the sorptivity and RCPT tests. After 24 h in an area having an average temperature of about 27 ± 4 °C, samples were taken out and labelled for easier identification. The samples were submerged completely using potable water and cured for 7, 28, and 56 days using a plastic treatment chamber. Each specimen was assembled and cured in a laboratory per ASTM-C192/C192M-18 (ASTM-C192/192 M).

3.3 Evaluation Practice

According to the BS EN 12350–8 (BS.EN:12350–8, 2010) and BS EN 12350–12 (EN.BS:12350–12, 2010) regulations, flowability and passing ability of fresh materials were assessed using the slump flow and J-ring flow tests. Experimental methods for verifying all fresh qualities are shown in Fig. 2a, b.

Fig. 2

Test setup on HSSCC of (a) slump flow (b) J-ring flow (c) compressive strength (d) splitting tensile strength (e) water permeability (f) sorptivity (g) RCPT (h) rebound hammer and (i) UPV assessment

The compressive strength test was performed in line with ASTM C39 standards (ASTM-C39, 2010) to evaluate the cylinder specimen in this investigation. To examine the influence of FA, MP, and SF regarding the compressive strength of HSSCC, cylindrical molds with dimensions of 100 × 200 mm were selected after undergoing water curing for durations of 7, 28, and 56 days. The cylinder specimen used in this investigation was tested using the splitting tensile test in compliance with the ASTM C496 standard (ASTM-C496, 2011). This would be conducted as the same procedure as the compressive strength test.

Fig. 2e depicts the procedure used to evaluate water permeability following BS EN 12390–8 (EN.BS-12390–8, 2019). HSSCC specimens were subjected to sorptivity testing following ASTM-C 1585–13 (ASTM-C1585, 2013), as highlighted in Fig. 2f. According to the ASTM C1202 (ASTM-C, 1202, 2019) Rapid Chloride Penetration Test (RCPT), a concrete specimen saturated with water is exposed to a 60 V DC voltage for 6 h, incorporating the apparatus and cells indicated in Fig. 2g.

The compressive strength for each NDT sample was assessed using the digital rebound hammer test following ASTM C805 (ASTM-C805, 2018), which is illustrated in Fig. 2h at 28, and 56 days after curing the process. Ultrasonic pulse velocity (UPV) experiments were carried out to analyze the pulse velocities at 28 and 56 days following ASTM C597 (ASTM-C597, 2009), as can be seen in Fig. 2i. In Table 4, the RCPT, the rebound number, and the UPV were utilized to define the charge passed as well as concrete quality in structures.

Table 4 RCPT, rebound number and velocity criterion for concrete

Microstructural properties were assessed using a scanning electron microscope (SEM) in compliance with ASTM C1723-16 (ASTM-C, 1723, 2016). The hardened concrete specimens were analyzed under a scanning electron microscope to understand the compactness of the interfacial transition zone. The concrete specimens were cut into 5 $\times$ 5 mm sizes and soaked in isopropanol for 72 h, followed by drying at 40 °C. Later, the concrete specimens were vacuum desiccated for 24 h coated with 5 nm carbon and analyzed at 15 keV voltage and beam intensity of 14.0 under an SE detector at high vacuum (S. A. Rahman et al., 2023a, 2023b, 2024).

3.4 Predictive Machine Learning Ensemble Model

Researchers have developed ensemble methods to enhance prediction accuracy, which merges several machine learning (ML) models into a single predictive model (Al Daoud, 2019). Ensemble methods, like boosting and bagging, which combine RF and GB algorithms, provide satisfactory outcomes. Equations 1, 2, 3, 4 were implemented to evaluate the impact of the model utilizing multiple metrics. Fig. 3 depicts the framework executed in this investigation to generate both single-based machine learning and ensemble model algorithms.

$${\text{R}}^{2} \, = \,1 - \frac{{\sum\limits_{i = 1}^{n} {(Y - Y_{i} )^{2} } }}{{\sum\limits_{i = 1}^{n} {Y_{i}^{2} } }}$$

(1)

$${\text{MAE}}\, = \,\frac{1}{n}\sum\limits_{i = 1}^{n} {\left| {Y - Y_{i} } \right|^{2} }$$

(2)

$${\text{RMSE}}\, = \,\sqrt {\frac{1}{n}\sum\limits_{i = 1}^{n} {(Y - Y_{i} )^{2} } }$$

(3)

$$\text{MSE }=\frac{1}{n} \sum_{i=1}^{n}{(Y-{Y}_{i})}^{2}$$

(4)

where R² = correlation coefficient, MAE = mean absolute error, RMSE = root mean square error, and MSE = mean square error.

Fig. 3

Flowchart diagram for the proposed ML approach

The random forest (RF), established by Leo Breiman in 2001 by integrating classification and regression trees with bagging, is one type of decision tree that delivers the most enhanced presentation when predicting the outcome (Breiman, 2001). When used as a predictive estimator, RF considers extensive randomized decision trees and updates them through data from various subsets of the training data. By smoothing out the variations and biases of individual trees, this approach lessens overfitting and enhances generalization.

Gradient boosting (GB) progressively creates an ensemble by repeatedly training a new model to emphasize improperly organized training data from previous approaches. Jerome H. Friedman discovered gradient boosting, a method for developing strong prediction models that can efficiently deal with complicated patterns and linkages in the information (Friedman, 2002). GB is nowadays a widely applied algorithm because of its capacity to attain accuracy levels and manage multiple data categories.

With a dataset of mixed compositions and experimental values, the accompanying Python script uses scikit-learn to train and assess two ML models: RF and GB. The models are trained and utilized for prediction-making after the data has been loaded and prepared. The forecasts from both models are then averaged to provide an ensemble prediction. This procedure is carried out in a Python environment using libraries like scikit-learn for machine learning methods, pandas for data processing, and numpy for mathematical operations. The syntax is procedural, meaning that the script is organized sequentially, beginning with the import of required libraries and ending with the presentation of the results. Splitting the dataset into training, testing, and validation subsets is an important step in this procedure. The function mentioned in the supplied script divides the dataset at random into training and testing sets in a predetermined proportion. Through the use of data splitting methods such as train_test_split, the Python code may improve the overall resilience and dependability of the ML process and guarantee a more accurate model assessment.

4 Results and Discussion

4.1 Effects of the Fresh Characteristics of HSSCC

4.1.1 Slump Flow and J-Ring Test

Fig. 4 illustrates the J-ring and slump flowing for the reference mix and several additional SCC combinations. The slump rates appeared within 675–740 mm, per the accepted criteria governing HSSCC slump flow established by the European Federation of National Associations Representing for Concrete (EFNARC) (EFNARC, 2005). Slump flow analysis showed that mixture S20F10M10 had an optimal value of 740 mm instead of the control mixture’s 675 mm. Additional analysis noticed the slump’s flow rates for each of the mixes, including 0, 5, 10, 15, and 20% MP, in addition to the steady 10% FA and 20% SF, were 1.85, 7.40, 9.60, 5.90, and 5% higher than the control mixture S20F0M0, respectively. The increase in performance may be associated with the finer particles of MP, which generate less friction internally and occur in a smoother flow. The maximum slump rate was recorded at 20% substitution, as Choudhary et al. (2019) found that the flow rate of slump enhanced with each additional MP substitution as well as exceeding 655 mm.

Fig. 4

Slump flow and J-ring flow rates of HSSCC specimens

In contrast, the J-ring flow value enhancement was detected when marble powder was included in the HSSCC combinations. It was found that the optimal J-ring flow measurement for the S20F10M10 combination appeared to be 715 mm, which came very close to the slump value of 740 mm. When compared to the control mix, the J-ring flow measurements increased by 4.18, 11.68, 19.17, 16.59, and 15% in a combination that included 0, 5, 10, 15, and 20% replacement of MP alongside 10 and 20% FA and SF, accordingly. Particles of the small size of MP, FA, and SF fill out the pores along with voids among coarse particles forming a more homogeneous mixture that can flow easily inside the J-ring with no separating. An almost identical investigation was performed by Belaidi et al. (2016), who found that adding natural pozzolana and marble powder as alternatives to OPC improved the passing ability of SCC mixtures.

4.1.2 Predictive Analysis of ML Algorithms on Fresh Properties

The effectiveness of the generated models was evaluated by their ability to quantify the slump flow and HSSCC J-ring assessment. In addition, a well-trained model can generate accurate predictions of slump flow and J-ring values for the input data. The intricate relationships between input parameters and the corresponding slump flow and J-ring values were effectively captured in the models by employing ML methodologies. This approach enhances the accuracy and reliability of concrete quality assessment by allowing for more exact quantification of slump and J-ring flow measurement. By carefully assessing the model's performance, cross-validation, and regularization techniques also aid in avoiding overfitting. Fig. 5a, b represents the outcomes of a regression evaluation done on experimentation data and predictive data of slump flow and j-ring analyses utilizing single-based machine learning algorithms like RF and GB and ensemble models employing this approach. Fig. 5a displays the R² values for the RF, GB, and ensemble models applied to estimate the slump flow values, which were 0.955, 0.992, and 0.985 correspondingly, indicating that the prediction results of current models had a satisfactory performance in predicting the respective experimental findings. As a result, these suggested models for slump flow have excellent correlation coefficients. According to Azimi-Pour et al. (2020), large-volume fly ash substitution in SCC resulted in a nearly equivalent coefficient of regression value (R² = 0.974) in support vector machine (SVM) analysis.

Fig. 5

Correlation between experimental and ML predicted outcomes on (a) slump flow and (b) J-ring assessment

However, Fig. 5b displays the impressive correlation coefficient achieved by the ML-based method when implemented in the prediction of J-ring evaluations. Additionally, most of the data points are clustered along the line of best fit this time around. However, the regression coefficients for the RF, GB, and their ensemble models are all approximately 0.977, 0.987, and 0.981. This means the proposed models have shown the best accuracy with the current investigation.

4.2 Effects of the Hardened Characteristics of HSSCC

4.2.1 Compressive Strength Test

According to different statistical methods, the findings of the compressive strength test are examined in Table 5 in terms of mean strength, standard deviation, coefficient of variation (COV), standard error, and the lower and higher range at a 95% confidence interval. Within a confidence interval of 95% ranging from 30.62 to 32.66 MPa, the minimum compressive strength measured was 31.64 MPa after 7 days of curing. Maximum compressive strength was obtained at 35.17 MPa, with a confidence rate of 95% ranging from 34.71 to 35.64 MPa. The minimum compressive strength seemed 45.21 MPa after 28 days, with 95% confidence varying from 43.33 to 47.09 MPa. A 56-day experiment displayed a minimum amount of compressive strength regarding 46.63 MPa with a confidence rate of 95% ranging from 45.41 to 47.45 MPa. Maximum strength was obtained at 53.57 MPa after 56 days, with a 95% confidence interval of 48.89–58.25 MPa. The COV in the sample data was also rather wide, ranging from 0.81 to 6.31%. The maximum variance was 3.38, and the standard error was 2.39 while assessing compression.

Table 5 Statistical analysis of the compressive strength test outcomes

Fig. 6 illustrates the compressive strength and percentages of strength variation among SCC mixtures at 7, 28, and 56 days with changing percentages of MP and constant percentages of SF and FA substitutes. It was demonstrated that the SCC blends with up to 15% MP and constant 10% FA and 20% SF exhibited superior compressive strengths than the control concrete. In several mixes, the greatest compressive strength was obtained when 10% MP, 10% FA, and 20% SF were replaced for SCC. The S20F10M10 mix was observed to have an optimal compressive strength of 35.17 MPa, 50.40 MPa, and 53.57 MPa, superior to reference mixes by 10.39, 10.92, and 14.68% for 7, 28 days, and 56 days, respectively. Better C–S–H gel formation and increased compressive strength are the outcomes of adding MP, which also increases the stiffness and decreases permeability inside the concrete form together with FA and SF. The lowest compressive strength was observed in the S20F10M20 blend, which consisted of 20% MP, 10% FA constant, and 20% SF. Its compressive strength decreased by 0.70, 0.51, and 0.59% at 7, 28, and 56 days, respectively, in comparison to the reference mix (S20F0M0). As MP concentration increases, interparticle contact and homogeneity may reduce, causing reduced strength. According to Khodabakhshian et al. (2018) and Uysal and Yilmaz (2011), its pore-filling influence was responsible for the enhanced compressive strength of mixtures when MP was used to replace cement up to 10%.

Fig. 6

Hardened properties of compressive strength test and the variation of percentage on HSSCC mixtures

4.2.1.1 Predictive Performance of Bagging and Boosting ML Algorithms on Compression

When utilizing ML techniques, such as random forest, gradient boosting, and their ensemble models to estimate concrete compressive strength from a dataset containing 230 experimental and literature review samples, each data point contains specific characteristics representing the composition of the concrete mixture as well as the corresponding strength measurement. The aforementioned characteristics include numerical values for a range of constituent substances, including fine aggregate, coarse aggregate, cement, water, and possible admixtures incorporated into the concrete composition. Further, qualitative variables are frequently incorporated as supplementary characteristics, including curing conditions, the age of the concrete at the time of testing, and the environmental conditions that were prevalent during the testing process. Every data point in the dataset reflects a unique mix of these attributes as well as the measured concrete compressive strength, which serves as the target variable for the ML models. ML algorithms analyze these data points in great detail, finding complex relationships between the input variables and concrete strength. Fig. 7 depicts the boxplots of the experimental and predicted correlation. Mean and standard deviation are shown, along with other statistical measures, such as median, interquartile range, minimum, and maximum. The methods are assessed and contrasted to determine the best prediction strategy that generates enhanced outcomes with excellent accuracy. The results revealed that the distribution of the predicted and experimental outcomes for no model was significantly skewed. All median lines were within their respective boxes, suggesting that there was likely no distinction across the datasets. Furthermore, no outliers were identified in either the RF or GB models.

Fig. 7

Box plot exhibiting the data distribution for compressive strength with machine learning algorithms

Fig. 8 illustrates that the RF and GB approaches provide accurate predictions when contrasted with the experimental outcomes and literature review datasets. Although it provides valuable insights, the established correlation between concrete properties and predictive models has deficiencies. These include possible discrepancies in forecasting precision due to dataset qualities, issues associated with overfitting, difficulties in capturing all the variables of concrete behavior, lack of generalization to diverse situations, and the requirement for cautious validation and interpretation of outcomes due to inherent biases in ML techniques. To attain precise predictions while preventing overfitting, it is imperative to conduct thorough preprocessing of the dataset, which encompasses addressing missing data and normalizing features. By dividing the dataset into training and testing sets, precise model evaluation is facilitated, with the assistance of cross-validation methods. Consistently evaluating the performance of the model and adjusting hyperparameters are essential to derive precise predictions from the dataset without inadvertently overfitting. In the RF and GB models, the correlation coefficient was 0.93 and 0.97, respectively, while ensemble models generated a value of 0.94. However, it is clear from the prediction graph, where the values are clustered along the prediction line, that the predicted outcomes for the model run with GB were rather close to the actual statistics. Using bagging analysis, Azimi-Pour et al. (2020) found that substituting significant amounts of SCC with fly ash had a pretty comparable coefficient of regression value.

Fig. 8

Correlation between experimental and predicted compressive strength outcomes of bagging and boosting models

From Table 6, it is clear that the GB model has a much higher R² value compared to the RF and ensemble model. As a result, it may be inferred that the GB model is highly predictive. In terms of MAE, MSE, and RMSE the GB model performed the most effectively, with values of 1.76, 4.87, and 2.21, respectively.

Table 6 Performance criteria of the developed bagging and boosting methods

4.2.2 Splitting Tensile Strength Test

Table 7 reflects the splitting strength results of the various MP proportions combined with a constant percentage of FA and SF components throughout all SCC samples, based on statistical evaluations such as compressive strength tests. The findings of the samples’ deviation ranged from 0.0424 to 0.495, with a COV of 0.84–11.22% and a standard error of 0.03–0.35.

Table 7 Statistical analysis of the splitting tensile strength test outcomes

Fig. 9 illustrates the outcomes of tensile strength splitting and percentage variation off SCC mixtures with the addition of MP, FA, and SF integration during the curing ages of 7, 28, and 56 days. The progression of splitting tensile strength seems comparable to the pattern of compressive strength, where strength decreases as the MP percentage increases. Nevertheless, replacing cement with MP combined with constant FA and SF typically increased splitting tensile strength by up to 15% in contrast to the reference mixture. As shown, this cementitious material decreases strength when used excessively. S20F10M20 had 17.32, 15.19, and 15.59% greater splitting tensile strengths than the control mix at 7, 28, and 56 days. As a result of lessening the concrete's porosity and increasing its packing densities, MP can increase the overall splitting tensile strength of the concrete. The S20F10M20 mix displayed the lowest splitting tensile strength, with declines of 2.36, 2.27, and 6.62% in the curing times of 7, 28, and 56 days in comparison to the reference mixture. This is because a higher concentration of MP results in more pore space and weaker interparticle bonds, all of which have lower strength. Khodabakhshian et al. (2018) in TVC and Sharma and Khan (2017b) in SCC revealed nearly parallel test outcomes. Comparable to the current investigation, Ali et al. (2022) found that combining polypropylene fibers with marble powder in SCC increased the split tensile strength by 16.92%.

Fig. 9

Hardened properties of splitting tensile strength test and the variation of percentage on HSSCC mixtures

4.2.2.1 Predictive Performance of Bagging and Boosting ML Algorithms on Tension

In employing ML methods like RF, GB, and ensemble models to forecast the splitting tensile strength of concrete from a dataset comprising 230 experimental and literature review datasets, each data point encapsulates distinct attributes reflecting the composition of the concrete mixture and the associated splitting tensile strength measurement. Each experimental sample in the dataset has a different mix of these characteristics in addition to the measured concrete splitting tensile strength, which is used as the dependent variable in ML models. Boxplot representations of the experimental and predicted results are presented in Fig. 10. None of the models had extremely erroneous distributions of predicted and experimental findings. Since each median line was located inside its box, this implies that there most likely was no significant difference between the datasets. Both the RF and GB models did not reveal any outliers.

Fig. 10

Box plot exhibiting the data distribution for splitting tensile strength with machine learning algorithms

Fig. 11 reveals that compared to experimental results, both the RF and GB methods seem to produce suited predictions. Potential discrepancies in forecasts caused by the quality of the dataset, concerns about overfitting, restricted applicability, and the necessity for careful validation owing to inherent biases in machine learning are all limitations of the correlation between concrete properties and predictive models. For accurate predictions that prevent overfitting, it is vital to perform exhaustive preprocessing of the dataset, which includes restoring features and addressing missing data. With the use of cross-validation methods, separating the dataset into a training set and a testing set allows for a more accurate model assessment. Predicted results for the model conducted with GB were very close to the actual outcomes, as seen by the prediction graph, in which the values are organized near the prediction line. de-Prado-Gil et al. (2022) observed that the correlation coefficient of the regression analysis value of the SCC generated with recycled aggregates had a highly comparable value when applying extreme gradient boosting (XG Boost).

Fig. 11

Correlation between experimental and predicted splitting tensile strength outcomes of bagging and boosting models

Table 8 demonstrates that the GB model outperforms the RF and ensemble models regarding R² value. This suggests that the GB approach has a significant level of predictive ability. The GB model fared the best, with values of 0.21, 0.08, and 0.29 for MAE, MSE, and RMSE. The algorithms of the GB framework are responsible for its exceptional performance.

Table 8 Performance criteria of the developed bagging and boosting methods

4.3 Effects of the Durability Characteristics of HSSCC

4.3.1 Water Permeability Test

The water penetration depth of several concrete specimens is illustrated in Fig. 12 after 56 days of the curing age. The specimen with 10% MP, 10% FA, and 20% SF had the lowest water penetration depth of all the mixtures at approximately 14 mm, with a decrease of 39.13% from the control mixture. As a result, the water permeability might vary depending on the SF, FA, and MP combination. Including MP significantly lowers the water penetration depth by up to 15% of the replacement. The addition of MP, FA, and SF to concrete enhanced packing density, which improved durability by filling internal voids between fine aggregates and decreased water penetration depth. The assessment made after water is continuously circulated through concrete specimens of MP for 72 h is satisfactory for construction activities in high-humidity, moist, or high-atmospheric-pressure regions. Gesoğlu et al. (2009) noticed that the quaternary blend of SF, FA, and blast furnace slag (BFS), as well as the binary cement containing SF, had the lowest water penetration depths in SCC compared to the control mixture, thus being nearly the same as this proceeding research. The possibility of a water retention access may cause several kinds of durability concerns, such as freeze–thaw damage, corrosion of structural strengthening, and chemical violence is minimized when the depth of water penetration is kept lower than 25 mm. Due to the depth of water penetration having to exceed 25 mm during the permeability test, it is concluded that the specimens have become weakened (Chung et al., 2013).

Fig. 12

Water penetration depth of the HSSCC mixtures

4.3.2 Sorptivity Test

The findings of the water absorption of the HSSCC at 56 days of curing are shown in Fig. 13. Nearly identical patterns could be observed in all sorptivity test outcomes. After 56 days, adding MP, FA, and SF drastically lowered the absorption rate. The S20F10M20 mixture had a maximum water sorptivity over the 56-day curing period. In a sorptivity test, adding MP to SCC mixes at 5% and 10%, along with consistent amounts of 20% SF and 10% FA, reduced water absorption by 3.57% and 14.29%, respectively; however, it enhanced 14.28% and 42.85% when 15% and 20% MP were incorporated, compared to the control mix. HSSCC likely had better durability in the sorptivity test than any other SCC mixes due to using 10% MP, constant 10% FA, and 20% SF. The incorporation of this SCM reduces the sorptivity coefficient, which in turn causes the concrete to absorb less water, resulting in increased resistance to water penetration and improved durability. Furthermore, incorporating MP in higher concentrations improves concrete porosity, allowing additional paths for water to penetrate and enhancing the sorptivity coefficient. Sharma and Khan (2017a) observed that SCC incorporating copper slag exhibited a substantial decrease in water absorption, with sorptivity readings reduced by up to 60%, resulting in significantly enhanced impermeability compared to this study's results. According to Md (2019), steel fiber-reinforced SCC (SFRSCC) cumulative water absorption increased as time progressed and also displayed equivalent results compared to this investigation.

Fig. 13

Sorptivity test of the HSSCC mixtures on (a) capillary absorption (b) initial absorption and (c) summarized coefficients at 56 days

4.3.3 Rapid Chloride Penetration Test (RCPT)

Fig. 14 illustrates the findings of chloride penetration tests for all SCC combinations regarding the amount of total charge passed after curing for 56 days. The investigation found that incorporating 0–15% MP along with steady amounts of 10% and 20% FA and SF in the SCC mixture led to a decrease in chloride penetration, in contrast to incorporating 20% MP in addition to the same amounts FA and SF, which increased chloride penetration in comparison with the control mix. The RCPT values reduced by 5.75, 22.41, 37.36, and 20.69% when MP was replaced at 0, 5, 10, and 15% together with a constant 10% FA and 20% SF, whereas MP at 20% replacement indicated an increase of 2.30% instead of the control mixture. The pozzolanic activity of these substances may enhance the durability of concrete by causing the formation of additional C–S–H as well as the generation of pores in the concrete matrix. On the contrary, a higher concentration of MP may also accelerate the production of additional reaction products, making it easier for chloride ions to circulate and enhance the RCPT values. The graph demonstrates that the mix S20F10M10 had the lowest charge travelling through it at 981 Coulombs, exhibiting that it was more durable than other SCC mixtures. Following the ASTM C1202 standards, the passing charges were low to very low, which might further reduce the concrete's porosity (ASTM-C, 1202, 2019). Givi et al. (2010) stated that the high concentration of SiO₂ in rice husk ash (RHA) established an extremely reactive pozzolanic substance used to formulate more C–S–H gel, enhancing concrete formation.

Fig. 14

RCPT test of the HSSCC mixture

4.4 Effects of the Non-Destructive (NDT) Characteristics of HSSCC

4.4.1 Rebound Hammer Test

The statistical outcomes of the digital rebound hammer assessment analysis are summarized in Table 9, where the mean rebound number (RN) was found to vary from 22.66 to 32.20. With a COV that ranged from 0.0108 to 0.0458, the standard deviation for all values fluctuated between 0.3253 and 1.1314.

Table 9 Statistical analysis of the rebound number test results

The mean rebound number (RN), which was determined through analysis of digital rebound hammer assessment, can be depicted in Fig. 15. The test was performed on a specimen curing for 28 and 56 days. The SCC surface layer is regarded as a “good layer” up to a 15% substitution of the MP, coupled with a constant 10% FA and 20% SF, respectively. Despite the porosity of the ITZ within the concrete samples, it can be observed that the RN value of the control was lower than that of all other SCC mixes except the S20F10M20 mix. The S20F10M10 mix had the greatest RN at 28 days, which is 31.9, and at 56 days, which is 32.2, whereas the S20F10M20 mix had the lowest RN at 22.6 and 26.3 at the curing times of 28 days and 56 days, respectively as compared to the reference mix. According to Singh et al. (2017), the outcomes of this experiment were extremely comparable with the findings of the rebound hammer test of waste marble powder substitutes.

Fig. 15

Rebound hammer test of HSSCC specimens

4.4.2 Ultrasonic Pulse Velocity (UPV) Test

The statistical results for flexural strength for SCC specimens containing different MP content percentages and consistent percentages of FA and SF are exhibited in Table 10. With a COV between 0.00429 and 0.01692, the standard deviation for UPV values encompassed 19.799 to 70.711.

Table 10 Statistical analysis of the UPV test results

As depicted in Fig. 16, the specimen underwent UPV testing at 28 and 56 days with varying MP percentages alongside consistent FA and SF percentages. In comparison, the variable concentration of MP alongside the constant concentration of FA and SF exhibited UPV values fluctuating between 4136 and 4929 m/sec at 28, and 56 days. A control mix, S20F0M0, demonstrated an impressive UPV value of 4065 m/sec after 28 days and 4580 m/sec after 56 days, indicating that pulse velocity enhances with the duration of more curing ages and enhances the grading quality of concrete “excellent” following ASTM C597 criteria (Astm). The optimum UPV value was noticed in the S20F10M10 mix, whereas the pulse velocity was 4285 m/sec and 4929 m/sec for the curing periods of 28 and 56 days, respectively. In contrast to the control mix, the pulse velocity values in a mixture including 5–20% replacements of MP along with constant 10% FA and 20% SF increased by 0.65, 5.61, 7.62, 6.70, and 2.21% during the 56-day curing period. As a result, such an investigation demonstrated that a higher proportion of replacement resulted in a denser structure and enhanced concrete density, resulting in a higher pulse velocity. Uysal and Yilmaz (2011) reported an equivalent outcome using marble slurry powder (MSP). According to their findings, MSP using up to 10% enhanced UPV values, whereas higher MSP substitution with cement caused lower UPV values. Uysal and Sumer (2011) also found comparable outcomes with nearly equivalent pulse velocities achieved by incorporating FA into SCC mixtures as a 25% cement replacement.

Fig. 16

UPV test of HSSCC specimens

4.5 Relationship between DT and NDT Properties of HSSCC

4.5.1 Compressive Strength Versus Rebound Number (RN)

This investigation additionally recommends one variable in the equation for regression for determining the compressive strength of concrete according to several regression analyses. The correlation associated with compressive strength (f'c) and rebound number (RN) can be illustrated in Fig. 17. A modest percentage of variability and high stability of the findings can be observed by the coefficient of regression (R²) value, which is approximately 82.67%. The graphical representations of the points were acquired from multiple investigations on rebound number and compressive strength from Sunayana and Barai (2017), Datta et al. (2022), Domingo and Hirose (2009), Singh et al. (2017), and De Almeida (1991). The association with the destructive strength and rebound number value of the SCC specimens, which include varying percentages of MP alongside consistent percentages of FA and SF, is likewise expressed in this equation. Furthermore, the predicted interval and 95% confidence interval function are determined to illustrate the observed relations precisely. Nevertheless, because of varying mixes of concrete, Sunayana and Barai (2017), Datta et al. (2022), Domingo and Hirose (2009), and Singh et al. (2017) displayed variations beyond the confidence interval of 95% level region. According to the findings of De Almeida (1991) and the outcomes of this investigation, the expected compressive strength deviates by an average of almost 3%.

Fig. 17

Correlation between the compressive strength and rebound number of HSSCC

4.5.2 Compressive Strength Versus Ultrasonic Pulse Velocity (UPV)

The correlation between the compressive strength and UPV of each combination is represented in Fig. 18. Relatively poor results can be observed by the coefficient of regression (R²) value, which is about 78.73%. The dispersion of every data series from the typical trend line is highlighted. The substitution of varying MP percentages was mainly responsible for the dispersion of the data points. In this scenario, the parameters impacting validations tend to be so bound that it might be complicated to accomplish a confidence level of 95% intervals higher than 20% during precise strength anticipation of in situ concrete, regardless of optimal circumstances and a particular validation (Madandoust et al., 2010). The graphical points were extracted through multiple studies on UPV and compressive strength by Belagraa et al. (2015), Sunayana and Barai (2017), Vo et al. (2021), Datta et al. (2022), and Singh et al. (2017). However, Datta et al. (2022), and Vo et al. (2021) exhibit variations beyond the 95% confidence level region as an outcome of multiple concrete mixtures. The analyses carried out by Singh et al. (2017) and the present investigation revealed differences by an average of almost 19.03%. On the contrary, this investigation and the findings of Belagraa et al. (2015) demonstrated a nearly 4% average divergence from the expected compressive strength.

Fig. 18

Correlation between the compressive strength and UPV of HSSCC

4.6 Effects of the Microstructural Characteristics of HSSCC

Fig. 19a–d displays SEM micrographs of certain HSSCC mixtures whereas Fig. 19a illustrates the surface morphology of the control mixture. The control mix has noticeable surface fractures and poor interfacial transition zone (ITZ). As opposed to the control mixture, the HSSCC mixture having MP, SF, and FA had enhanced ITZ and a narrower crack width, indicating better mechanical performance. Fig. 19b illustrates an observation of micropores on the surface of the S20F10M0 mixture. SEM micrographs of HSSCC mixtures containing MP revealed a higher ITZ and the absence of any observable crack. However, mechanical performance declined as a consequence of OPC dilution caused by increased MP substitution. HSSCC mix S20F10M10 exhibited outstanding surface morphology, which is apparent in Fig. 19c. The superior mechanical performance may be attributed to the compact concrete matrix with fewer pores in the HSSCC mix consisting of 10% MP, 20% SF, and 10% FA. Fig. 19d displays the SEM micrograph of mix S20F10M20. In this micrograph, reduced ITZ and minimal visible fractures were observed, with the S20F10M20 combination appearing to exhibit an even more porous and less dense structure compared to other HSSCC combinations. Singh et al. (2017) observed findings for MP substitution that were pretty comparable to the findings of the present investigation.

Fig. 19

SEM images of (a) S20F0M0 (b) S20F10M0 (c) S20F10M10 and (d) S20F10M20

4.7 Environmental Impact Evaluation

OPC generates from 75 to 90% of the global CO₂ emission that concrete produces since it is the main binder component of concrete (Khair et al., 2024; Rahman et al., 2022). Alternative binders have been developed to lessen the dependency on OPC due to rising concerns about its potential for worldwide warming (kg CO₂-eq/kg material). Although using SCMs reduces the overall embedded greenhouse gas (GHG) emissions of building materials, the embedded GHG of various types of waste differs (S. Rahman et al., 2023a, 2023b). MP, SF, and FA were partly used as binding agents alternative to OPC for the current investigation. This research assesses the entire emissions of CO₂, taking into account the comparable CO₂ exposures for every material to emphasize the implications of FA, MP, and SF upon the entire embodied GHG emissions. Table 11 compiles the carbon content embodied data for every mixture of concrete after analyzing the emissions of CO₂ of every ingredient per 1m³ of concrete and summing those numbers together employing Eq. (5).

$${\text{CO}}_{{{\text{2Concrete}}}} \, = \,\sum\nolimits_{i = 1}^{n} {\left( {{\text{CO}}_{{\text{2i - e}}} \, \times \,\,{\text{W}}_{{\text{i}}} } \right)}$$

(5)

Table 11 Estimated CO₂ emissions from concrete

Fig. 20 reveals that overall CO₂ emissions were lowered when OPC was partially replaced with different percentages of MP alongside a steady percentage of FA and SF. The control concrete S20F0M0 mixture produced 455.90 kg CO₂/m³ using a more significant cement proportion. Cement produces roughly 89% embodied carbon. As contrasted with the reference mixture, the total amount of carbon embodied dropped by 10.67, 15.68, 20.69, 25.70, and 30.70% when OPC partially substituted MP at 0, 5, 10, 15, and 20% with a constant proportion of 20% SF and 10% FA. Despite the environmental benefits, the significant reduction in compressive strength resulted from greater use of MP and mineral admixtures. In such a scenario, the HSSCC mixture S20F10M10 offers the advantage of being used as an improved mechanically sound and environmentally friendly concrete. This study closely aligns with Khodabakhshian et al. (2018), indicating a notable reduction in global warming potential (GWP) indices with escalating incorporation ratios of MWP and SF.

Fig. 20

The entire CO₂ emissions and possible global warming reduction percentage

5 Conclusion

High-strength self-compacting concrete (HSSCC) may greatly benefit from the partial substitution of cement with marble powder (MP), silica fume (SF), and fly ash (FA). Based on the findings of the experiment, the following inferences may be concluded:

• Combining SF, FA, and MP facilitates the development of HSSCC. The collective use of these SCMs was highlighted to enhance HSSCC’s workability, leading to higher flowability and filling ability in the context of fresh characteristics. Evidence also revealed that SCM additions reduced concrete water requirements, resulting in enhanced density.

• Regarding durability characteristics, including MP, SF, and FA significantly reduced the water penetration depth and absorption rate. In the RCPT test, the charge that was transmitted on the concrete was in the low to very low range, which could have further reduced its porosity.

• S20F10M10 demonstrated increases in compressive strength of 10.39, 10.92, and 14.68% on days 7, 28, and 56, respectively, in comparison to the control mix. In terms of splitting tensile strength, it performed 17.32, 15.19, and 15.59% better than the control mix, respectively.

• The ML approach demonstrated the superiority of RF, GB, and their ensemble models in predicting fresh and mechanical characteristics. While GB and ensemble models have shown more reliability than RF algorithms across various statistical evaluations, RF algorithms remain viable for forecasting concrete strength.

• The replacement of 10% MP to a steady 10% FA and 20% SF resulted in UPV values above 4.5 km/s and rebound index values between 30 and 40, which indicates excellent quality and substantial compressive strength.

• The SEM images of HSSCC generated with a combination of 10% MP, 20% SF, and 10% FA revealed negligible voids, superior aggregate packing, more effective cement paste, and a dense concrete matrix.

• The HSSCC mixture S20F10M10 reduced CO₂ emissions without sacrificing mechanical efficiency.