Optimization of recycled rubber self-compacting concrete: Experimental findings and machine learning-based evaluation

Abstract

This research aims to assess the rheological and mechanical characteristics of Self-compacting concrete (SCC) incorporating waste tire rubber aggregates (WRTA) as an interim substitute for coarse aggregates. However, the standard experimental modeling approach has significant obstacles when it comes to overcoming the nonlinearity and environmental susceptibility of concrete parts. Therefore, linear regression (LR) and extreme gradient boosting (XGBoost) were used as two standard single machine learning (ML) models to predict the aforementioned rubberized SCC features. In this study, conventional coarse aggregates were supplanted with WRTA at 0%, 5%, 10%, and 20% to uncover the optimal proportion of coarse aggregates substituting rubber. To find the optimum amount of WRTA to use as a substitute, the study follows the impacts of rubber on the self-compacting rubberized concrete's (SCRC) rheological and mechanical characteristics. The consequences on fresh properties were investigated by the slump flow, J-ring, and V-funnel tests, while compressive and splitting tensile strengths tests were conducted to assess mechanical properties. Increasing WRTA test outputs indicated a deterioration in workability and hardened qualities. While a 10% swapping ratio is deemed feasible for producing SCRC, optimal results were achieved by reducing environmental impacts and efficiently managing a significant volume of rubber tire waste with a 5% substitution of rubber within the coarse aggregates. The research findings indicated a noticeable decrease in fresh properties as the WRTA content increased. Notably, after 28 days, a 10% WRTA substitution led to a 34% reduction in compressive strength and a 28% decrease in splitting tensile strength, satisfying ACI standards. Furthermore, XGBoost demonstrated superior predictive performance with the highest R² values, outperforming the LR model and affirming its efficacy in delivering more accurate predictions.

1. Introduction

Concrete is a product of cementing medium, which is a product of hydraulic cement and water reaction [1]. Researchers aimed to create an innovative mixture from the combination of the positive effects of self-compacting concrete (SCC) and rubberized concrete (RC) Self-compacting rubberized concrete (SCRC). SCRC has many potential benefits to the construction sector owing to its lightweight and ductile traits. Thomas and Gupta [2] observed that with chemical admixtures, workable RC can be prepared without the need to increase water quantity. RC can be prepared or cast in a traditional method of conventional concrete without any additional procedure [[3], [4], [5]]. Previous research observed Significant differences in rubber replacement SCRC and conventional rubberized concrete [6]. For both cases, a declination in workability and mechanical strength was observed. However, rubber aggregates develop a fragile connection with the surrounding binder matrix resulting in a weak interfacial transition zone. As rubber aggregates react more than the binding matrix, cracks are generated within this zone [7]. For conventional concrete, the replenishment of coarse aggregates with rubber aggregates showed a more significant decrease in compressive strength, about 61% for 75% replacement, as opposed to the displacement of fine aggregates [3]. Similarly, adding rubber to conventional concrete has a detrimental effect on workability and mechanical strength [8,9].

Several research studies on supplementing fine aggregates with rubber aggregates in self-compacting concrete have illustrated a significant decrease in mechanical properties [[10], [11], [12], [13], [14]]. Werdine et al. [15] found a deterioration in the compressive strength with an upsurge in rubber granulometry. However, silica fume (SF) is considered one of the best supplementary cementing materials (SCM). Significant increases in mechanical properties were observed after incorporating silica fume in concrete because of its high pozzolanic reactivity [[16], [17], [18]]. Chu et al. [19] replaced 15% fine aggregates with crumb rubber to determine the cyclic characteristics of beam-column joints. The fatigue capacity of SCC with the insertion of rubber aggregates increased by 15% [20]. Rahman et al. [21] observed a reduction of dynamic modulus up to 10%–20% with added rubber aggregates. In contrast, by observing the chlorine penetration, a progressive increase was seen with the addition of rubber aggregates and increased water absorption because of the presence of crumb rubber [22]. Similar results for water absorption were found by Thomas and Gupta [2]. With a difference in ultimate bond strength of just 5%, crumb rubber concrete (CRC) specimens performed similarly to conventional concrete (CC) specimens in the elemental tests, according to Yi et al. [23]. Hassanli et al. [24] claimed that using rubber in unconfined and non-prestressed concrete reduces its strength; nevertheless, when lateral prestress and fiber-reinforced polymer (FRP) confinement are coupled, CRC may significantly boost its axial stiffness and strength. Abd-Elaal et al. [25] examined the efficacy of treating rubber aggregates of various diameters (0.425 mm–5.0 mm) and reported that this procedure improved the compressive strength by as much as 60.3% when relatively small rubber particles were utilized.

Replacement by rubber aggregates also makes the concrete less dense because of the rubbers' low density [8]. Significantly improved dynamic loading behavior, impact resistance, and fracture toughness were observed for SCRC [[26], [27], [28]]. For a lower proportion of replacement with rubber at concrete, NaOH and KMnO₄-treated crumb rubber improved significantly, while 40% replacement with cement-treated crumb rubber enhanced the compressive and flexure strength by 64% and 33% [29]. Amiri et al. [30] replaced the cement with waste rubber powder and aggregates with recycled aggregates to evaluate the properties. Upon application of Herschel-Buckley and modified Bingham models, shear thickening behavior was observed, with the addition of higher torque for replacement with rubber aggregates [31]. Furthermore, it initiates progressed thermal and acoustic insulation in homes containing small equipment in addition to floors of buildings [32]. Yung et al. [12] observed a 17% increase in surface resistance by adding rubber aggregates passing through the #30 sieve, indicating high electrical resistance properties. By contrast, Záleská et al. [33] observed reduced volumetric heat capacity and thermal conductivity due to rubbers’ lower heat capacity. Youssf et al. [34] found that heat-treated rubber with CRC for 2 h at 200 °C and then treated with 100% magnetized water (MW) for 24 h showed promising performance. With a rubber component of 40%, these circumstances led to a 74% recovery of compressive strength and an increase in impact resistance of 2.2 times at first fracture and 92% at final failure. Su and Wu [35] demonstrated that fly-ash-containing concrete treated with MW had a 15%–25% improvement in compressive strength. Eltayeb et al. [36] applied flexural stresses to profiled steel composite walls filled with FC and polypropylene fibers. The findings demonstrated that the flexural capacity in FC was enhanced by the use of fibers.

Several researchers have used computer-aided forecasting systems to anticipate the attributes of the composite frame and substances, including multi-scale technique regressions and machine learning (ML) techniques [[37], [38], [39], [40]]. Recently, self-compacting recycled aggregate concrete (SRAC) was investigated by de-Prado-Gil et al. [41], who used four ML techniques comprising extreme gradient boosting (XGBoost), gradient boosting (GB), categorical gradient boosting (CatBoost), and extra trees regression (ExtraTrees), with 381 assortments of data for assessing the splitting tensile strength. Compared to other models, they concluded that XGBoost performed the best, with an R² value of 0.84. In another study, Mohammadi Golafshani et al. [42] used two machine learning (ML) models, the M5P model tree and the multi-gene expression programming (MGEP) to estimate the compressive strength of rubberized concrete (RC) using an ensemble of 712 samples.

Even though a substantial amount of research has been initiated experimentally to analyze the various fresh and mechanical properties of SCRC, there was not a single article that could be found in the available literature that developed machine learning models (XGBoost and LR) for predicting the fresh and mechanical properties of rubberized concrete. In this body of work, prior ML algorithms were only concerned with forecasting the individual properties like compressive and splitting tensile strength in recycled aggregate concrete and rubberized concrete. Therefore, this investigation aims to use two explainable ML techniques to make predictions about the fresh and mechanical properties of concrete made using discarded tire rubber. Furthermore, this research involves reusing discarded tire rubber, which helps mitigate environmental consequences via waste management. Throughout the investigation, the effect of rubber on the rheological and mechanical properties of SCC with the replacement of coarse particles with waste rubber tire aggregates was tracked, and the optimal quantity of rubber aggregates in Self-compacting rubberized concrete was identified. Thus, by adopting the proposed new ML models, researchers in the future should be able to streamline the design of concrete mixtures containing rubberized recycled aggregate to attain the appropriate mechanical properties that might be a cost-saving approach in terms of materials, labor, and without using heavy structural machinery for experimentation.

2. Experimental design

2.1. Raw materials

Ordinary Portland Cement (OPC) was used as the main binder for the concrete mixtures, which conformed to the ASTM C150 Type I, having a specific surface area of 388 m²/kg. OPC was partially replaced with silica fume with a proportion of 7.5% of the OPC's weight to achieve high strength and durability. The chemical constituents of OPC and silica fume are described in Table 1.

Table 1.

Chemical composition of ordinary portland cement and silica fume.

Ingredients	OPC (% of ingredients)	Silica Fume (% of ingredients)
CaO	65.57	2.63
SiO2	19.01	89.94
Al2O3	5.29	0.51
Fe2O3	3.38	0.65
MgO	2.06	1.13
NaO2	0.87	0.21
K2O	0.63	0.47
SO3	2.14	0.17
LOI	1.24	4.36

The locally available pit sand known as Sylhet sand, which is free from organic or vegetable matter and chemically inert, was used as fine aggregates. The fineness modulus was found to be 2.69 for fine aggregates. The gradation curves for fine, coarse, and rubber aggregates are illustrated in Fig. 1. On the contrary, black stone chips retained through an ASTM standard sieve size of 19 mm were used as a coarse aggregate for all concrete batches. Numerous tests were conducted based on ASTM C127 [43] and ASTM C128 [44] to determine the physical properties of fine and coarse aggregates, such as specific gravity, bulk density, water absorption, etc., as shown in Table 2.

Fig. 1.

Gradation curve for fine aggregate.

Table 2.

Material properties of fine and coarse aggregates.

Properties	Fine Aggregate	Coarse Aggregate	Rubber
Loose unit weight (Kg/m3)	1391.81	1493	763
Compacted unit weight (Kg/m3)	1564	1671.72	771
Bulk specific gravity	2.36	2.72	1.13
Bulk SSD specific gravity	2.40	2.75	1.15
Apparent specific gravity	2.46	2.81	1.18
Absorption (%)	1.62	1.18	1.78

Waster tires were collected from local garages and pits and cut into pieces through a time-consuming and laborious effort [45]. The maximum size of these aggregates was found to be 19 mm. Moreover, these waste rubber tire aggregates (WRTA) were partially replaced with natural coarse aggregates. Moreover, Superplasticizer (SP) was used to emphasize the flowability of SCC. It is a polycarboxylate-based ether. Fresh potable water was used for all concrete mixes.

2.2. Mix proportion and concrete production

Overall, a total of four concrete mix proportions of SCC were prepared by replacing coarse aggregates with different proportions of WRTA. First, several trial mixtures incorporated with SP and varying water-to-binder ratio addition were conducted to specify the control mixture of this experiment with a ratio of 1: 1.036: 1.76 by the weight of the binder, coarse aggregates, and fine aggregates. A water-to-binder ratio of about 0.39 was used, while the SP content remained at 1% of the total mass of binder content (Cement + Silica Fume). For recognition purposes, the control mixture was defined as Control, where the rest were labeled as SCRC5, SCRC10, and SCRC20, where WRTA replaced the coarse aggregates by 5%, 10%, and 20%. The mixed design for this study is described in Table 3.

Table 3.

Mix proportion for the study.

No.	Mix ID	Cement (kg/m3)	Silica Fume (kg/m3)	Water (kg/m3)	Coarse Aggregates (kg/m3)	Sand (kg/m3)	Rubber (kg/m3)
1.	Control	511.5	38.5	214.5	570	970	0
2.	SCRC5	511.5	38.5	214.5	541.5	970	28.5
3.	SCRC10	511.5	38.5	214.5	513	970	57
4.	SCRC20	511.5	38.5	214.5	456	970	114

2.3. Concrete production, preparation, and curing

An electric-operated concrete mixture was used to construct all the concrete batches with a systematic approach consisting of different steps. All the dry ingredients were put in the mixture after weighing following the requirements and were stirred till a firm mix was prepared. Furthermore, water and SP are gradually mixed with these dry ingredients till flowability is visible. At the final step, WRTA is mixed with 3 min of mixing time.

The overall experiment was divided into two major test matrixes consisting of rheological and mechanical properties of SCC with four different batches of SCRC. First, the slump flow, J-ring, and V funnel tests were conducted on mixed designs to determine the fresh properties. Compressive strength and tensile strength were determined for 7 days and 28 days of curing to investigate the hardened properties. For this sake, cylindrical specimens of 100 mm × 200 mm were used. The molds were first lubricated properly with lubricants during this specimen preparation process. Then, after filling up the molds, they were identified properly for future identification. After 48 h, the specimens were demolded carefully so that no harm was done to their surface. Curing tanks made of plastic were used to cure those specimens for 7 days and 28 days at room temperature (25 °C) for the conduction of mechanical properties tests.

2.4. Testing of SCC and SCRC

2.4.1. Rheological properties

Materials, mix methods, mix composition, testing procedures, and other environmental factors usually influence the rheological properties of concrete. The process is done from the mixing till the molding of the specimens. These tests include the determination of the slump flow, T₅₀₀ time, and V funnel time to measure and compare the flowability of concrete. The blocking index was measured to determine the viscosity and passing ability of SCC. Furthermore, T₅₀₀ time allowed us to visualize the horizontal spreading ability without any obstacles, indicating greater filling capacity. Slump flow tests, V-funnel tests, and J-ring tests were conducted according to EN 12350 [46]. The passing ability and flowability of the concrete were determined by the slump flow test and J-ring test, while the blocking index was observed by conducting the V-funnel test illustrated in Fig. 2 (a), 2 (b), and 2 (c), respectively.

Fig. 2.

Instrumental setup for (a) Slump flow test (b) J-ring test, and (c) V funnel test.

2.4.2. Mechanical properties

On the contrary, after appropriate curing of the cylindrical specimens for 7 days and 28 days compressive strength and splitting tensile strength were determined. Following the ASTM C39/C39 M [47] compressive test was done, while a splitting tensile strength test was carried out following ASTM C496/496 M [48]. Both laboratory test setups are illustrated in Fig. 3(a) and (b). Also, failure patterns and cracks in the specimens were identified through microscopy.

Fig. 3.

Laboratory test setup for (a) Compressive strength test and (b) Splitting tensile strength test.

Every sample was subjected to a three-times-repeated testing procedure to ensure reliable and valid outcomes. It was crucial to repeat this process three times to catch any possible variances or outliers in the data. Finally, each sample's representative and aggregated result was created by averaging the results acquired from each test iteration.

2.4.3. Machine learning methods

Machine learning (ML) is the ability of computers to evaluate information and gain knowledge about complex patterns within the data without being explicitly programmed [49]. Here, Equation (1), (2), (3), (4) measure how well a model fits the data, and these measures are the coefficient of determination (R²), the Root Mean Square Error (RMSE), the Mean Square Error (MSE), and the Mean Absolute Error (MAE) [50]. The following explains the two powerful ML approaches utilized in this research to estimate the properties of SCC made with recycled rubber tires. The process flow employed in the modeling study is outlined in Fig. 4.

$$\text{RMSE}=\sqrt{\frac{1}{n}\sum _{i=1}^n{\left(Y-Y_i\right)}^2}$$ (1)

$$\text{MSE}=\frac{1}{n}\sum _{i=1}^n{\left(Y-Y_i\right)}^2$$ (2)

$$\text{MAE}=\frac{1}{n}\sum _{i=1}^n{\left|Y-Y_i\right|}^2$$ (3)

$${\mathrm{R}}^2=1-\frac{\sum _{i=1}^n{\left(Y-Y_i\right)}^2}{\sum _{i=1}^n{Y_i}^2}$$ (4)

Fig. 4.

Flow chart of the proposed method.

2.4.3.1. Linear regression (LR)

y = w₁x₁ + w₂x₂+ w₃x₃+ ⋯ + w_nx_n + b (5)

Linear regression is an illustrative regression procedure used to model the linear connection between a dependent variable and a set of independent variables. Fast learning, strong explanatory ability, and comparable performance to other algorithms are all hallmarks of linear regression [51]. This technique is useful when there is a moderate degree of correlation between the independent variables, but all of them have a linear connection with the dependent variable. The LR model can be formulated as follows:

Where y is the outcome (depending on the value of x), x is the independent variable (determining the value of y), w is the weight for x, and b is the bias.

2.4.3.2. Extreme gradient boosting (XGBoost)

It is an ensemble learning method based on trees that adhere to the boosting protocol [52]. The idea behind this technique is to modernize what remains by adding new learners to the method, each of which is matched to the leftovers of the preceding learners. Decision trees, in addition to gradient boosting, form the basis of extreme gradient boosting. While the results of only one decision tree are easy to decipher, users of XG can ascertain the weights given to each input variable. It was found by Cui et al. [53] that compared to traditional methods, the XGBoost approach provides more accurate predictions of compressive strength in concrete.

3. Experimental results and discussion

3.1. Rheological properties of SCRC

The outcomes of the rheological properties test are described in Table 4.

Table 4.

Results obtained from the rheological properties test.

Mix	Slump Flow (mm)	T500 time (sec)	J-ring diameter (mm)	Blocking index (mm)	V funnel time (sec)
Control	745	2.29	705	6.25	3.64
SCRC5	720	2.34	700	11.25	4.11
SCRC10	705	2.88	685	11.25	4.55
SCRC20	670	3.21	645	15	4.78

3.1.1. Slump flow and J ring flow

Fig. 5 illustrates the J ring and slump slow rate for all four concrete ix designs. It is seen from Table 4 and Fig. 5 that all slump flows were within the permissible range of 550 mm–750 mm. The slump flow diameter for the control mixture was 745 mm, indicating the control mixture's good flow ability. However, the slump flow diameter gradually decreased with the increments of rubber aggregates' proportion. Around 3.3 %, 5%, and 10 % decrease in slump flow were observed for 5%,10%, and 20% of WRTA compared to the control mixture. A similar type of reduction was observed by Si et al. [54], where the reduction of slump flow for rubberized concrete was about 5%–14%. For 5%–25% replacement with rubber, up to 30% reduction was observed by Güneyisi [55]. On the other hand, the J-ring diameter showed similarities with slump flow diameter reduction. J-ring diameter gradually decreased with the increment of rubber aggregates. The lowest J-ring diameter was for SCRC20, about 645 mm, indicating its least passing ability among all the mix designs. There was a fall of about 0.7%, 2.8%, and 8.5% decrease in J ring diameter compared to the control mix for 5%, 10%, and 20% replacement of coarser aggregates with rubber. Bušić and Miličević [56] conducted a study incorporating 5%–30% waste tire rubber and claimed a decrease in J-ring diameter similar to this study. Aslani et al. [13] observed that all the slump flow diameters for SCRC were within the 550–650 mm limit. On the other hand, similar to this experiment, all the slump flow diameters were greater than 600 mm, according to Bignozzi and Sandrolini [32]. However, it was concluded that rubberized compacting concrete needs more superplasticiser than conventional self-compacting concrete [32]. It was also observed that jagged edges of rubber aggregates acted as a disturbing factor to affect the concrete consistency. On the other hand, rubber replacement more than 20% was not able to make a successful SCC mix [16]. Like previous research, 20% replacement showed the lowest results, which were close to the norm's margin. A comparison between slump flow diameter and J ring showed that the increment of rubber aggregates leads to a bigger difference between slump flow and J-ring diameter [57].

Fig. 5.

Relationship between slump flow diameter and J ring diameter.

3.1.2. V-funnel time and T₅₀₀ time

Fig. 6 describes a comparison between V-funnel and T₅₀₀ time for all different mix designs. It is seen from Fig. 6 and Table 4 that the T₅₀₀ time of different mixtures was very acceptable. All the values are well within the range of 5 s, and as per standard, they were greater than 2 s, satisfying all the requirements. The control mix obtained the least time to reach the 500 mm circle. However, the highest was noticed for the SCRC20 mix, which was around 3.21 s. This is due to the jagged and coarser rubber aggregates. WRTA creates an interlocking structure that affects the flow ability of concrete mixtures, resulting in a longer time to reach a 500 mm circle [13]. On the contrary, all V-funnel tests for different mixtures resulted in the admissible limit of 6–12 s. The control mixture showed good passing ability and took the least time, about 3.64 s, among all the four mix designs. However, the SCRC20 showed the least passing ability and the highest time to clear the v-funnel. It became obvious from the result that the progression ability decreased by increasing the proportion of rubber. There was an increase in V funnel time of about 2.18%, 25.7%, and 40.1% for 5%, 10%, and 20% replacement of rubber, respectively. The extra rubber aggregates precipitated concrete segregation due to the vulnerable bond between WRTA and the cement matrix. In previous research, it was seen that by increasing the proportion of CR from 0% to 20%, about a 95.5% increase in T500 time was observed and similarly, the T50J and V-funnel times were increased by 65% and 110% [16]. With the increasing amount of rubber and decreased slump flow, the T₅₀₀ time varied from 1.15 to 2.36 s [58]. However, because of the use of chemical admixtures, an opposite incident was observed where the T500 time decreased with the increased rubber aggregates, indicating an increase in workability [59]. On the contrary, the incremented proportion of rubber aggregates resulted in increased V-funnel time. Güneyisi [55] investigated rubberized self-compacting concrete with 0%–25% waste tire rubber and disclosed a 60% increment in V-funnel time, which is relevant to the experimental results obtained in this study.

Fig. 6.

Relationship between slump T₅₀₀ time and V-funnel time.

3.1.3. Blocking index and passing ability

The blocking index (B_J) for four different mixtures is illustrated in Fig. 7. It is seen from Table 4 and Fig. 7 that the control mixture showed a blocking index of 6.25 mm and a good passing ability. However, with the increase of WRTA in SCRC, the blocking index increased, representing a reduction in passing ability. Due to the jagged rubber aggregates, there was a small amount of segregation during the J-ring test. The highest blocking index was for SRCR 20, with an amount of 15 mm, which showed the least passing ability through obstacles. In comparison with the T₅₀₀ time, both B_J and T₅₀₀ time increased with the rubber aggregates. Similar to experimental results, Bušić and Miličević [58] observed that the control mixture without rubber showed a blocking index of 6.25 mm exactly, similar to the experiment.

Fig. 7.

Comparison of blocking index and T500 time.

In contrast, for 20% replacement by crumb rubber, it increased to 12 mm, which is close to the experimental result. This can be due to high friction between the WRTA and coarse aggregates. The lower density of WRTA can also act as an attributing factor for this reduction [16]. So, experimental results were like the previously conducted research indicating a reduction in passing ability.

3.1.4. Machine learning methods on fresh properties

The ML models' ability to reliably estimate the slump flow, J-ring flow, V-funnel, and L-box of SCC was used as the criterion to accept or reject the models. Recently, Sonebi et al. [60] conducted with the use of the super vector model produced approximately 98% accuracy in fresh properties of SCC. In this research work, initial findings confirm that the suggested models are beneficial for predicting the fresh features of SCRC. However, Fig. 8, Fig. 9, Fig. 10 showcase the conclusions of the regression study done on practical data as well as forecasts for slump flow, J-ring flow, and V-funnel applying the LR and XGBoost approaches respectively. Fig. 8 highlights the accuracy of the suggested models to quantify the slump flow, where XGBoost (R² = 0.99) and LR (R² = 0.98) both reveal an elevated correlation coefficient. Likewise, Fig. 9 exhibits the comparison between the ML approach's anticipated values for the J-ring flow and the experimentally determined values. As can be observed the LR-based method only managed 0.93 R² accuracy, whereas the XGBoost-based model obtained 0.99 R². It indicates that the XGBoost-based model slightly surpassed the LR-based strategy in terms of accuracy. However, as can be witnessed in Fig. 10, XGBoost is nearly 20% more reliable in predicting V-funnel values than the LR model.

Fig. 8.

Measured versus predicted slump flow.

Fig. 9.

Measured versus predicted J-ring flow.

Fig. 10.

Measured versus predicted V-funnel time.

The XGBoost model is notable for its innovative approach and exceptional predictive capabilities, as illustrated in Fig. 8, Fig. 9, Fig. 10. When dealing with data that has complicated associations, the ensemble learning approach XGBoost outperforms conventional LR models. It works by incrementally merging underperforming learners, usually decision trees, and improving their performance via boosting. As a result, XGBoost can better understand the dataset and provide precise predictions. Within the provided regression analysis framework, XGBoost demonstrated significantly stronger correlation coefficients for slump flow, J-ring flow, and V-funnel values, outperforming LR in each of the three cases.

3.2. Mechanical properties

3.2.1. Effects of WRTA on compressive strength of SCRC

Obtained test results for the compressive strength of all the mix designs are presented in Table 5 concerning several statistical parameters such as standard deviation, coefficient of variation (COV), mean strength, standard error, and the lower and upper range of 95% confidence interval for both 7 days and 28 days. Three consecutive specimens were tested, and the mean was taken for each data point. It is seen from the data that the compressive strength for every SCRC mix was within the limit of 16.04 MPa–32.04 MPa after 7 days, while for 28 days, the strength increased to a range of 23.43 MPa–32.04 MPa. The test results deviated from 0.30 to 0.71. On the contrary, the corresponding COV is 0.013%–0.036%. After 7 days of curing, the lowest compressive strength was seen for SCRC20 with 16.04 MPa with a respected lower limit and upper limit of 95% confidence interval of 15.63 MPa and 16.45 MPa, respectively. By contrast, the highest compressive strength was seen for the control mix with 32.04 MPa with lower and upper limits of the confidence interval of 31.59 MPa and 32.49 MPa. On the other hand, after 28 days of curing, SCRC20 showed the least strength with 23.43 MPa with upper and lower confidence intervals of 23.98 MPa and 22.88 MPa, respectively, while control showed the best results with 37.98 MPa with upper and lower confidence intervals of 38.57 MPa and 37.39 MPa respectively.

Table 5.

Statistical findings of compressive strength.

Mix	Days	Mean strength (MPa)	Standard deviation	COV (%)	Standard error	95% confidence interval
						Lower Limit	Upper Limit
Control	7	32.04	0.45	0.014	0.259	31.59	32.49
	28	37.98	0.59	0.015	0.341	37.39	38.57
SCRC5	7	22.03	0.30	0.013	0.173	21.84	22.42
	28	27.64	0.66	0.02	0.381	26.98	28.30
SCRC10	7	19.58	0.71	0.036	0.409	18.87	20.29
	28	25.23	0.47	0.018	0.271	24.76	25.70
SCRC20	7	16.04	0.41	0.026	0.237	15.63	16.45
	28	23.43	0.55	0.023	0.318	22.88	23.98

Fig. 11 illustrates the obtained results from the compressive strength test for the mix designs. As expected, it was seen that the compressive strength decreased with the increment of rubber percentage in the mix design. The control mixture showed the best results, while SCRC20 gained the least strength. By evaluating normalized strength, it was seen that the strength reduced by 32%, 37%, and 50% for 5%, 10%, and 20% replacement of WRTA, respectively at 7 days, while for 28 days, the reduction was 28%, 34% and 39% for 5%, 10%, and 20% replacement compared to the control mix design. The feeble bond between the rubber and cement matrix may be the reason behind this shrinkage, which initiated the formation of an interfacial transition zone. Due to the feeble zone, the rubber aggregates often act as voids affecting the strength [13]. On the other hand, Reda Taha et al. [27] concluded that the strength reduction might be because the rubber aggregates act as soft aggregates. Though for some proportions of rubber, the decline in strength was not that great, when the ratio of rubber is increased, it undergoes a drastic fall of up to 50% of strength compared with the control mixture. Similar findings were found in the previous research. It is seen that coarser crumb rubber aggregates affect compressive strength more compared to fine aggregates [61,62]. Aslani et al. [13] observed that the 10 mm crumb rubber replacement with coarse aggregates showed the lowest compressive strengths among the series after 28 days of curing (25.57–15.13 MPa). This is due to a larger proportion of volume being occupied by coarser aggregates, which play an important role in the strength of SCC compared to the fine aggregates. There was a reduction of about 64% from 73.1 MPa to 26.4 MPa for 0%–25% replacement by rubber aggregates [55]. Werdine et al. [15] found 10% replacement with a particle size of 0.6 mm–4.8 mm to have a compressive strength of 50.57Mpa after 28 days of curing. Meanwhile, Li et al. [57] observed a reduction of 33.5% for 90 kg of rubber aggregates, while the declination of strength was linear with the increment of rubber aggregates. According to Hilal [62], as rubber aggregates are soft compared to natural aggregators, the bond between the surrounding cement matrix and rubber aggregates decreases. It was found that coarse rubber aggregates had a vastly negative impact compared to fine rubber aggregates [63]. So, the experimental findings are like the previous research.

Fig. 11.

Comparison of compressive strength test.

3.2.1.1. Machine learning on compressive strength

Fig. 12 is a representation of the intended and practical consequences of compressive strength through SCC boxplots. Boxplots are commonly used to represent data transportation, and this method typically includes a five-number summary that consists of the minimum, first (lower), median, third (upper), and greatest scores. The analytical summaries provide information related to the mean, median, interquartile range, minimum, and maximum. It is reasonable to infer that there was little difference between the groups, given that the average lines for all data sets were enclosed inside their respective boxes (see Fig. 10). Distributions of results showed a medium dispersion, with interquartile ranges of roughly 22–30 MPa.

Fig. 12.

Box plot showing the data distribution for compressive strength.

Previous research subjected the proposed ML models, specifically the LR model, to substantial scrutiny. Nevertheless, the substantial originality of this study resides in the implementation of XGBoost. With its advanced ensemble learning technique, XGBoost successively combines weak learners and boosts their performance, setting it apart from typical LR models. As a result, XGBoost can capture complex nonlinear correlations in the data, which improves its prediction power. The complexity of predicting the compressive strength of concrete is a significant challenge, but XGBoost can handle it. However, regarding precision, when compared to the LR model, XGBoost performed significantly better in predicting compressive strength (see Table 6). Similarly, Fig. 13 presents analytic results comparing experimental and anticipated results, showing a 0.97 correlation coefficient (R²), which is consistent with XGBoost's performance. Compared to the prior model, the LR model conversely produced a low coefficient (R² = 0.78). But generally, the proposed ML models operate admirably, as seen by the overall results. This development represents a significant advancement in the accuracy of predictive modeling for properties of this nature. The effective prediction of concrete strength using XGBoost, which provides a more refined and dependable substitute for traditional linear models, constitutes the study's contribution to the field. According to Nguyen et al. [64], the XGBoost approach achieves an accuracy of 97% for predicting the compressive strength of concrete composites.

Table 6.

Performance of the developed models.

Model	Linear Regression	Extreme Gradient Boosting
R2	0.78	0.97
RMSE	3.01	1.16
MSE	9.07	1.36
MAE	2.94	0.76

Fig. 13.

Measured versus predicted compressive strength.

3.2.1.2. Failure and crack pattern

The failure patterns were observed and depicted in Fig. 14 during the experiment. Several failure patterns were observed with the varying rubber proportion within the concrete mixture. After implementing the compression strength test, all the cylindrical specimens were closely observed and inspected to identify the failure nature and the formation of cracks. The failure patterns varied from shear cracks followed by multiple cracks around the specimen. First, the control specimen showed shear cracks with concrete spalling. A few cracks were visible on the upper portion visualized in Fig. 14(a). SCRC5 underwent concrete spalling from various sides, visualized in Fig. 14(b). On the other hand, the SCRC10 specimen showed concrete spalling from the surface with wedge formation and multiple crack formation on its body, shown in Fig. 14(c). As having the lowest strength among all the others, SCRC20 underwent a vast number of cracks containing shear cracks, while full wedge formation was observed at the start of the load cycle, as shown in Fig. 14(d). The weak interfacial zone within the rubber and cement matrix creates a void within the concrete. These voids react faster than the cement matrix during loading, initiating cracks [13]. It can be said that the crack formations of these specimens are not consistent with the varying rubber proportions.

Fig. 14.

Failure and crack patterns after compressive strength test.

3.2.2. Effects of WRTA in splitting tensile strength of SCRC

Table 7 shows the splitting tensile strength of all the mixtures concerning different statistical parameters such as standard deviation, coefficient of variation (COV), mean strength, and the lesser and higher range of 95% confidence boundary for 7 days and 28 days. Each detail was obtained from the average of the three consecutive experimented samples. It is seen from the data that the tensile strength among all SCRC mixes was within the range of 2.89 MPa–1.83 MPa after 7 days, while for 28 days, the strength increased to a range of 2.097 MPa–3.78 MPa. There was a deviation of 0.05–0.25 for all test results. Meanwhile, the COV (%) varied from 0.018 to 0.07. After 7 days of curing, the lowest tensile strength was seen for SCRC20 with 1.83 MPa with a respected lower limit and upper limit of 95% confidence interval of 1.73 MPa and 1.93 MPa, respectively. By contrast, the highest tensile strength was seen for the control mix with 2.89 MPa with lower and upper limits of the confidence interval of 2.66 MPa and 3.12 MPa. On the other hand, after 28 days of curing, SCRC20 showed the least strength with 2.097 MPa with upper and lower confidence intervals of 2.02 MPa and 2.174 MPa, respectively, while control showed the best results with 3.78 MPa with upper and lower confidence intervals of 3.95 MPa and 3.61 MPa respectively.

Table 7.

Statistical findings of splitting tensile strength.

Mix	Days	Mean strength (MPa)	Standard deviation	COV (%)	Standard error	95% confidence interval
						Lower Limit	Upper Limit
Control	7	2.89	0.23	0.07	0.132	2.66	3.12
	28	3.78	0.17	0.04	0.098	3.61	3.95
SCRC5	7	2.56	0.15	0.05	0.086	2.41	2.71
	28	3.14	0.25	0.07	0.144	2.89	3.39
SCRC10	7	2.28	0.13	0.05	0.075	2.15	2.41
	28	2.74	0.05	0.018	0.028	2.69	2.79
SCRC20	7	1.83	0.10	0.05	0.057	1.73	1.93
	28	2.097	0.07	0.03	0.040	2.02	2.174

Fig. 15 compares the obtained results from the splitting tensile strength test for all the mix designs. It was seen that the tensile strength showed similar characteristics to compressive strength. It decreased with the enhancement of rubber proportion in the mixture. The control mixture showed the best results, while SCRC20 gained the least strength. Compared to the control mixture, about 12%, 22%, and 37% reduction in splitting tensile strength was observed for 5%, 10%, and 20% replacement of WRTA, respectively at 7 days, while for 28 days the reduction was 17%, 28% and 45% for 5%, 10% and 20% replacement with WRTA. This is due to the same reason as compressive strength. Also, rubber particles react more than the cementitious part in its surroundings during loading. Failure is initiated due to the cracks being generated within the layer of the interfacial transition zone [13,63]. Several researchers also claimed the strength declined with the increment of the proportion of rubber in concrete mixtures. 10 % incremental increase of crumb rubber for 2 mm, 5 mm, and 10 mm rubber replaced designs resulted in a decline in tensile strength with a proportion of 12%, 8%, and 16%, respectively [13].

Fig. 15.

Comparison of splitting tensile strength test.

On the contrary, Fig. 16 compares splitting tensile strength to compressive strength ratio percentage of SCRC with corresponding compressive strength. At 7 days, the average coefficient of regression value was found to be R² = 0.9966; at 28 days, the average coefficient of regression value was found to be R² = 0.9138. Other researchers also concluded that the splitting tensile strength of these cylindrical specimens was roughly around 10% of the compressive strength [62]. However, comparing the compressive strength and splitting strength, it was seen that the declination of splitting tensile strength is lower than compressive strength. It's also similar to the previous research showing that the decrease in the tensile strength was lower than the declination in compressive strength for SCRC [7].

Fig. 16.

Relationship of splitting tensile strength to compressive strength ratio with compressive strength for 7 days and 28 days.

3.2.2.1. Machine learning on splitting tensile strength

The spatial distribution of tensile strength measurements is portrayed by employing a box plot similar to the one applied to depict compressive strength data (see Fig. 17). Based on the box lengths (interquartile ranges), the insights were not very dispersed, lying somewhere between 2.3 and 3.1 MPa. However, outcomes derived from implemented models are shown in Table 8.

Fig. 17.

Box plot showing the data distribution for splitting tensile strength.

Table 8.

Performance of the developed models.

Model	Linear Regression	Extreme Gradient Boosting
R2	0.97	0.99
RMSE	0.092	0.013
MSE	0.014	0.009
MAE	0.105	0.064

Fig. 18 also displays the R² values, measuring how well the models predict the results. Due to the superior predictive power, XGBoost (R² = 0.99) is preferred over LR (R² = 0.97). Additionally, both provided models had a good correlation coefficient, demonstrating their broad usefulness for strength prediction. The effectiveness of XGBoost in detecting complex patterns and nonlinear correlations in the data is shown by this significant enhancement. In addition to its superior R² value, XGBoost's sophisticated ensemble learning methods and boosting iterations are the main reasons for its popularity. Although it produces respectable outcomes, the LR model is less precise than XGBoost. The research's ground-breaking contribution is XGBoost, which was hand-picked for its versatility and outstanding effectiveness in forecasting concrete strength. For researchers and practitioners aiming to use sophisticated ML methods for improved concrete material prediction, this comprehensive understanding of the XGBoost model is crucial. By adopting the XGBoost method, Nguyen et al. [64] achieved a precise rate of 98% for tensile strength measurement.

Fig. 18.

Measured versus predicted splitting tensile strength.

3.2.2.2. Failure and crack pattern

Almost familiar failure patterns were seen when the rubber replacement in the concrete mixture was varied. After implementing the tensile strength test, all the cylindrical specimens were closely observed and inspected to identify the failure nature and formation of cracks. First, the control specimen showed some pure tensile failure with some cracking. The crack was introduced in the middle along the periphery of the diameter, ultimately turning into a brittle failure. On the other hand, SCRC5, SCRC10, and SCRC20 specimens showed pure tensile failure in the same process as the control specimen. Similar to compressive strength, the initiation of these cracks is an effect of void within the weal bond of WRTA and cement matrix. These voids react faster than the cement matrix during loading, initiating cracks. It can be said that the crack formation of these specimens is more consistent than the compressive strength test with the varying rubber proportion showing pure brittle failure with cracks.

3.3. Linkage between mechanical properties and analytical prediction

3.3.1. Splitting tensile strength vs compressive strength

Through a scatter diagram, Fig. 19 depicts the relationship between splitting tensile strength and compressive strength. The projected regression model of SCRC is illustrated by a straight path in Fig. 19 and displayed in Equation (6). The coefficient of determination R² of these relations is about 85%, indicating a very strong correlation.

ƒ_ct = 0.0817 ƒ_c’ +0.5821 (6)

here, ƒ_ct = mean splitting tensile strength, ƒ_c’ = mean compressive strength.

Fig. 19.

Splitting tensile strength vs compressive strength.

Confidence intervals were calculated for both equations and were presented in Fig. 19. These actual values of splitting tensile strength were also contrasted with the calculated strengths of various standards illustrated in Fig. 19. The AS 3600 [65], ACI 363 R [66], ACI 318 [67], and CEB-FIP [68] give models [Eqns. (7)–(10)] to represent the relationship between splitting tensile and compressive strength.

ƒ_ct = 0.4 √ƒ_c’ (AS 3600) (7)

ƒ_ct = 0.56√ƒ_c’ (ACI 318) (8)

ƒ_ct = 0.59√ƒ_c’ (ACI 363R) (9)

ƒ_ct = 0.3 (√ƒ_c’)^2/3 (CEB-FIP) (10)

After analyzing the linear regression model, it was concluded that most of the code standards lie within the upper and lower confidence intervals. In most cases, the values were near the experimental values of this experiment. However, a little scatter was observed for AS 3600 [65]. This phenomenon means that a conservative design can be obtained using existing code standards regarding tensile strength for SCRC. On the other hand, it also indicates that SCRC is more tensile than compression. By contrast, Fig. 20 compares the experimental and estimated values for splitting tensile following compressive strength with the varying code of standard analytical formulation mentioned in Eqs (2)–(5). It is observed that the normalized experimental and predicting data provides good estimation by most of the standard codes. However, a slight scatter was observed for AS 3600 [65] and ACI 363R [66] in this experiment. The prediction for the experimental ratio of splitting tensile strength helps determine the relation with compressive strength, which also helps identify the experimentation without conducting, which saves different factors associated with this experiment, such as labour, time, and material cost.

Fig. 20.

Predicted and experimental splitting tensile strength of SCRC.

4. Conclusion

The influence of varying proportions of WRTA in an SCC mixture to produce SCRC with moderate strength was studied. The experimental results of rheological and mechanical properties are summarized below:

• ⁃All the SCRC mixtures demonstrated commendable workability, successfully meeting and surpassing the acceptable limits for passing ability and flowing characteristics.
• ⁃An increment in the proportion of WRTA percentage in SCRC resulted in a decrease in workability. All slump and j-ring flow diameter values satisfied EN 12350.
• ⁃The increment of replacement by WRTA resulted in a decrease of slump flow up to 10% for 20% replacement of coarse aggregate with WRTA compared to the control mixture. Similarly, the addition of WRTA decreased the J-ring slump flow value, indicating a decline in passing ability.
• ⁃The compressive strength dropped by 32%, 37%, and 50% for 5%, 10%, and 20% WRTA substitution at 7 days and 28%, 34%, and 39% for 28 days compared to the control mix design.
• ⁃Similar compressive strength patterns were seen in the splitting tensile strength test findings. The decrease over 28 days was 17%, 28%, and 45% for WRTA substitution at 5%, 10%, and 20%, respectively.
• ⁃The concrete's mechanical properties test fulfilled the ACI requirements with a 10% WRTA substitution of coarse aggregate.
• ⁃The failure pattern varied from shear cracks following multiple cracks around the edges to full wedge formation with increased WRTA proportion after compression. However, pure tensile failure with multiple cracking was observed during the splitting tensile strength test.
• ⁃ML predicted findings showed that both LR and XGBoost were adept at anticipating fresh and mechanical traits. Regarding the correlation coefficient, the XGBoost approach proved more reliable and provided better estimation with the experimental results than the LR model; however, both models are useable alternatives for estimating concrete strength.

From all the perspectives studied, the optimum percentage of WRTA as a partial replacement of coarse aggregates in SCC is 10%. However, the 20% substitute mixture can be used for several kinds of urban civil works like water drainage, highway blocks, dividers, and landscaping projects.

Ethical consideration

We have reviewed Ethics in Publishing as well as Heliyon's Ethics and Editorial Policies for this research.

Funding

There are no specific funding sources for this research.

Data availability statement

Data should be available upon request.

CRediT authorship contribution statement

Md. Habibur Rahman Sobuz: Conceptualization, Methodology, Formal analysis, Investigation, Validation, Data curation, Supervision, Writing - original draft, Writing - review & editing. Limon Paul Joy: Conceptualization, Methodology, Validation, Data curation, Writing - original draft, Writing - review & editing. Abu Sayeed Mohammad Akid: Formal analysis, Validation, Data curation, Writing - original draft, Writing - review & editing. Fahim Shahriyar Aditto: Formal analysis, Investigation, Writing - review & editing. Jannat Ara Jabin: Formal analysis, Data curation, Writing - review & editing. Noor Md. Sadiqul Hasan: Formal analysis, Writing - review & editing. Md Montasor Meeraz: Formal analysis, Writing - review & editing. Md. Kawsarul Islam Kabbo: Investigation, Writing - review & editing. Shuvo Dip Datta: Investigation, Writing - review & editing.

Declaration of competing interest

The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper.