Machine-Learning-Driven Analysis of Wear Loss and Frictional Behavior in Magnesium Hybrid Composites

Abstract

The wear loss and frictional characteristics of magnesium-based hybrid composites reinforced with boron carbide (B₄C) particles and graphite filler were the main subjects of the investigation. Key parameters, including reinforcement content (0–10 wt%), applied load (5–30 N), sliding speed (0.5–3 m/s), and sliding distance (500–3000 m), were varied. Data-driven machine learning (ML) algorithms were utilized to identify complex patterns and predict relationships between input variables and output responses. Five distinct machine learning algorithms, Artificial Neural Network (ANN), Random Forest (RF), K-Nearest Neighbor (KNN), Gradient Boosting Machine (GBM), and Support Vector Machine (SVM), were employed to analyze experimental tribological data for predicting wear loss and coefficients of friction (COFs). The performance evaluation showed that ML models effectively predicted friction behavior and wear behavior of magnesium-based hybrid composites using tribological test data. A comparison of model performances revealed that the Gradient Boosting Machine (GBM) provided superior accuracy compared to other machine learning models in predicting both wear loss and the coefficient of friction. Additionally, feature importance analysis indicated that the graphite weight percentage was the most significant influence in predicting the coefficient of friction and wear loss characteristics.

1. Introduction

The demand for high-performance and lightweight materials in the automotive and aerospace sectors has led to substantial research and development activities targeted at developing advanced magnesium composites and optimizing their manufacturing procedures [1]. Magnesium composite materials are particularly appealing due to their ability to be tailored to enhance specific material properties. Currently, magnesium and its alloys are utilized for weight reduction for exotic parts due to its lower density as compared to aluminum [2]. However, magnesium and its alloys’ applications are restricted for their poor resistance to creep at elevated temperatures, along with their limited strength, modulus, and inadequate wear and corrosion resistance [3]. Wear is one of the most common issues in industrial components, which shortens the lifespan of magnesium-based parts, making magnesium unsuitable for components like gears, pistons, cylinders, and bearings [4,5,6]. Although wear primarily affects the surface, it can significantly impair the mechanical performance of engineering components leading to reduced precise tolerances, degraded quality of surface finishes, and structural failure. To address these challenges, various particle reinforcements have been added to magnesium and its alloys, with reports detailing the improved properties of these composites [7,8]. In the literature, various reinforcement particles such as borides (TiB₂ [9,10,11,12], ZrB₂ [13,14]), carbides (B₄C [15,16,17], SiC [18,19,20], TiC [21,22]), oxides (Al₂O₃ [23,24], TiO₂ [25,26]), nitrides (TiN [27,28], BN [29,30]), carbon-based materials (GNPs [31,32], CNT [33,34]), and waste materials (e.g., eggshell [35,36], fly ash [37,38]) have been extensively applied in Mg matrix composites. Due to such advancements, magnesium composites present a strong alternative to aluminum, iron, and steel for many applications. Hybrid composites incorporate at least two reinforcement materials into the matrix, and exhibit significantly enhanced properties, providing a high level of design flexibility depending on the type and size of particles used. Boron carbide stands out as a material of interest due to its low density (2.51 g/cm³), remarkable chemical resistance, and exceptional hardness. The addition of boron carbide (B₄C) to a magnesium matrix considerably enhances the composite’s interfacial bonding, flexural strength, hardness, and wear resistance. Additionally, a comprehensive review of the literature indicates that the addition of solid lubricants like graphite, hexa-boron nitride, molybdenum sulphide, or other similar materials significantly enhances wear resistance [39]. Graphite, in particular, is used as a solid lubricant due to its availability and has shown superior results in metal matrix composites (MMCs) reinforced with B₄C, Al₂O₃, and SiC.

Despite numerous attempts to define rules of wear and related friction, tribology remains predominantly empirical due to its complexity. However, the large volume of data on surface interactions, material wear, and friction from tribological studies has paved the way for analysis based on data-driven approaches. The recent progress in computational power has accelerated the development of data-driven analytical methods, offering innovative insights. AI and ML techniques are now useful in “Big Data” analysis to discover patterns and correlations that are difficult to detect using conventional methods. Machine learning techniques have previously been applied in wear loss experiments across various studies. In research conducted by Batista et al. [40], the linear regression (LR) algorithm was employed to predict wear resistance. Another study utilized the support vector regression (SVR) method to predict the thickness of the abrasion-resistant chrome layer [41]. Wang et al. [42] investigated the forecasting of tool wear using the Gaussian-mixture-based regression approach. ANN is one of the most widely employed techniques for forecasting wear loss [43]. A study by Foster forecasted wear loss through the ANN model or algorithm [44]. Tan et al. applied both SVR and ANN models to forecast wear loss and friction coefficients [45]. Although disciplines such as biology, chemistry, and medicine have widely adopted AI and ML procedures as trusted tools, their application in materials science remains a relatively recent development [46]. Recent progress in applying machine learning to predict material and tribological properties [47,48,49,50,51,52,53] has opened up exciting new possibilities in both tribology and materials science. Conventional analysis, which typically emphasizes individual experimental results, is often insufficient for achieving a comprehensive insight into tribological behavior, especially because of the complex factors influencing friction and wear. Machine learning models can be developed on investigational tribological information to forecast behavior across various material mixtures and influencing factors. Sophisticated machine learning models like SVM and ANN can detect patterns within the dataset and adapt the learning processes to generate correct predictions, providing broader insight into tribological behavior.

This study begins with a conventional examination of the wear and friction behavior of magnesium-based hybrid composites concerning tribological test parameters. We then evaluate the performance of five different machine learning approaches (ANN; SVM; KNN; RF; and GBM) in forecasting tribological nature and provide a comparative study of the various material and other input factors that impact the tribological properties of these composites.

2. Materials and Methods

The technique of solid metallurgy (powder metallurgy) employed to manufacture both composites and hybrid composites and wear testing methods are depicted in Figure 1.

Magnesium (purity > 98%) with an average particle size of 45 μm was used as the base (matrix) material. Graphite (Gr) and boron carbide (B₄C), both with a purity exceeding 99% and an average particle size of 37 μm, were used as reinforcement materials. The study involved the fabrication of eight different composite types, as detailed in

Table 1. Reinforcement percentage of the hybrid composite.

Sl No	Mg Powder (wt%)	B4C Powder (wt%)	Gr. Powder (wt%)
1	100	0	0
2	95	5	0
3	90	10	0
4	95	0	5
5	90	0	10
6	90	5	5
7	85	10	5
8	80	10	10
9	85	5	10

.

The materials were mixed using planetary ball milling at 150 rpm for 4 h, with acetone as a wetting agent to prevent oxidation. The blended powders were then heated at 120 °C for 1 h to remove moisture. Next, the dried powder was pressed at 620 MPa in a hydraulic press to form 13 mm diameter green samples (billets), with the die wall lubricated using zinc stearate. The compacted samples were sintered at 550 °C for 30 min in a programmable furnace under an argon atmosphere (200 mL/min flow rate) to prevent oxidation. The samples were then allowed to cool to room temperature within the furnace. Wear tests on the top surface of the composite samples have been carried out with a reciprocating tribometer (Ducom CM-9104).

To ensure the reliability of the input parameters used in the machine learning models, we have carefully evaluated the accuracy and confidence intervals of the measured data. The applied load, sliding speed, and sliding distance were precisely controlled using the instrument’s built-in load cell and speed regulator, ensuring repeatability. The wear loss measurements were obtained using a precision balance with an accuracy of ±0.001 g, while the coefficient of friction (COF) was recorded using a tribometer with an accuracy of ±0.01. Each measurement was repeated five times, and the mean values along with 95% confidence intervals were calculated to quantify measurement uncertainty. The standard deviation remained within an acceptable range, ensuring minimal variability in the experimental data. Furthermore, outlier detection and data normalization were performed during preprocessing to enhance data quality and avoid biases in model training. These steps reinforce the reliability of the dataset, ultimately improving the robustness of the ML predictions.

2.1. Mechanisms of Wear and Friction in Magnesium, Boron Carbide and Graphite Hybrid MMCs

Friction and wear are essential components of tribological research as they play a major role in material deterioration and energy loss in dynamic systems. Determining the wear and friction mechanisms in multi-phase metal matrix composites, like those made of graphite, boron carbide, and magnesium, is still a difficult task. The wear loss and friction mechanisms in magnesium hybrid composites are examined in this section.

Friction is the resistance between two sliding surfaces and is quantified by the coefficient of friction (COF), denoted as μ [54,55]. The COF is determined using an equation that relates the frictional force (F_f) to the normal load (N).

$$\mathsf{\mu }={\displaystyle \frac{{\mathrm{F}}_{\mathrm{f}}}{\mathrm{N}}}={\displaystyle \frac{{\mathrm{F}}_{\mathrm{a}}+{\mathrm{F}}_{\mathrm{d}}}{\mathrm{N}}}={\displaystyle \frac{{\mathrm{F}}_{\mathrm{a}}}{\mathrm{N}}}+{\displaystyle \frac{{\mathrm{F}}_{\mathrm{d}}}{\mathrm{N}}}={\mathsf{\mu }}_{\mathrm{a}}+{\mathsf{\mu }}_{\mathrm{d}}$$

(1)

The coefficient of friction (COF) consists of two components: μ_a, which arises from adhesive forces at the molecular level in metallic, covalent, ionic, and van der Waals bonds [56], and μ_d, which results from microscale and macroscale deformation of surface irregularities, influenced by mechanical and material properties. The adhesion component of friction depends on the actual contact area between asperities on sliding surfaces.

Hybrid Metal Matrix Composites (MMCs), such as magnesium reinforced with boron carbide (B₄C) and graphite (Gr), exhibit distinct tribological properties. These composites provide enhanced wear resistance, lower friction, and improved mechanical performance compared to single-reinforcement composites. Understanding their wear and friction behavior requires knowledge of tribology and materials science.

In Mg-based hybrid MMCs, friction is influenced by two main factors, adhesion and plowing, where the metal matrix contributes to adhesive friction, but the inclusion of B₄C and Gr alters this interaction. Graphite acts as a solid lubricant, forming a transfer film that reduces adhesion and friction, while B₄C enhances hardness and deformation resistance, minimizing surface damage. In the plowing effect, hard B₄C particles penetrate the softer sliding counter face, increasing friction. The extent of this effect depends on B₄C hardness and applied load, as described by the following equation.

$${\mathrm{F}}_{\mathrm{p}}={\displaystyle \frac{\mathrm{k}·{\mathrm{A}}_{\mathrm{p}}}{2}}$$

(2)

where F_p = plowing friction force, k = hardness of the counter face, and A_p = projected area of the plowing particle.

Deformation is another factor affecting wear, occurring when surface asperities bend under load. Due to magnesium’s softness, it undergoes significant plastic deformation, but B₄C reinforcement and graphite lubrication help mitigate this effect. Wear in magnesium–B₄C–graphite hybrid MMCs is influenced by abrasive, oxidative, and delamination wear mechanisms [57]. Abrasive wear occurs when B₄C particles either embed into the material or act as free-moving debris, causing surface scratches and material loss. Oxidative wear results from high surface temperatures forming oxide layers, which may either protect the material or break down into wear debris. Graphite’s lubricating properties help lower surface temperatures and reduce oxidative wear. Delamination wear occurs due to subsurface cracks caused by cyclic loading and sliding. The presence of B₄C improves resistance to crack initiation and growth, while graphite reduces surface stresses and minimizes crack formation.

2.2. ML to Predict Tribological Characterization

ML models are developed to predict the tribological behavior of magnesium hybrid composites under dry conditions. This section covers data collection, pre-processing, parameter optimization, and performance enhancement methods for the ML models.

2.2.1. Collection of Data

Since a sizable and pertinent dataset improves predicted accuracy, data collecting is essential to creating a successful data-driven machine learning model. However, it is expensive and time-consuming to produce enough tribological data through trials. Furthermore, the predicted performance of the model on a variety of datasets may be impacted by biases introduced by depending solely on data from a single experimental scenario. Data on material deterioration (wear loss) and friction coefficient were gathered from dry-condition tests using a steel counter-face to address this. The coefficient of friction (COF) and wear loss were predicted using a tribological dataset of 154 sample data points.

2.2.2. Key Input Variables and Output Metrics

As shown in Table 1, the ML models or algorithms need nine material variables as input parameters. Three tribological test factors are also taken into account: sliding distance, sliding speed, and applied load. Every input variable is dealt with as a numerical number, including material compositions and other tribological test factors. The coefficient of friction (COF) and wear loss are used as the target outputs or responses.

2.2.3. ML Models or Algorithms

In supervised ML, regression analysis is employed to establish the relationship between input variables and output metrics, facilitating the generation of numerical results. This study utilizes five ML models/algorithms, ANN, KNN, RF, SVM, and GBM, to analyze tribological data of magnesium composites, aiming to predict tribological behavior by establishing relationships between input and output parameters. The analyses are performed using MATLAB R2024a.

K Nearest Neighbor (KNN): KNN predicts outcomes for a new data point by referencing the closest data points (neighbors) in the training dataset. The parameter “k” indicates the number of closest data points used for prediction, while the parameter “n” impacts the complexity of the approach, requiring optimal selection based on data characteristics. KNN performance also depends on whether uniform or proximity-based weights are applied to the neighbors.

Support Vector Machines (SVMs): Kernelized SVMs work well for both regression and classification tasks, particularly when dealing with high-dimensional, complicated datasets. SVM manages non-linear correlations between features by converting input data into higher-dimensional space using kernel functions, which allows for variable decision limits. The regularization parameter (C) and kernel coefficient (gamma) affect the model’s complexity and performance, and they must be modified in tandem to maximize outcomes.

Artificial Neural Networks (ANNs): Based on the human brain’s architecture, ANN uses interconnected neuron layers to recognize patterns. ANNs capture complex data relationships, with optimization parameters like learning rate, activation function, and the number of hidden layers and neurons. This study uses Multi-Layer Perceptron (MLP) regression to predict wear loss and the coefficient of friction (COF).

Random Forest (RF): RF works with an ensemble of decision trees to process large datasets with many input variables, improving predictive accuracy through averaging predictions from individual trees. This helps mitigate overfitting, making RF more accurate than many other machine-learning models. Optimization involves tuning parameters such as max_features and n_estimators.

Gradient Boosting Machine (GBM): GBM also employs an ensemble of decision trees, with each tree addressing a subset of the dataset. Errors from earlier trees are corrected in subsequent iterations, enhancing prediction accuracy. The model’s complexity is influenced by the learning rate and the number of boosting iterations (n_estimators), requiring simultaneous adjustment to improve performance.

Data cleaning (pre-processing) plays a crucial role in developing a successful ML model. Key steps include managing outliers, handling missing data, splitting the dataset into training and validation subsets, and normalizing the data through scaling methods. Anomalies and gaps in the dataset were identified and resolved. To avoid any biases or unwanted trends, data shuffling was performed. Using MATLAB R2024a, the data were split into two exclusive sets: a training set and a test set. The training set was used to construct and validate the ML models, while the test set was employed to evaluate model performance on new data. Using a typical data splitting technique, 25% of the data were set aside for testing and 75% for training. In the training set, 25% was set aside for validation and 75% for training. The machine learning methods used 86, 29, and 39 data points (out of 154 total data points) for training, validation, and testing to forecast wear loss and the coefficient of friction (COF), respectively. To improve the performance of machine learning algorithms, feature scaling, the last stage of data pre-processing, brings all variables into line with a uniform scale.

To ensure data reliability, preprocessing involved outlier detection using the interquartile range (IQR) method, where data points falling outside 1.5 times the IQR were considered potential outliers. Missing data, if present, were handled using linear interpolation to maintain dataset integrity. The dataset was systematically split into training (75%) and test (25%) sets, with an internal validation set (25% of the training data) used for hyperparameter tuning and model refinement. The model performance was assessed using key evaluation metrics, including RMSE, R², and MAE, ensuring a rigorous validation process. This structured approach enhances model reliability by minimizing biases and ensuring meaningful generalization.

2.2.4. Optimization of Different ML Models’ Parameters

To get the best prediction performance, several ML models must be carefully improved. Section 3 discusses the parameters that need to be optimized for each model. Cross-validation techniques and parameter modification are used to determine the best forecasting model configurations. The prediction models are run several times and various parameter ranges are tried during the optimization procedures.

Table 2. Optimization parameters for wear loss.

Model	Particular Parameters
ANN	Alpha = 0.01; Activation Function: tansig (1st hidden layer); logsig (2nd hidden layer); purelin (output layer); Hidden layers = (10, 5)
KNN	n_neighbors = 10; Weights = ‘uniform’
RF	n_estimators = 100; max_features = 2
SVM	Kernel Function: linear; Kernel Scale (Gamma): 969.092; C (Box Constraint): 359.5019
GBM	n_iestimators: 100; Max Depth: 1; Learning Rate: 0.1

and

Table 3. Optimization parameters for COF.

Model	Particular Parameters
ANN	Alpha = 0.01; Activation Function: tansig (1st hidden layer), logsig (2nd hidden layer), purelin (output layer); Hidden layers = (8, 4)
KNN	n_neighbors = 10; Weights = ‘uniform’
RF	n_estimators = 120; max_features = 3
SVM	Kernel Function: polynomial; Kernel Scale (Gamma): 1; C (Box Constraint): 1
GBM	n_iestimators: 100; Max Depth: 1; Learning Rate: 0.1

list the characteristics that are chosen and perform the best in estimating wear loss and the coefficient of friction. The grid search strategy, another optimization tool used in conjunction with the ML algorithms, is implemented using MATLAB scripts. The grid search technique evaluates model performance with various parameter settings, identifying the configurations that yield the most accurate results.

3. Results and Discussion

This section discusses the variables that influence the tribological behavior of magnesium composites and the results of ML studies for predicting wear loss and coefficient of friction (COF). Additionally, it evaluates the effectiveness of different ML models and examines the importance of various features in predicting these parameters.

3.1. Factors Affecting the Tribological Behavior of Magnesium Composites

The wear loss of composite samples is reduced by the inclusion of boron carbide (B₄C) reinforcements, as shown in Figure 2, and wear loss is further reduced by the addition of graphite. Generally, composites with higher hardness exhibit better wear resistance, with both wear resistance and hardness of magnesium increasing as the B₄C content rises. However, the addition of B₄C and graphite to magnesium lowers hardness while simultaneously reducing wear. Analysis of the data in Figure 1 indicates that the inclusion of graphite (a solid lubricant) significantly reduces wear between sliding surfaces. Despite this, the combination of 10 wt% B₄C and 10 wt% graphite in a hybrid composite results in higher wear loss than the combination of 10 wt% B₄C and 5 wt% graphite, even though the additional graphite contributes to improved lubrication. These findings suggest that a higher proportion of graphite may increase localized pressure, leading to matrix cracking and distortion during sliding.

Figure 2a shows the wear loss in a sample tested under varying load conditions, ranging from 10 to 25 N, helping to understand the material’s behavior under solid surface interactions, particularly at higher loads. It is evident that wear loss increases as the load increases. The Mg–10B₄C–5Gr hybrid composite exhibits the lowest wear loss at 25 N. Figure 2b presents wear loss data for composites at different sliding speeds, from 0.5 to 3.0 m/s, revealing that wear loss decreases with increasing sliding speed, regardless of reinforcement content. Higher speeds may initially separate little wear debris containing graphite which may reduce friction and lead to reduced wear loss. Additionally, the shorter contact duration between the pin and disc at higher speeds contributes to reduced wear loss. Figure 2c illustrates wear loss at varying sliding distances, from 500 m to 3000 m. The increase in distance correlates with extended sliding over a hard surface, raising interface temperature, which weakens the specimens and leads to a higher wear rate. The analysis indicates that wear loss decreases significantly with higher graphite content at increased sliding distances, suggesting improved wear resistance due to the synergistic effects of B₄C and graphite reinforcements.

Figure 3 illustrates the variation of the coefficient of friction (COF) with different reinforcements. The COF increases with the addition of the harder B₄C reinforcement. It is also observed that the COF of the sample reinforced with graphite particles is lower than that of the magnesium matrix reinforced with B₄C particles. Among the two reinforcements, the Mg–10Gr composite shows the lowest COF, while the Mg–10B₄C composite exhibits the highest COF. As shown in Figure 3a, the COF increases with load (from 5 N to 30 N) for all composite combinations. The even distribution of reinforcing particles helps manage the load, reducing the contact area at the disc–pin interface and lowering the COF. In contrast, for pure Mg, the increased contact between the counter disc and Mg pin causes the COF to rise with increasing load. Under acceptable load conditions, the COF of Mg-Gr composites is lower than that of Mg-B₄C and Mg-B₄C-Gr hybrid composites. Figure 3b demonstrates the variation of COF for composites over a range of sliding distances, from 500 m to 3000 m. At shorter sliding distances, the COF remains low but increases as the sliding distance increases. The COF of Mg and Mg-B₄C composites is higher than that of composites reinforced with graphite (Mg-Gr) and those hybridized with both B₄C and graphite (Mg-B₄C-Gr). Figure 3c shows a steady decrease in COF as the sliding velocity increases (from 0.5 m/s to 3 m/s), mainly due to the formation of a Mechanically Mixed Layer (MML) on the surface of the counter disc.

3.2. Model Performance Evaluation

Performance metrics provide a quantitative evaluation of how ML models align with real data. For supervised learning in regression tasks, the R² (coefficient of determination) is a key metric, indicating the proportion of data variation explained by the model. R² values range from 0 to 1, where 0 indicates no correlation and values above 0.9 signify highly satisfactory model execution. Values between 0.70 and 0.90 represent adequate performance. Other error metrics, such as Mean Absolute Error (MAE), Mean Squared Error (MSE), and Root Mean Squared Error (RMSE), are also essential for assessing model accuracy, with lower values indicating better prediction accuracy.

Table 4. Different ML methods’ performance values for predicting wear loss.

Model	MSE	RMSE	MAE	R2
ANN	1.0596 × 10−6	0.0010294	0.00083431	0.75565
KNN	1.2339 × 10−6	0.0011108	0.000835	0.7858
RF	1.031 × 10−6	0.0010154	0.00080914	0.72001
SVM	7.8438 × 10−7	0.00088565	0.00065783	0.80112
GBM	4.4567 × 10−7	0.00066758	0.00048194	0.88914

presents the performance values of the ML models predicting material loss in magnesium hybrid composites. The R² values for validation data range from 0.72001 to 0.88914 across five ML models. The MAE values range from 0.00048194 to 0.000835, the MSE ranges from 4.4567 × 10⁻⁷ to 1.2339 × 10⁻⁶, and RMSE values range from 0.00066758 to 0.0011108, all indicating high prediction accuracy. The Gradient Boosting Machine (GBM) model achieved the highest R² value of 0.88914, suggesting nearly 90% accuracy in predicting wear loss. Additionally, GBM exhibited the lowest MSE, RMSE, and MAE values, indicating its superior performance.

Table 5. Different ML methods’ performance values for predicting COF.

Model	MSE	RMSE	MAE	R2
ANN	0.00068406	0.026155	0.02114	0.80428
KNN	0.0042372	0.065094	0.057691	0.55682
RF	0.0019593	0.044264	0.036311	0.7472
SVM	0.0020778	0.045583	0.039353	0.66813
GBM	0.0011032	0.033214	0.024421	0.83406

shows the performance metrics for predicting the coefficient of friction (COF). R² values for the test sets range from 0.55682 to 0.83406, with MAE values between 0.02114 and 0.057691, MSE ranging from 0.00068406 to 0.0042372, and RMSE varying from 0.026155 to 0.065094, all of which are notably low. These results suggest satisfactory COF prediction accuracy. The GBM model achieved the highest R² value of 0.83406, indicating 83.4% prediction accuracy, with MAE, MSE, and RMSE values of 0.0011032, 0.033214, and 0.024421, respectively, confirming the model’s strong predictive accuracy.

3.3. Impact of Independent Variables and Correlation Heatmap

The feature-importance characteristic of the GBM model evaluates the significance of input variables in predicting wear loss, assigning each a score between 0 and 1, with higher scores indicating greater importance. Figure 4a illustrates the feature importance in wear loss prediction for magnesium composites, showing that all independent variables contribute. Graphite weight percentage received the highest score, identifying it as the most critical factor due to its solid lubricant properties, which reduces surface adhesion, abrasive particles, and material transfer. While boron carbide and magnesium enhance structural strength, graphite’s lubrication minimizes wear, compensating for magnesium’s inherent softness. The interaction between boron carbide (hard particles) and graphite (lubricant) improves wear resistance, though graphite has a more significant effect. Sliding distance ranked as the second most significant factor, as wear loss increases with prolonged surface interaction, leading to friction-induced heat that softens the magnesium matrix. Applied load and sliding speed also influence wear loss, though their impact is moderated by graphite’s friction-reducing properties and the cumulative effects of sliding distance. Boron carbide weight percentage contributes to hardness and wear resistance but plays a secondary role compared to graphite’s direct lubrication effect.

Figure 4b shows the feature importance chart for predicting the coefficient of friction (COF) in magnesium-based hybrid composites. All independent variables contribute to the COF prediction, with none scoring zero. Graphite wt% has the highest importance due to its lubricating properties, which significantly reduce friction between sliding surfaces. Other variables such as boron carbide wt%, magnesium wt%, and operational factors like sliding distance and load play secondary roles. Graphite wt% directly lowers COF because of its solid lubrication behavior, making its high feature importance consistent with tribological principles. The applied load is the second most important variable for COF prediction. As the load increases, softer materials like magnesium may deform more, altering friction. This can lead to plastic deformation of surface asperities, raising friction as more material contacts the surfaces. In hybrid composites, the reinforcement type, like boron carbide, influences this interaction, though graphite helps mitigate its effects. However, when the load surpasses a certain threshold, COF may slightly increase due to material deformation and adhesion forces. Additionally, higher loads generate more frictional heat, which can degrade the graphite layer, further increasing COF. Therefore, despite the presence of graphite, the applied load affects friction by altering the thermal and mechanical properties of the material. Other factors, such as sliding speed, sliding distance, and boron carbide wt%, play secondary but relevant roles in COF prediction. Sliding speed affects frictional heat generation, which can soften the matrix material or wear down the graphite layer, but this effect is less significant than load. Sliding distance increases COF by extending surface interaction time, though graphite’s lubrication reduces its impact over time. Boron carbide wt% increases composite hardness, reducing material deformation, but it does not directly influence friction as graphite does.

The MATLAB heatmap function was utilized to visualize the relationships between parameters in the dataset, as shown in Figure 5. A value approaching “1” indicates a strong positive relationship, where a change in one feature typically causes a corresponding change in the other. A correlation value of “−1” represents an ideal inverse relationship, where an increase in one variable leads to a decrease in the other. A correlation close to “0” suggests little to no relationship between variables.

The heatmap reveals the correlations between wear loss and COF with other factors, such as sliding speed, applied load, sliding distance, and the weight percentages of different powders (Mg, B₄C, and Gr) added to the composite material. The wear loss shows a moderate negative correlation with the graphite powder weight percentage (−0.5776), indicating that higher graphite content reduces wear loss. Conversely, there is a moderate positive correlation between wear loss and magnesium powder weight percentage (0.5616), suggesting that more magnesium increases wear loss. The relationship between wear loss and applied load is weak but positive (0.228), meaning higher loads are generally associated with more wear loss. The correlation with sliding speed is negative (−0.3611), suggesting faster speeds reduce wear loss, while the sliding distance shows a positive correlation (0.4068), indicating wear loss increases over longer distances.

For the coefficient of friction (COF), a moderate positive correlation is found with applied load (0.3358) and Gr powder weight percentage (0.437), meaning COF tends to rise as applied load or graphite content increases. The correlation between COF and sliding speed is moderate and negative (−0.414), indicating that higher speeds typically result in lower COF values. There is also a slight negative correlation between COF and Mg powder (−0.3262), suggesting that magnesium may help reduce COF. The correlations with sliding distance and B₄C powder are weak, indicating minimal influence on COF.

3.4. Validation and Prediction of Different Models

A comparison of various ML models for forecasting wear loss and coefficient of friction (COF) is presented using the experimental testing dataset, as shown in Figure 6 and Figure 7. The results indicate that the Gradient Boosting Machine (GBM) achieves the highest accuracy in predicting both wear loss and COF compared to the other models. Figure 8 and Figure 9 display the predictions from different models alongside the experimental results for wear loss and COF. The experimental data shows significant variation and analysis of the plots confirms that the GBM model offers the most accurate predictions. The ML models developed in this study effectively predicted material loss and COF for composite materials with a satisfactory level of accuracy, utilizing various material properties and tribological factors.

3.5. Auto-Correlation Study for the GBM Model

The Autocorrelation Function (ACF) plot is employed to assess potential correlations in residuals and any unaccounted patterns, serving as a key diagnostic tool for analyzing time-dependent relationships in residuals and providing insights into the model’s performance during analysis [58]. The plot reveals that the residual correlation coefficients remain within the established confidence bounds, indicating minimal autocorrelation. This result suggests that the model has effectively captured and integrated temporal relationships. Conversely, coefficients falling outside these bounds would indicate residual autocorrelation, implying that the model may have missed certain patterns. In the case of the GBM model analysis, the plot shows that all autocorrelation coefficients remain within the expected confidence intervals, as shown in Figure 10. The absence of significant autocorrelation indicates that the model successfully accounts for time-based patterns, demonstrating its strong ability to capture the temporal relationships in the data.

4. Conclusions

This research focuses on predicting the coefficient of friction (COF) and wear loss for magnesium-based hybrid composites using various ML algorithms, based on six input parameters and two output parameters. The study yielded the following key conclusions:

• Magnesium hybrid composites were successfully produced through the powder metallurgy process, reinforced with varying weight percentages of boron carbide (B₄C) and graphite particulates. The tribological behavior (wear loss and COF) of these composites was examined concerning material properties and testing conditions, using both traditional and data-driven analysis.
• Five ML models were developed using 86 training, 29 validation, and 39 test data points (total: 154 data points) to predict wear loss and COF. Performance analysis confirmed that ML models effectively captured nonlinear relationships in tribological behavior. The Gradient Boosting Machine (GBM) outperformed other models in predicting both wear loss (R²: 0.89, MSE: 4.47 × 10⁻⁷, RMSE: 0.00066, MAE: 0.00048) and COF (R²: 0.83, MSE: 0.0011, RMSE: 0.0332, MAE: 0.0244).
• Feature importance analysis indicated that the type of reinforcement particles, particularly graphite particles, had the most significant effect on both wear loss and COF.
• For wear loss, the heatmap correlation showed that increasing graphite powder content and sliding speed led to decreased wear loss, while higher magnesium powder content, applied load, and sliding distance resulted in increased wear loss.
• In the COF analysis, the heatmap correlation revealed that sliding speed and graphite powder content contributed to a decrease in COF, while sliding distance, applied load, and magnesium powder content were associated with an increase in COF. The presence of B₄C powder had minimal impact on the COF.
• The study provides quantitative insights into the wear mechanisms of magnesium hybrid composites, demonstrating that ML models can serve as powerful predictive tools for tribological behavior. These findings highlight the potential of ML-driven approaches in optimizing material compositions for wear-resistant applications in automotive, aerospace, and biomedical industries.
• Integration of multi-objective optimization techniques with machine learning models can further assist in identifying the optimal composition and process parameters for achieving superior wear resistance and friction performance.