Mechanical and tribological response of red brick dust filled hemp-epoxy composites using response surface methodology and machine learning approach

Abstract

The current study focuses on the wear and mechanical properties of epoxy-based hybrid composites that contain hemp fiber and red brick dust (RBD). The hand lay-up method is used for making composites using hemp fiber (40 wt%) reinforced epoxy with varied RBD loading up to 15 wt%. The prepared samples are assessed by investigating their physical, mechanical and tribological behavior. The results revealed that the density increased from 1.257 to 1.338 g/cm³ with increasing RBD content, while tensile and flexural strengths decreased by approximately 13.3 and 7.4%, respectively, at 15 wt% filler loading. In contrast, impact strength and hardness improved significantly, showing enhancements of 144 and 152%, respectively, compared to neat epoxy. The sliding wear tests are also carried out in accordance with the L₃₀ model of response surface methodology (RSM). According to reports, the two most important factors affecting the composites’ wear rate are RBD content and sliding velocity. The developed RSM models showed good agreement with experimental results, with prediction errors within ± 5%. Additionally, an efficient model for predicting the wear characteristics of the RBD-filled hemp-epoxy composites is developed using three machine learning (ML) models: decision tree (DT), random forest (RF), and gradient boosting machine (GBM). Among them, the GBM model demonstrated superior performance, achieving the highest accuracy with R² values of 0.9542 for specific wear rate and 0.9879 for coefficient of friction, along with the lowest prediction errors. Moreover, the investigation revealed that filler content and normal load are the most important variables affecting the estimation of specific wear rate and COF, respectively.

Introduction

Nowadays, the utilization of solid wastes as filler in polymer matrix composites (PMCs) has gained attention over the past few years because of their easy availability at low cost^1–5. Recently, Khoaele et al.⁶ reviewed the opportunities and difficulties associated with the manufacturing of composites with polymer as matrix and waste wood as reinforcement. The development of composites made of polyester and waste marble dust was investigated by Nayak and Satapathy⁷. They observed that these composites could be useful in wear-related engineering applications due to their enhanced hardness and impact strength. Smoleń et al.⁸ highlighted the incorporation of milled waste of CFRP as fillers in polyester composites. In their analysis, they revealed that the CFRP-containing compositions had higher viscosity and flexural strength than the neat resin. Alkaron et al.⁹ studied the thermal and chemical properties of eggshell powder reinforced polyamide 12 composites with different ratios (3, 5, and 10 wt%). The results for eggshell/polyamide composites indicated that 5 wt% filler composites had the potential for better eggshell particle distribution. Rattanapan et al.¹⁰ worked on the mechanical characteristics of polypropylene composites filled with waste cement. In comparison with conventional cement, the waste cement shows a more significant impact. The implementation of the waste generated from the demolition was utilized for the development of hybrid composites with a combination of polymer and wood by Keskisaari et al.¹¹. The flexural strength and modulus of the developed material were reduced with the inclusion of mineral fillers. However, the impact strength increased when a portion of the wood was replaced with waste filler. The presence of solid wastes as filler in PMCs has also improved the wear resistance. It is a key challenge to meet the design specifications of the manufactured components under sliding situations. Gupta et al.¹² recently discussed the sliding wear response when Kota stone dust (KSD) is added to the epoxy resin. They found that adding KSD to epoxy resin significantly enhanced the resistance against sliding wear. Ray et al.¹³ also investigated the utilization of eggshell powder in epoxy and studied the wear behavior of the developed composites. Pati and Satpathy¹⁴ reported the tribological characterization of epoxy composites reinforced with glass fiber and micro-sized red brick dust. The sliding velocity was the most influential control factor which affects the material loss of the composite material. Nayak and Satapathy^15,16 reviewed the performance of polyester/waste marble dust composites. From the experimentation, it is concluded that the most influential factor which affects the wear rate is the filler loading and the velocity of the sliding. Purohit et al.¹⁷ also reported a similar finding that the filler loading, which is LD sludge, is the prime factor which governs the polymeric composites developed by them. The mechanical and abrasive wear behavior of epoxy/LD slag composites was evaluated by Lohiya et al.¹⁸. They revealed that the specific wear rate of the developed material under dry sliding conditions reduces with increased hardness. Maluga et al.¹⁹ used both granite and marble dust for the development of polymer composites. The developed composites were investigated for their mechanical and tribological properties. The analysis revealed the better mechanical properties of waste granite/marble dust reinforced composites as compared to unfilled epoxy. They also found improved tribological behavior for the combination. Ajibade et al.²⁰ described the wear rate model for agro-waste-filled epoxy composites and found that the model was very appropriate for the proposal of brake pads.

Recent studies on natural fiber bio-composites show strong potential for improved mechanical and environmental performance. A composite made with Mimosa cinerea fiber and ginger waste filler in a Poly lactic acid (PLA) matrix showed higher tensile strength and reduced swelling due to better fiber-matrix bonding and fewer voids²¹. Another study on banana fiber reinforced plasma-treated polyethylene reported improved viscoelastic and thermal stability because of enhanced interfacial interaction. Surface treatments such as NaOH and plasma treatment play a key role in improving compatibility and overall composite performance. These findings highlight the importance of fiber treatment and hybrid filler use in developing efficient and sustainable bio-composites²². Cement mortars reinforced with natural fibers such as hemp improve flexural and compressive strength and reduce shrinkage. These bio-based composites also offer better dimensional stability and provide a sustainable alternative to synthetic fiber-reinforced materials^23,24.

In practical applications, predicting the wear performance of functional composites is essential for minimizing failures and maintenance expenses. One possible solution lies in finding correlations between the different factors affecting wear rate using a data-driven approach. This method uses a variety of mathematical techniques to minimize wear rate based on experimental data. Machine learning and artificial intelligence are found to be powerful tools in this process, enabling the prediction of performance outcomes in non-linear complex dependency among the input and the output data^25,26. Recently, researchers have applied data-driven machine learning to forecast the wear behavior of metal matrix and polymer matrix composites^27,28. Russel et al.²⁹ used machine learning techniques to predict the wear behavior in graphene oxide reinforced hybrid polymeric composites and concluded that the decision tree model accurately predicted the wear rate and coefficient of friction. Singh et al.³⁰ developed predictive models like Artificial neural networks (ANN), RF and GBM to forecast the COF of carbon nanotube reinforced aramid-epoxy composites and reported that the GBM model showed higher accuracy. Hasan et al.³¹ compared various machine learning models to envisage the friction and wear behavior of graphene-reinforced Al-based composite and found that out of all the considered ML models, GBM provided the highest accuracy for COF prediction, and the ANN model provided the best prediction for wear rate. Wang et al.³² evaluated the tribological properties of PTFE composites and concluded that the excellent outcome is delivered by a gradient boosting regression model. Sheikh et al.³³ employed four different ML models, i.e. DT, RF, support vector machine (SVM) and k-nearest neighbour (KNN) models for the wear rate forecast of cenosphere-filled aluminium composites and highlighted the efficacy of the random forest model. Singh et al.³⁴ performed a similar study and found that the extreme gradient boosting model outperforms other models. Mahakur et al.³⁵ concluded that the SVM model showed the highest accuracy and least errors. Deshpande et al.³⁶ attempted to apply different ML algorithms to forecast the specific wear rate of PTFE/glass composites and concluded that the GBM model stands out as the best model.

Though the majority of research has been done on a large range of filler and fiber reinforced PMCs, a few reports on reinforcing waste filler (red brick dust (RBD)) and natural fiber (hemp fiber (HF)) in an epoxy-based system, particularly from a tribological perspective. Although machine learning models are widely used in other fields of research and engineering, their application in material science has been limited, especially in predicting the tribological performance of the hybrid composite system. Therefore, the current investigation aims to bridge the gap by investigating the epoxy hybrid composite system with respect to the mechanical and sliding wear behavior. The significance of this work lies in the utilization of a sustainable waste material (RBD) in combination with natural fibers to develop eco-friendly composite systems with improved functional performance. Furthermore, the novelty of this study resides in the integration of experimental analysis, response surface methodology (RSM), and multiple machine learning models (DT, RF, and GBM) to accurately predict tribological behavior. This combined approach not only enhances the understanding of wear mechanisms but also provides a reliable predictive framework for the design and optimization of hybrid polymer composites.

Materials and methods

Experimental details

For the development of polymer composites in the present work, epoxy (Lapox L-12) resin is used along with the HY 951 (TETA) hardener. Both materials were supplied by Atul Ltd., Mumbai, India. Hardener is used for curing the epoxy resin, where epoxy and hardener are combined in a 10:1 ratio as per the manufacturer’s recommendation. The hemp fiber is used as a primary reinforcement material whose content is kept constant in all sets of the composites. The micro-sized red brick dust is used as a secondary filler material in the current work. The same is collected from the nearby red brick industry, which is then sieved to get an almost uniform-sized particulate. Both phases are combined via the hand lay-up method for the development of the composite body. The samples are prepared at three different loading levels of red brick dust and a constant content of hemp fiber. The composition used for the fabrication of three sets of composites is presented in Table 1. The prepared samples are presented in Fig. 1.

Table 1.

Sample composition.

Sl. no	Hemp fiber (wt%)	RBD content (wt%)	Epoxy resin (wt%)
1	40	0	60
2	40	7.5	52.5
3	40	15	45

Fig. 1.

Hemp fiber reinforced epoxy composites consisting of (a) 0 wt%, (b) 7.5 wt%, (c) 15 wt% of red brick dust.

The compatibility of the various phases and wear-related damage is examined using a Carl Zeiss scanning electron microscope. The composite density is calculated using a simple water immersion method. The ASTM D 792 − 91 standard is followed for conducting the trials. The theoretical density must be determined using a correlation created by Agarwal and Broutman (Eq. 1) in order to compute the voids created during the composite production³⁷.

1

Here, W is the percentage of weight, ρ is the density, and f, p, and m stand for fiber (Hemp), filler (RBD), and matrix (epoxy), respectively. The percentage deviation between the measured density and the calculated density gives the content of voids. The water uptake by the composites is measured following the ASTM D 570 standard. Under mechanical characterization, tensile tests (ASTM D638), compressive tests (ASTM D695) and flexural tests (ASTM D790) are performed with an Instron Universal testing machine, whereas hardness (ASTM D-2240) is measured in Shore D number using a PosiTector Durometer³⁸. For each material variant and test condition, three specimens were tested, and the reported results represent the average values.

The sliding wear test is conducted as per ASTM G99³⁹ standard using a well-known DUCOM pin-on-disc tribometer. Table 2 provides information on the different control factors and their levels used in the wear tests. To design the experiments using the different levels of the selected control factors, a well-established response surface method (RSM) technique in the MINITAB 18 software was used. RSM is a statistical tool integrated with mathematical techniques, often used to design and optimize processes where one output response is dependent on several input variables⁴⁰. It establishes a correlation between the output response and the input variables through regression equations, which helps in identifying the optimal control factors and can also be used for future predictions^41,42 In RSM, the common designs used are the central composite design (CCD) and the box-behnken design (BBD). In the present research, the experiments were designed using a CCD of the L₃₀ model of RSM with two continuous factors in the MINITAB 18 software⁴³.

Table 2.

Control factors and levels for performing experimentation.

Control factors	Units	Levels
A: Sliding Velocity	cm/s	157	275	393
B: Normal Load	N	5	12.5	20
C: RBD Content	wt%	0	7.5	15
D: Sliding Distance	m	600	1500	2400

According to the design, 31 combinations of the levels of selected control factors were obtained as displayed in Table 5. The 31 combination of test runs was selected following Eq. (2)⁴⁴.

Table 5.

Experimental and predicted coefficient of friction values of the composites corresponding to the combination of test parameter levels as per the L₃₀ model of RSM.

Sl. no.	Sliding velocity (cm/s)	Normal load (N)	RBD content (wt%)	Sliding distance (m)	Coefficient of friction		Error (%)
					Experimental	Predicted
1	157	20.0	0.0	600	0.617	0.610	1.14
2	393	12.5	7.5	1500	0.584	0.575	1.56
3	157	5.0	15.0	600	0.576	0.572	0.69
4	275	5.0	7.5	1500	0.588	0.566	3.88
5	393	20.0	15.0	600	0.592	0.593	0.16
6	275	12.5	7.5	1500	0.624	0.642	2.80
7	275	12.5	7.5	1500	0.632	0.642	1.55
8	393	5.0	15.0	2400	0.535	0.536	0.18
9	393	5.0	0.0	600	0.426	0.423	0.70
10	275	12.5	7.5	600	0.651	0.634	2.68
11	393	20.0	0.0	600	0.532	0.534	0.37
12	275	20.0	7.5	1500	0.689	0.685	0.58
13	275	12.5	7.5	2400	0.682	0.673	1.33
14	157	5.0	0.0	2400	0.512	0.505	1.38
15	275	12.5	15.0	1500	0.659	0.643	2.48
16	157	5.0	0.0	600	0.437	0.450	2.88
17	393	5.0	0.0	2400	0.459	0.463	0.86
18	275	12.5	0.0	1500	0.565	0.556	1.61
19	157	12.5	7.5	1500	0.672	0.656	2.43
20	157	20.0	15.0	2400	0.753	0.750	0.4
21	275	12.5	7.5	1500	0.615	0.642	4.20
22	275	12.5	7.5	1500	0.621	0.642	3.27
23	275	12.5	7.5	1500	0.637	0.642	0.77
24	393	20.0	0.0	2400	0.564	0.562	0.35
25	393	20.0	15.0	2400	0.617	0.615	0.32
26	157	20.0	0.0	2400	0.648	0.654	0.91
27	157	5.0	15.0	2400	0.613	0.622	1.44
28	275	12.5	7.5	1500	0.629	0.642	2.02
29	275	12.5	7.5	1500	0.659	0.642	2.64
30	157	20.0	15.0	600	0.705	0.712	0.98
31	393	5.0	15.0	600	0.496	0.502	1.19

2

where,

k = number of selected control factors (4 in the present research).

n = number of centre point replicates (Generally 3–6 or more for lower error estimation).

In the present research, as there are four control factors, the number of test runs considering 7 centre point replicates is 31.

Tribological tests of the composite pins were conducted using this combination of well-established test parameters. Emery papers with grit size P400 were glued using Araldite on the surface of a steel disc having a diameter of 100 mm to be used as the counter surface. The composite pins were polished properly before the wear test, and the average roughness of the pin surface was measured to be approximately 0.52 ± 0.18 μm. The loss of mass after each test was calculated by subtracting the test specimen mass after the wear test from the mass before the wear test. The specific wear rate was then calculated using Eq. 3.

3

where,

Δm = mass loss (g).

ρ = density of the composites (g/cm³).

L = sliding distance (m).

N = applied load (N).

Predictive modelling using machine learning

Machine learning models are primarily classified into three types, namely, supervised, unsupervised and reinforcement learning. In the present work, wear data consists of continuous labelled data sets; as a result, supervised machine learning regression models are used in this work. Tree-based machine learning techniques are effective in predicting the tribological behavior of materials⁴⁵. These models can handle non-linear, complex relationships in tribological systems and are effective even with smaller data sets and are flexible, which can be implemented in classification as well as regression problems⁴⁶. Using these advantages, this study aims to predict the specific wear rate and COF of RBD-filled hemp-epoxy composites under different test conditions using various machine learning (ML) algorithms, including DT, RF and GBM. The implementation of these algorithms is carried out in Python, utilizing key libraries such as “Pandas,” “NumPy,” “Scikit-learn,” and “Matplotlib”. These ML models are discussed below in detail.

A decision tree is a supervised learning technique that uses a non-parametric approach to create decision rules from data attributes, allowing it to accurately predict the target values. Previous researchers have shown that the DT model handles and predicts tribological data well^47,48. As its name suggests, it comprises roots, nodes and leaves. The decision-making process follows a tree-like structure, which starts from a root node and then proceeds to the leaf node through branches based on defined conditions at each node. The outcomes are represented through decision nodes and leaf nodes²⁶. To get a more effective prediction than a DT model, methods like the random forest (RF) model increase prediction accuracy by combining multiple trees. It is also called a bagging model, which selects data randomly and can manage highly complex, non-linear data sets with several input features and can reduce overfitting problems. RF model functions like a democratic voting system, where the final prediction is the average of the decisions made by individual trees, differentiating it from a single decision-making model⁴⁹. Similar to random forest, the gradient boosting machine model also consists of multiple decision trees, but the process of predicting data is different to that of the RF model. In this model, regression trees are built one after another, step by step. GBM iteratively optimizes residual errors, enhancing predictive performance by sequentially eliminating weak learners, and this process is called the boosting mechanism⁵⁰. This model is generally used in solving problems related to materials science due to its capability to handle intricate dependencies. The deployment of these models involves various stages, which are discussed below step by step:

Data acquisition and pre-processing

In order to create machine learning models, wear tests are conducted on hemp-epoxy composites loaded with RBD. Four input variables: normal load, sliding distance, sliding velocity, and RBD composition and two output variables: specific wear rate and coefficient of friction make up the data set. Since the information makes it clear that the input and output variables use a continuous labelled data set, supervised machine learning regression models are employed to predict the material’s wear behavior. To increase the accuracy and reliability of the model, the wear data is cleaned and pre-processed to correct outliers and missing values. The “Standard Scaler” function is used to standardize the data set because the input parameters had multiple numerical scales. The data set is split into training and testing data sets in a 75:25 ratio following pre-processing. The training set is used to develop the model, while the test set is used to assess its performance.

Hyperparameter tuning of the ML models

After training and testing the data, parameter tuning is a crucial step to maximize the predictive performance and improve the prediction accuracy of the model. In order to find the configuration that maximizes accuracy, this approach entails fine-tuning the settings using a trial-and-error method in Python. Three-fold cross-validation combined with a grid search algorithm is employed to explore and optimized a set of parameters to achieve the highest accuracy. The final optimized parameters are selected based on the highest value of the coefficient of determination and the lowest value of errors. Table 3 shows the selected optimal parameters which contributed to improved prediction performance of each machine learning model.

Table 3.

Control factors and levels for performing experimentation.

ML model	Optimized parameters for SWR
DT	criterion = “friedman_mse”, splitter = “best”, min_samples_split = 3, max_depth = 3
RF	n_estimators = 100, criterion = “friedman_mse”, max_depth = 3
GBM	n_estimators = 7, learning_rate = 0.4, max_depth = 2

ML model	Optimized parameters for COF
DT	criterion = “squared_error”, splitter = “best”, min_samples_split = 2, max_depth = 4
RF	n_estimators = 100, criterion = “squared_error”, min_samples_split = 2
GBM	n_estimators = 5, learning_rate = 0.2, max_depth = 5

Performance assessment criteria for ML models

A variety of performance metrics are used to assess each machine learning model’s performance. The coefficient of determination (R²), mean squared error (MSE), root mean squared error (RMSE), mean absolute error (MAE), and mean absolute percentage error (MAPE) are the five statistical metrics that are typically used to evaluate it. Equations (4)–(8) are used to assess these metrics.

4

5

6

7

8

Results and discussion

Density and void

Figure 2 shows the density of the developed hemp-epoxy composites. The density of epoxy was measured to be 1.2 g/cm³. At the first stage, an epoxy/hemp fiber mat composite is prepared where the hemp fiber content is kept at 40 wt%. Since the density of hemp fiber (1.48 g/cm³) is more than epoxy resin, the composite shows a higher density value. The measured value of density for the combination is 1.257 g/cm³. With the insertion of red brick dust, the density further increases, as red brick dust also has a higher density value than epoxy resin. The density of red brick dust is 2.30 g/cm³. The density increases with red brick dust content, and for a loading of 15 wt%, the density reaches 1.338 g/cm³. This value is 11.5% higher than the density of neat epoxy and 6.4% higher than the density of epoxy/hemp fiber composite. It is also observed from the figure that the calculated density of the composites is greater than the experimental counterpart. The difference in density between the two values is primarily due to the presence of voids in the actual composites. The air voids have less density than the other phases, and because of that, the actual composites have a lower density value. The percentage difference between the experimental and the theoretical density gives the air voids present inside the composites, as given in Table 4. For the epoxy/hemp fiber composite, the void content present is 3.16%, which increases with the loading of red brick dust. The maximum void content present inside the composite is 4.93% for a red brick content of 15 wt%.

Fig. 2.

Density and void content of epoxy-based composites.

Table 4.

Density and voids of the hemp-epoxy composites with RBD content.

RBD content (wt%)	Density (g/cm³)		Void fraction (%)
	Theoretical	Measured
0	1.298	1.257	3.17
7.5	1.351	1.296	4.24
15	1.407	1.338	4.93

Water absorption rate

Figure 3 presents the behavior of the samples when immersed in water. The water absorption rate at different filler loadings and immersion time is presented in this figure. The figure shows that the water uptake increases with filler loading and also with the immersion time. Also, the unfilled epoxy shows the minimum water absorption rate, which is measured to be 0.85% after an immersion time of 168 h. There is a sudden increment in water absorption rate for epoxy/hemp fiber composites. For 40 wt% hemp fiber, the maximum water uptake rate is 7.83%. The increase in water uptake rate with the inclusion of hemp fiber is primarily due to the hydrophilic nature of the natural fiber. Hemp fiber is a natural fiber which can absorb more water, and because of that, its inclusion in epoxy resin results in higher absorption of water. Later, it is observed that with the inclusion of red brick dust, the water absorption further increases. For 7.5 wt% red brick dust, the water absorption rate slightly increases to 8.15%. The red brick is less hydrophilic than the hemp fiber, and because of that, there is a slight increase in water absorption rate. The water absorption rate increases with the content of red brick dust. For 15 wt% of red brick dust, the water absorption rate reaches 8.85%. It is further observed that the rate of water uptake increases with the immersion time as well. For the initial hour of immersion, the water uptake is high and then same decreases and reaches a saturation level, after a specific duration of immersion. Initially, the composite had more space to accommodate the moisture. Whereas with an increase in immersion time, the availability of space decreases, and the rate of water uptake decreases. The saturation time for neat epoxy is 120 h, whereas that for reinforced composites is 144 h.

Fig. 3.

Water absorption rate of epoxy-based composites.

Mechanical properties

The mechanical properties of the hemp-epoxy composites reinforced with RBD are studied, and the outcomes are presented in Figs. 4 and 5. The tensile strength and flexural strength of unfilled epoxy are 30.5 MPa and 36.8 MPa, respectively. With the inclusion of the hemp fiber mat, the strength increases appreciably. With 40 wt% hemp fiber, the tensile strength increases to 60.43 MPa, and the flexural strength increases to 89.26 MPa (Fig. 4). Contrary to that, the inclusion of red brick dust reduces the tensile and flexural strength of the sample as a function of filler loading. With 15 wt% of red brick dust, the tensile strength reduces to 52.38 MPa, and the flexural strength reduces to 82.65 MPa. A reduction in the measured properties is observed, likely caused by filler particles clustering within the epoxy matrix. Such clustering introduces localized stress concentration zones, making the composite more susceptible to failure under tensile or flexural loads. When these loads are applied, the regions with elevated stress tend to separate first, allowing cracks to initiate and travel through the matrix, ultimately leading to the composite’s breakdown. Similar observations are reported in the literature, which support this explanation⁵¹.

Fig. 4.

Tensile and flexural strength of the composites with RBD content.

Fig. 5.

Impact strength and hardness of the composites with RBD content.

Unlike the behavior seen in tensile and flexural strengths, the impact strength of the epoxy composite increases when hemp fiber mat and red brick dust are incorporated (Fig. 5). The base epoxy records an impact strength of 2.2 J/cm², which rises to 7.54 J/cm² with 40 wt% hemp fiber mats. This improvement occurs because the fibers introduce energy-absorbing mechanisms such as fiber pull-out, fiber–matrix debonding, and fiber breakage during impact. When red brick dust is added, the impact resistance improves even further. At 7.5 wt% red brick dust, the impact strength reaches 11.93 J/cm² and at 15 wt%, it increases to 18.41 J/cm². These values represent an enhancement of 144% compared to the epoxy/hemp composite. This rise can be attributed to the ability of red brick dust particles to hinder crack growth. Under impact loading, micro-cracks naturally form and spread through the epoxy matrix; however, the presence of brick dust forces the cracks to deflect, change direction, or slow down, resulting in a higher energy-absorbing capacity and overall improvement in impact strength.

The impact of the inclusion of hemp fiber and red brick dust in the epoxy matrix is also presented in Fig. 5. The hardness of neat epoxy is 19.92 HV. With 40 wt% hemp fiber mats, the hardness reaches 38.49 HV. The hardness increases noticeably with the inclusion of red brick dust, mainly due to the higher hardness of red brick dust than the polymer matrix. With the inclusion of 15 wt% of the red brick dust particulates, the hardness increases to 50.26 HV, registering an increment of 152% over neat epoxy. The enhancement in hardness can be attributed to the presence of rigid filler particles, which strengthen the epoxy matrix by limiting its plastic deformation. When a compressive force is applied, the distinct phases within the composite are compacted more tightly, leading to stronger interfacial contact between them. This improved contact facilitates better stress transfer across the interface. This results in increasing the material’s resistance to penetration. As a result, the overall hardness increases.

Sliding wear behavior

The coefficient of friction values of the epoxy composites consisting of various amounts of RBD, as obtained from the experimental investigation using the combination of test parameter levels as per the L₃₀ model of RSM, are presented in Table 5, and the respective specific wear rate values are presented in Table 6. To ascertain the influence of selected test parameters on the COF and SWR of the composites, the respective outcome values along with the corresponding combination of test parameter levels were analysed using the RSM model. The outcomes of the analysis in terms of analysis of variance are presented in Table 7 for COF and Table 8 for SWR. The P-values of the model for both COF and SWR are seen to be below 0.05, and the lack of fits are above 0.05, which confirms that both models are significant. Moreover, the ANOVA table for COF (Table 6) shows that all the control factors have significant effects on the COF as their p-value falls below 0.05. Similarly, Table 7 shows that all the control factors other than sliding distance have significant effects on the SWR. This is because at higher sliding distances, a polymer transfer film is formed on the counter face that restricts the direct contact of the composite pin with the counter face, thereby reducing the wear loss. During the analysis in RSM, the model established a relation between the selected input parameters presented in Table 7 for COF and Table 8 for SWR with the outcomes. The obtained relation for COF is provided as Eq. 9 and that for SWR as Eq. 10. The terms consisting of higher orders of the control factors and having negligible effects on the outputs were neglected from the equations.

Table 6.

Experimental and predicted specific wear rate values of the composites corresponding to the combination of test parameter levels as per the L₃₀ model of RSM.

Sl. no.	Sliding velocity (cm/s)	Normal load (N)	RBD content (wt%)	Sliding distance (m)	Specific wear rate (mm3/Nm)		Error (%)
					Experimental	Predicted
1	157	5.0	15.0	600	1.017	1.107	8.85
2	275	20.0	7.5	1500	2.499	2.573	2.96
3	393	5.0	0.0	600	2.437	2.598	6.61
4	393	20.0	15.0	600	2.588	2.754	6.41
5	275	12.5	7.5	2400	2.485	2.536	2.05
6	157	5.0	0.0	2400	3.376	3.523	4.25
7	393	20.0	0.0	600	3.997	4.036	0.98
8	275	12.5	7.5	600	2.125	2.335	9.88
9	393	5.0	15.0	600	2.062	1.895	8.10
10	275	12.5	7.5	1500	2.679	2.562	4.37
11	157	20.0	0.0	600	3.821	3.574	6.46
12	393	12.5	7.5	1500	2.976	2.894	2.76
13	157	5.0	0.0	600	3.042	3.022	0.66
14	275	5.0	7.5	1500	1.682	1.869	11.12
15	393	5.0	15.0	2400	2.638	2.498	5.31
16	157	12.5	7.5	1500	1.878	2.022	7.67
17	393	20.0	0.0	2400	4.807	4.331	9.90
18	275	12.5	7.5	1500	2.362	2.363	0.04
19	393	5.0	0.0	2400	2.737	2.894	5.74
20	157	20.0	0.0	2400	2.487	2.273	8.60
21	157	20.0	15.0	600	1.318	1.281	2.81
22	157	5.0	15.0	2400	0.935	1.005	7.49
23	275	12.5	0.0	1500	3.255	3.304	1.51
24	275	12.5	15.0	1500	1.647	1.810	9.90
25	275	12.5	7.5	1500	2.465	2.363	4.14
26	393	20.0	15.0	2400	3.016	3.355	11.24
27	157	20.0	15.0	2400	1.732	1.685	2.71
28	275	12.5	7.5	1500	2.481	2.363	4.76
29	275	12.5	7.5	1500	2.395	2.363	1.34
30	275	12.5	7.5	1500	2.488	2.363	5.02
31	275	12.5	7.5	1500	2.457	2.363	3.83

Table 7.

Analysis of variance table for the coefficient of friction.

Source	DF	Adj SS	Adj MS	F-Value	P-Value
Model	14	0.174951	0.012497	48.84	0.000
Sliding Velocity	1	0.029444	0.029444	115.08	0.000
Normal Load	1	0.064201	0.064201	250.92	0.000
RBD Content	1	0.034322	0.034322	134.14	0.000
Sliding Distance	1	0.006845	0.006845	26.75	0.000
Square	4	0.035032	0.008758	34.23	0.000
Sliding Velocity*Sliding Velocity	1	0.001809	0.001809	7.07	0.017
Normal Load*Normal Load	1	0.000656	0.000656	2.56	0.129
RBD Content*RBD Content	1	0.004665	0.004665	18.23	0.001
Sliding Distance*Sliding Distance	1	0.000380	0.000380	1.49	0.241
2-way interaction	6	0.005108	0.000851	3.33	0.025
Sliding Velocity*Normal Load	1	0.002401	0.002401	9.38	0.007
Sliding Velocity*RBD Content	1	0.001892	0.001892	7.40	0.015
Sliding Velocity*Sliding Distance	1	0.000240	0.000240	0.94	0.347
Normal Load*RBD Content	1	0.000400	0.000400	1.56	0.229
Normal Load*Sliding Distance	1	0.000144	0.000144	0.56	0.464
RBD Content*Sliding Distance	1	0.000030	0.000030	0.12	0.735
Lack-of-fit	10	0.002864	0.000286	1.40	0.354

Table 8.

Analysis of variance table for coefficient of friction.

Source	DF	Adj SS	Adj MS	F-Value	P-Value
Model	14	18.3116	1.30797	9.05	0.000
Sliding Velocity	1	3.4252	3.42522	23.69	0.000
Normal Load	1	2.2324	2.23238	15.44	0.001
RBD Composition	1	9.3976	9.39756	65.01	0.000
Sliding Distance	1	0.1812	0.18120	1.25	0.279
Square	4	0.5016	0.12539	0.87	0.505
Sliding Velocity*Sliding Velocity	1	0.0235	0.02352	0.16	0.692
Normal Load*Normal Load	1	0.0518	0.05181	0.36	0.558
RBD Composition*RBD Composition	1	0.1247	0.12470	0.86	0.367
Sliding Distance*Sliding Distance	1	0.0139	0.01391	0.10	0.760
2-way interaction	6	2.5737	0.42895	2.97	0.038
Sliding Velocity*Normal Load	1	0.7859	0.78588	5.44	0.033
Sliding Velocity*RBD Composition	1	1.0252	1.02516	7.09	0.017
Sliding Velocity*Sliding Distance	1	0.4837	0.48372	3.35	0.086
Normal Load*RBD Composition	1	0.1440	0.14402	1.00	0.333
Normal Load*Sliding Distance	1	0.0410	0.04101	0.28	0.602
RBD Composition*Sliding Distance	1	0.0939	0.09394	0.65	0.432
Lack-of-fit	10	2.2515	0.22515	22.00	0.121

9

10

These equations are used to predict the COF and SWR of the composites for the given combination of test parameter levels. The predicted values of COF corresponding to the combination of test parameter levels are presented in Table 5, and those of the values for SWR are presented in Table 6. The comparisons of experimentally obtained and predicted values of COF are plotted in Fig. 6, and of SWR are plotted in Fig. 7. It is evident from Tables 5 and 6, as well as Figs. 6 and 7, that the predicted and experimental values coincide well because the error never went beyond 5%. Therefore, any combination of test parameter levels within the experimental domain can be used to make additional predictions using these equations.

Fig. 6.

Comparison of experimental and predicted values of coefficient of friction.

Fig. 7.

Comparison of experimental and predicted values of specific wear rate.

The standardized effects of each control factor and their interactions on the COF of the composites are shown in Fig. 8. The average COF of the composites is found to be most significantly impacted by normal load, followed by sliding velocity and RBD content. Sliding distance also has a noticeable effect on the COF, but, as compared to the other control factors, its effect is marginal. A few interactions associated with RBD content and sliding velocity were also noticed to have an observable effect on the COF, as the effect lines representing them are above the average standardized effect line (2.12), as witnessed from the figure. Similarly, Fig. 9 shows the standardized effects of the individual control factors and their interactions on the SWR of the composites. RBD composition, sliding velocity and normal load in the respective sequence are observed to have a detrimental impact on the SWR. As discussed above, sliding distance has a negligible impact on the SWR and is thus removed from the histogram. Only a few interactions associated with sliding velocity are observed to have a notable impact on the SWR. Figures 10 and 11 are the 3-dimensional surface plots displaying the combined effects of control factors on the COF and SWR of the composites. The approximate variation in COF and SWR at any combination of two different control factors within the experimental domain can be seen from these plots. From the experimental findings and predictions, it can thus be concluded that among all the control factors, RBD content has the maximum significant effect on both COF and SWR.

Fig. 8.

Standardized effects of individual control factors and their interactions on the coefficient of friction.

Fig. 9.

Standardized effects of individual control factors and their interactions on the specific wear rate.

Fig. 10.

Surface plot displaying the combined effects of control factors on the coefficient of friction of the composites.

Fig. 11.

Surface plot displaying the combined effects of control factors on the specific wear rate of the composites.

Predictive modelling using machine learning

Although RSM is used to predict the specific wear rate and coefficient of friction of the RBD-filled hybrid composites, it has limitations in predicting the tribological properties of these composites under varying conditions and beyond the experimental domain. Therefore, to further improve the accuracy of prediction, three machine learning models are employed for analysis, learning and prediction of wear and friction data obtained from experiments.

Evaluation of model performance

The specific wear rate and coefficient of friction of hemp-epoxy composites filled with RBD are predicted using each machine learning model. The optimized parameters in Table 3 are used to predict the output parameters, and the anticipated outcomes are compared with the corresponding experimental values. For each of the three ML models, Fig. 12 displays the comparison curve between the experimental and anticipated specific wear rates. It is clear from Fig. 12a that the DT model demonstrated weaker performance, with less accuracy and wider deviations from the zero-line error. Whereas the RF and GBM models demonstrate more precise prediction than the DT model with a greater number of data points near the zero-line error, as observed in Fig. 12b and c, respectively. This happened because RF and GBM use a combination of multiple decision trees and therefore perform better than a single decision tree. Moreover, the GBM model outperforms the RF model with good agreement between the experimental and predicted outcomes. This is because GBM iteratively optimizes residual errors, enhancing predictive performance by sequentially eliminating weak learners. Similar observations have been reported by Hasan et al.³¹ and Whang et al.³² for the prediction of the wear rate of the metal-matrix composites. The experimental vs. predicted COF for RBD-filled hybrid composites for DT, RF and GBM models are shown in Fig. 13. Similar to the case of specific wear rate, the GBM model is the standalone model among all the machine learning models.

Fig. 12.

Comparison of experimental and predicted specific wear rates for RBD-filled hybrid composites using (a) DT model, (b) RF model and (c) GBM model.

Fig. 13.

Comparison of experimental and predicted COF for RBD-filled hybrid composites using (a) DT model, (b) RF model and (c) GBM model.

Comparison of ML models

The effectiveness of ML models is assessed using different performance criteria as discussed in Sect.  2.2.3. Among these, R² is considered a key indicator for assessing regression model performance. This metric, which ranges between 0 and 1, represents the proportion of variance in predicted outcomes explained by the independent variables. An R² value between 0.7 and 0.9 signifies a good level of model performance, while values exceeding 0.9 indicate an exceptionally reliable and well-performing model.

The performance metrics for the prediction of SWR and COF of RBD-filled hemp-epoxy composites are shown in Table 9. It is clearly observed that for the prediction of SWR, the R² value ranges from 0.83698 to 0.95422 for all three models. The value of the coefficient of determination of all three models is greater than 0.8, indicating that all the machine learning models performed fairly well. The GBM model is the standalone model with R² = 0.9542, MSE = 0.0304, RMSE = 0.1745, MAE = 0.1130 and MAPE = 4.6261. The random forest model stands as the second-best model with R² = 0.9174, MSE = 0.0549, RMSE = 0.2343, MAE = 0.1378 and MAPE = 5.6231, followed by the decision tree model with R² = 0.8369, MSE = 0.1084, RMSE = 0.3293, MAE = 0.2753 and MAPE = 11.2126. These results indicate that the GBM model showed the highest accuracy with the least errors and, therefore, is the best performing model in predicting the specific wear rate of RBD-filled hemp-epoxy composites.

Table 9.

Performance of all ML models.

ML model	R 2	MSE	RMSE	MAE	MAPE
Specific wear rate
DT	0.8369	0.1084	0.3293	0.2753	11.2126
RF	0.9174	0.0549	0.2343	0.1378	5.6231
GBM	0.9542	0.0304	0.1745	0.1130	4.6261
Coefficient of friction
DT	0.8498	0.00087	0.0295	0.0229	3.9336
RF	0.9808	0.00011	0.0105	0.0069	1.1548
GBM	0.9879	0.000069	0.0084	0.0049	0.8259

Table 9 also shows the performance metrics for the prediction of the COF of hybrid composites. The R² value varied from 0.8498 to 0.9879 for all three machine learning models. The acquired MSE (0.000069–0.00087), RMSE (0.0084–0.0295), MAE(0.0049–0.0229) and MAPE (0.8259–3.9336) of all three models remain considerably low, which indicates that these models can effectively predict the COF of the hybrid composites. Further, among the three models, the GBM model outperforms other models since it learns from the mistakes of preceding trees and makes corrections in the subsequent trees, resulting in improved accuracy^50,52.

Feature importance

The feature importance signifies the contribution made by each input parameter during the prediction of the output response. The feature importance plot can only be plotted for decision tree-based models using the scikit-learn library in Python. In this plot, each feature is assigned a feature score, which varies between 0 and 1. While a feature score of zero denotes no influence, a higher number suggests a stronger impact on output prediction. If each one of the input variables has a non-zero score, then all the input features contributed significantly to the prediction of the output. Figure 14 presents the feature importance plot of each input variable of different ML models for predicting the specific wear rate of RBD-filled hybrid composites. The feature scores of the decision tree model for the prediction of SWR indicate that the RBD composition and sliding velocity are the most important influencing variables, with feature scores of 0.56251 and 0.37699, respectively. The normal load contributed less with a feature score of 0.06049. Similar results are also observed from the feature scores of random forests and GBM models. In all three models, the sliding distance has a poor correlation with feature scores of 0.00000, 0.082071 and 0.012860, respectively, indicating that it has minimal effect on the SWR of RBD-filled hemp-epoxy composites.

Fig. 14.

Relative importance of input features for SWR prediction using (a) DT model, (b) RF model and (c) GBM model.

The importance plot obtained in the case of DT, RF and GBM models for predicting the COF (Fig. 15) shows that for all three models, the normal load has the highest feature score and therefore has contributed prominently in predicting COF, followed by RBD composition and sliding velocity with relatively low scores. In each of the three models, the sliding distance has minimal effect on the prediction of the COF of hemp-epoxy-based hybrid composites.

Fig. 15.

Relative importance of input features for COF prediction using (a) DT model, (b) RF model and (c) GBM model.

Morphologies of worn composite surfaces

The analysis of experimental findings using the L₃₀ model of response surface method resulted in the control factors normal load, RBD loading and the velocity of sliding being significant in affecting the coefficient of friction and wear rate of the hybrid composites. Thus, the morphologies of the composite worn surfaces with the optimum combination of these control factors, as displayed in Fig. 16, are studied to ascertain the predominant mechanisms instigating the loss because of wear to the composites. Figure 16a and b display the micrographs of the epoxy/hemp fiber composites without any red brick dust. The worn surface displayed is for a composite undergoing wear for the sliding velocity of 157 cm/s under a 5 N load for 2400 m. The formation of wear tracks, plastic deformation, crater, micro-cracks, etc., pullout of fibers as well as breaking and removal of fibers, as witnessed from the micrographs, are the wear loss mechanisms for the composites worn with this combination of test parameters. The pullout of fibers from the matrix explains the weak bonding of the fiber with the matrix at the interface. In some cases, the improper wetting of the fiber with the matrix also restricts the participation of some fiber strings in the composite making, which is later removed from the composite by pullout during the sliding operation. Moreover, during sliding with some axial load, the fibers are acted upon by the repeated axial as well as tangential load, causing the loosening and easy removal of the fibers, leading to higher wear loss. Whereas the easy removal of fibers reduced the coefficient of friction. Also, during the removal of fibers, some portion of the matrix comes out from the composite body, which later forms a thin polymer layer on the wear region and covers the exposed fiber ends. Thus, the contact of the counter body and fiber reduces, thereby declining the coefficient of friction. Figure 16c and d depict the micrographs of the worn surface consisting of 7.5 wt% of RBD when worn with a velocity of 275 cm/s with a 12.5 N load for 1500 m. The red brick particles below the matrix surface are exposed because of the elimination of the top layer of the matrix. The fiber ends exposed from the composite body confirm the overlapping of the matrix layer on the fibers due to the repeated sliding under load. Figure 16d shows the breaking and removal of filler particles, explaining their participation in resisting the wear loss. The tracks of wear highlighted on the surface of the composite are developed due to the ploughing of the composite. The micrographs confirm the development of wear tracks, removal of wear debris and fibers, as well as the breaking and amputation of red brick dust as the possible mechanisms producing the wear loss.

Fig. 16.

SEM micrographs of the worn composite surfaces displaying the wear mechanisms (a,b) unfilled composite with 157 cm/s, 5 N, 2400 m, (c,d) composite consisting of 7.5 wt% RBD with 275 cm/s, 12.5 N, 1500 m, (e,f) composite consisting of 15 wt% RBD with 393 cm/s, 20 N, 2400 m, (g,h) composite consisting of 15 wt% RBD with 275 cm/s, 12.5 N, 1500 m.

Figure 16e and f illustrate the micrographs of the worn composite surface consisting of 15 wt% of RBD while tested with a velocity of 393 cm/s for 2400 m with 20 N loading. From the micrographs, the debonding of the fiber from the matrix and the fracture of the fiber can be clearly seen. During sliding, the matrix deforms and removes faster as compared to the fibers and thus the fibers endure a huge fraction of the total applied axial load and developed tangential load. Thus, the debonding of fiber with matrix, which is also called the interfacial fatigue, occurs near the fiber boundary. Moreover, the wear of the top matrix layer causes the improper distribution of stress among the hemp fiber and epoxy, as a result of which the fiber fracture occurs, which explains that the induced stress is higher than the maximum stress bearing capacity of the fibers. Furthermore, there is a difference in modulus between the fiber and the matrix, due to which the matrix deforms faster than the fiber, causing fiber-matrix de-bonding. The micrographs of the worn composite surfaces consisting of 15 wt% of RBD slide with a velocity of 275 cm/s for 1500 m with a load of 12.5 N are illustrated in Figs. 16 (g) and (h). At lower magnification (Fig. 16g), the micrograph shows a smooth appearance of the worn surface of the composite with few thinner wear tracks as compared to the other compositions. The micrograph confirms the proper incorporation and distribution of RBD particles within the composite. In general, the wear loss of polymer-based composites is significantly oriented to three different wear mechanisms, including micro-cutting, cracking and ploughing. The wear loss to the composites at this combination of test parameters, as evident from the micrographs is mainly due to the general wear loss mechanisms. The debonding of the matrix shown on the micrograph (Fig. 16h) is due to the propagation of micro-cracks, which may be initiated near the fiber boundary due to the repeated sliding at higher velocity. The presence of a large amount of filler particles on the worn surface explains the reduced wear loss and increased coefficient of friction as compared to the other composites. Thus, the formation of wear tracks and micro-cracks, removal of fibers, de-bonding and removal of matrix, etc., can be listed out to be the pronounced wear loss mechanisms at this combination of test parameters.

Conclusions

Epoxy-hemp fiber composites containing micro-sized RBD particles were fabricated successfully using a conventional hand lay-up process. The composite’s porosity exhibits an increase corresponding to the rise in RBD, with the maximum void content recorded at 4.93%. The hydrophilic hemp-epoxy composites’ water absorption behavior is found to be significantly impacted by the addition of RBD, with water absorption rates ranging from 7.83% to 8.85%. Mechanical properties showed contrasting trends, where tensile and flexural strengths decreased by approximately 13.3% and 7.4%, respectively, at 15 wt% RBD, whereas impact strength and hardness improved substantially by 144% and 152%, respectively. Their sliding wear behavior revealed that the red brick dust content and sliding velocity are the key parameters affecting the specific wear rate of the composite, whereas the normal load mainly governs the coefficient of friction. The developed RSM models showed strong agreement with experimental data, with prediction errors generally within ± 5%. The machine learning analysis further established that the gradient boosting machine (GBM) model provides the highest prediction accuracy, achieving R² values exceeding 0.95 for specific wear rate and 0.98 for coefficient of friction, outperforming decision tree and random forest models. The observed reduction in tensile and flexural strength with increasing filler content may be attributed to possible particle agglomeration and the formation of stress concentration regions within the matrix. Similarly, the improvement in impact strength and hardness is likely associated with the resistance offered by rigid filler particles to crack propagation and localized deformation. The enhancement in wear resistance may be explained by the presence of filler particles that contribute to load-bearing and reduce material removal during sliding. The key mechanisms of composite wear loss are micro-cracking, fiber fracture, crater, wear debris and fiber pullout under various sliding situations according to the morphologies of worn surfaces.

Author contributions

Devasri Fuloria: Formal analysis; Pravat Ranjan Pati: Writing - Review & Editing, Investigation; Sandip Kumar Nayak: Conceptualization, Visualization; Sourav Kumar Mahapatra: Software, Data Curation; Alok Agrawal: Methodology, Validation; Abhilash Purohit: Resources, Project administration; Gaurav Gupta: Writing - Original Draft, Supervision.

Funding

Open access funding provided by Vellore Institute of Technology.

Data availability

The datasets used and/or analysed during the current study available from the corresponding author on reasonable request.

Declarations

Competing interests

The authors declare no competing interests.