The Journal of
the Korean Journal of Metals and Materials

The Journal of
the Korean Journal of Metals and Materials

Monthly
  • pISSN : 1738-8228
  • eISSN : 2288-8241

Editorial Office





Fe-Cr composite coating, HVOF, Wear mechanism, Deep learning, Tribology, Wear pattern classification, Euro 7 compliance

1. INTRODUCTION

The transition of the automotive industry toward electric vehicles (EVs) has introduced new challenges in brake system design and particulate emission control [1]. Although EVs reduce tailpipe emissions, their increased mass intensifies the mechanical demands on brake systems, potentially exacerbating non-exhaust particulate matter (PM) emissions [2,3]. These brake-derived particles, which are typically in the respirable size range, pose significant environmental and health concerns owing to their high reactivity with biological tissues [4,5].

This challenge has been formally recognized in the new Euro 7 regulation (Regulation (EU) 2024/1257), which introduced specific limits for brake particle emissions for the first time [6]. The regulation stipulates that from November 2026, passenger cars and light commercial vehicles must satisfy PM10 emission limits of 7 and 3 mg km−1 for conventional powertrains and battery EVs, respectively, with stricter limits planned for implementation from January 2030 [1].

1.1. Background

In light of these stringent regulatory requirements, critical reassessment of current brake system materials is required. Conventional brake systems predominantly employ gray cast iron (GCI) discs because of their cost-effectiveness and adequate mechanical properties. However, GCI exhibits several significant limitations that are problematic in the context of modern applications and emerging emission standards [7,8]. This material exhibits a notable susceptibility to corrosion in humid environments and experiences accelerated wear when exposed to road salt. These conditions can lead to increased PM emission under corrosive conditions.

Furthermore, friction-induced temperature elevation during braking significantly affects the wear characteristics and braking performance [9]. This thermal effect creates a need for materials that can maintain stable friction coefficients across diverse operating conditions [10], particularly in demanding EV applications.

1.2. Current Solutions and Limitations

To address these challenges, various coating materials, including WC, Ni, and Cr, have been developed to enhance the wear resistance and corrosion protection [11]. These materials demonstrate superior mechanical properties and thermal stability compared to conventional GCI, making them promising candidates for brake-disc applications [12].

For example, high-velocity oxygen fuel (HVOF) coating techniques have garnered attention owing to their effectiveness when applied to brake discs [13,14]. The HVOF process involves heating and accelerating the coating particles through an HVOF jet to create dense, high-bonding coatings characterized by enhanced hardness and strong adhesion strength [15].

Fe-Cr-based composite materials have emerged as particularly promising candidates, offering an attractive combination of performance and cost-effectiveness [16,17]. These materials demonstrate favorable phase transformation characteristics during thermal spray processing, uniform microstructure development, and superior mechanical properties in the Fe matrix, while utilizing more economically viable raw materials [18,19].

1.3. Wear Analysis Challenges

Parallel to material development, the characterization of wear mechanisms presents its own set of challenges. Traditional wear mechanism analysis relies heavily on the expert interpretation of scanning electron microscopy (SEM) images, which poses significant limitations. This process involves considerable subjectivity in pattern recognition and is time-intensive, particularly when processing large datasets [20-22].

Although artificial intelligence approaches have shown promise for automated wear classification, existing studies have shown notable constraints. To overcome these, our study introduces a multimodal deep learning framework that integrates SEM image analysis with friction coefficient-derived features, addressing class imbalance, morphological variability, and subjectivity. This multimodal approach ensures higher robustness compared with previous methods. These include limitations in binary classification capabilities, restricted dataset sizes, and persistent challenges with class-imbalance issues [21,22].

1.4. Research Objectives

Considering the above-mentioned challenges, this study aimed to advance the understanding of Fe-Cr-based composite coatings and automated wear pattern analysis through an integrated approach. Consequently, this study had three interconnected objectives.

1. This study focused on tribological performance evaluation through the application of Fe-Cr-based composite coatings on GCI substrates using HVOF. This includes a systematic assessment using ball-on-disk testing and a comparison of the wear behavior against Si3N4 and WC counterfaces.

2. We developed advanced pattern recognition capabilities by implementing a deep ensemble convolutional neural network (CNN) model for wear classification. This involves creating a comprehensive, balanced dataset and implementing model averaging to enhance generalization.

3. The study addressed practical implementation by validating the method through classification reliability across diverse wear patterns, assessment of model performance under various tribological conditions, and establishment of an automated wear analysis framework.

The selection of Si3N4 and WC as counterface materials enabled a comprehensive evaluation of their contrasting properties. Si3N4 provides high-temperature stability and chemical inertness, whereas WC provides superior hardness and wear resistance. This strategic selection of materials directly supports the first objective of tribological performance evaluation, whereas the implementation of deep ensemble CNN modeling addresses the second objective of automated pattern recognition. By integrating these approaches, this study establishes a robust methodology for both coating performance optimization and wear mechanism classification in tribological applications, particularly for automotive brake systems that are required to meet Euro 7 standards.

2. Experimental

2.1. High-velocity Oxygen Fuel Spray Coating and Specimen Preparation

The substrate material selected for this study, GCI (FC200D, Myunghwa Ind. Co., Ltd., Republic of Korea), was machined into disk specimens with dimensions of 30 mm × 10 mm. Two novel Fe-Cr-based composite materials (Atometal Tech Korea Inc., Republic of Korea), named Composites A and B, were coated onto the substrates to enhance their glass-forming ability, which refers to the tendency of the powders to form amorphous-like or fine crystalline structures during the rapid solidification process in HVOF coating, potentially improving the coating's mechanical properties and wear resistance. These coating layers were deposited onto the substrates using an HVOF spraying process, performed by Shinhwa Metal Co., Ltd., Republic of Korea. The coating process was conducted according to the standardized industrial procedures of the company using powder particle size of 15–45 μm and target coating thickness >100 μm to ensure coating quality and reliability.

The surface preparation of the coated specimens was prepared by mechanical grinding and polishing processes. The final coating thickness was controlled to be greater than 100 μm, as measured using a coating thickness gauge (OM-8811FN, Optech, China), with an accuracy of ±2 μm. Measurement was obtained at five randomly selected points per specimen to ensure data precision. Figure 1(a) shows the SEM micrographs of the coating cross sections, illustrating the uniformity of the coating thickness and the quality of the interface between the coating and substrate.

Surface roughness was characterized using a surface profilometer (SURFTEST SJ-410, Mitutoyo Corporation, Japan) with a diamond stylus tip of 2 μm radius. Three measurements were obtained for each specimen to achieve a final Ra value in the range of 0.15–0.2 μm, which ensures optimal wear resistance and frictional performance while maintaining experimental consistency and cost efficiency.

Hardness measurements were performed (50× magnification) using a Vickers hardness tester (HV-114, Mitutoyo Corporation, Japan). The measurements were performed under an applied load of 1 kgf (9.81 N), with a dwell time of 10 s. Five measurements were obtained per specimen at 500 μm intervals to ensure representative sampling of the material.

2.2. Tribological Testing

Friction and wear behavior evaluations were performed using a ball-on-disk tribometer (MVP 110, Neoplus, Republic of Korea) in accordance with the ISO 20808 standard. This study employed two counterface materials: Si3N4 and WC. Both ball specimens were manufactured with a diameter of 12.7 mm and surface roughness Ra < 0.3 μm.

Test conditions were carefully controlled to ensure reliability. The tests were performed under a normal force of 20 N at a rotation speed of 100 RPM, corresponding to a line speed of 0.120 m s-1. The track radius was set to 11.5 mm for a total line distance of 1000 m (Fig 2(b)). Environmental conditions were maintained at 33 ± 2 °C with relative humidity of 50% ± 10%, in accordance with the ISO 20808 standard.

Wear-volume measurements were conducted using a high-precision surface profilometer (Alpha-Step ET3000, Kosaka Laboratory Ltd., Japan) (Fig 2(c)). The measurement parameters included a scanning speed of 0.1 mm s-1, stylus force of 300 μN, and horizontal resolution of 0.01 μm. The 0.01 μm digital scale embedded in the X-coordinate reading enabled high-accuracy horizontal measurement data sampling.

The wear-track profile was evaluated at four equidistant positions at 90° intervals around the wear track. The total wear volume (Vdisk) was calculated based on the methodology described in ISO 20808:2016 [23] as follows:

(1)
V disk = π R ( S 1 + S 2 + S 3 + S 4 ) / 4

where R represents the wear track radius (m) and S1–S4 denote the cross-sectional areas of the wear track (m2) at the four measured positions.

The specific wear rate (Ws) was subsequently determined using

(2)
W s = V disk P · L

where P represents the applied normal load (N), and L denotes the total sliding distance (m), as described in ISO 20808:2016 [23].

2.3. Classification of Wear Surfaces

2.3.1. Data Collection and Preprocessing

Microstructural analysis of the worn coating surfaces after tribological testing was performed using SEM (JSM-6390, JEOL, Japan). Energy-dispersive X-ray spectroscopy (EDS; Oxford Instruments, United Kingdom) was used for the elemental analysis of the wear surfaces. To enable systematic wear pattern classification, we collected a comprehensive dataset of 500 high-resolution SEM micrographs documenting various wear mechanisms observed under various test conditions (Table 1).

This dataset forms the foundation for the deep learning-based classification approach described in subsequent sections. The dataset comprised 500 SEM micrographs distributed across five distinct categories (0–4), representing different tribological testing conditions and surface morphologies. Class 0 represents untested surfaces, whereas Classes 1–4 correspond to specific material pair combinations under standardized testing conditions.

The dataset was systematically categorized into five distinct wear-surface types based on their morphological characteristics and underlying wear mechanisms, as listed in Table 1. Image quality consistency was maintained using standardized imaging parameters on the secondary electron mode and 300× magnification.

2.3.2. Data Management and Augmentation

Dataset partitioning was implemented using the train test split function of the Scikit-learn library with stratification to maintain balanced class distributions across all wear categories. Initially, 20% of the dataset (100 images) was allocated to the test set, and the remaining 400 images were split at an 80:20 ratio, resulting in 320 training and 80 validation images. The training set was used to teach the model by updating the weights through backpropagation, the validation set was used to tune hyperparameters and monitor overfitting, and the test set was used to evaluate the final model performance on unseen data. A consistent random seed number of 42 was applied throughout the splitting process to ensure reliability.

After partitioning, data augmentation was performed using the PyTorch transform function to improve the generalization ability of the model by artificially increasing the diversity of the training data. The augmentation operations included the following steps:

· Conversion to PyTorch tensor format: Transforming images into tensors for processing within PyTorch.

· Spatial resizing to uniform dimensions of 100 × 100 pixels: This ensures that all images are of consistent size to be input into the model.

· Normalization using ImageNet statistics: Standardizing the input data to match a common scale.

· Conversion to grayscale: Reducing image complexity by limiting color channels and focusing on surface wear features.

· Random sharpness enhancement by a factor of 2: Increasing image sharpness to enhance edge features, making patterns more distinguishable.

· Automatic contrast adjustment: Balancing contrast to improve the visual differentiation of features.

Data processing was conducted in batches of 10 images using DataLoader function of PyTorch library with two worker processes for efficient data handling. Subset random samplers were employed to maintain a stratified distribution of wear categories across all dataset splits during the training and evaluation phases. This ensured that the distribution remained consistent, and each mini-batch was representative of the overall dataset.

2.3.3. Model Architecture and Training

A VGG-style CNN architecture was implemented for wear surface classification, comprising six convolutional layers with batch normalization divided in three distinct stages:

· Initial stage:

- Two convolutional layers with 32 filters (3 × 3 kernels each), followed by batch normalization and rectified linear unit (ReLU) activation.

- A 2 × 2 max pooling (stride 2) layer to reduce spatial dimensions.

· Intermediate stage:

- Two convolutional layers with 64 filters (3 × 3 kernels each), followed by batch normalization and ReLU activation.

- A 2 × 2 max pooling (stride 2) layer to further reduce dimensionality.

· Final stage:

- Two convolutional layers with 128 filters (3 × 3 kernels each), followed by global average pooling to condense the feature maps.

- A fully connected layer (128 to 5 features) for final classification into five wear surface categories.

The model was trained using cross-entropy loss to handle multiclass classification and the Adam optimizer with default parameters (learning rate = 1 × 10-3). Training was conducted over 50 epochs with a batch size of 10, which balanced computational efficiency and model convergence.

To improve the classification robustness, an ensemble of five independently trained models was used, each initialized with a different random seed (seed = 2020). The final predictions were determined by majority voting across the ensemble members. The model checkpoints were preserved at the epoch with the highest validation accuracy, helping to prevent overfitting and improve generalization.

All models were implemented using PyTorch and trained on an NVIDIA Tesla T4 GPU (Google Colab) using CUDA 11.8. The complete dataset and source code are publicly available at https://doi.org/10.17632/m7vmhjbktv.1 for further reference and reproducibility. This computational approach complements the physical tribological testing previously described, creating an integrated workflow where material characterization leads directly to quantitative analysis of wear patterns, bridging the gap between traditional materials science and advanced computational methods.

3. RESULTS AND DISCUSSION

3.1. Characterization of Mechanical Properties

3.1.1. Hardness Analysis

Vickers hardness measurements revealed significant differences between the test materials (Table 2). The counterface materials demonstrated superior hardness characteristics, with Si3N4 and WC measuring (15.33 ± 0.47) and (16.40 ± 0.35) GPa, respectively. These values exceed those of the friction materials by approximately 2.3 times, ensuring geometric stability throughout the wear testing process [24].

The Fe-Cr-based composite coatings exhibited substantially enhanced hardness compared to that of the GCI substrate. The GCI substrate, Composite A, and Composite B achieved hardness values of (1.98 ± 0.15), (7.51 ± 0.20), and (6.52 ± 0.13) GPa, respectively. This significant improvement in the composites, which was approximately 3.2 times greater than that of the substrate, suggests an enhanced potential for wear resistance [25]. As demonstrated in subsequent tribological testing, this marked increase in hardness directly influences friction behavior and wear mechanisms, establishing a fundamental relationship between mechanical properties and tribological performance that forms the basis for our subsequent analysis.

3.1.2. Surface Topography

Surface roughness measurements confirmed a consistent surface finish across all the specimens, maintaining Ra values in the range of 0.15–0.2 μm (Table 2). This standardization enabled a direct comparison of the wear behavior based on intrinsic material properties rather than surface-finish variations [26,27].

3.2. Tribological Performance Analysis

3.2.1. Friction Behavior

The friction coefficient (μ) measurements revealed distinct material-dependent patterns across the test specimens (Fig 3). The GCI specimens exhibited a friction coefficient of 0.91 against Si3N4, which notably decreased to 0.36 against WC, representing a 60.4% reduction in friction against the WC counterface (Table 3). This significant difference highlights the strong dependence of friction behavior on counterface material selection.

Composite A demonstrated a different behavior, with a friction coefficient of 0.82 against Si3N4 and an unexpected increase to 0.90 against WC, representing a 9.8% increase in friction with the WC counterface. Composite B showed the most distinctive behavior, maintaining a friction coefficient of 1.03 against Si3N4. In contrast, it exhibited the highest observed friction coefficient of 2.00 against the WC counterface. Detailed analysis of the friction data revealed that Composite B with WC counterface demonstrated the highest variability (standard deviation: 0.19) and peak-to-peak amplitude (0.69), indicating greater tribological instability in this material combination.

3.2.2. Wear Rate Analysis

The specific wear rates exhibited clear material-dependent behavior patterns (Table 3). The GCI showed wear rates of 25.78 × 10-6m3N-1 m-1 against Si3N4 and 2.05 × 10-6m3N-1m -1 against WC. Composite A exhibited improved wear resistance with rates of 5.07 × 10-6 and 13.43 × 10-6 m3 N-1 m-1 against Si3N4 and WC, respectively. Composite B exhibited contrasting behavior, achieving a low wear rate of 7.37 × 10-6 mm3 N-1 m-1 against Si3N4, but an exceptionally high wear rate of 387.26 × 10-6 mm3 N-1 m-1 against WC, representing a dramatic 52-fold increase. This extreme tribological behavior can be attributed to the significant hardness mismatch between Composite B (~6.5 GPa) and WC ball (~16.4 GPa), creating a hardness ratio of approximately 2.5:1. The combination of high contact pressure and material incompatibility likely promoted severe adhesive wear mechanisms, characterized by material junction growth and subsequent fracture. Additionally, the friction coefficient data revealed exceptional instability for this pairing (μ = 2.00 ± 0.19), with the highest standard deviation and peak-to-peak amplitude (0.69) observed across all test combinations, indicating severe stick-slip behavior that would accelerate wear processes.

The relationship between the friction coefficient and wear rate reveals notable material-dependent characteristics. The GCI exhibited conventional behavior with a proportional relationship between these parameters. Composite A exhibited a moderate deviation from this proportionality, maintaining lower wear rates than its friction coefficients. Composite B demonstrated a unique counterface-dependent behavior, maintaining low wear rates against Si3N4 but exhibiting exceptionally high wear rates against WC.

These findings indicate that counterface material selection significantly influences tribological behavior, whereas composite coatings exhibit complex friction-wear relationships that deviate from traditional proportional patterns. More importantly, the results demonstrate that higher friction coefficients are not always correlated with increased wear rates, except in cases similar to those of Composite B with WC counterfaces, where both parameters increase. This non-linear relationship between friction and wear forms a critical foundation for understanding the tribochemical interactions and morphological changes observed in the subsequent wear surface characterization, in which specific wear mechanisms are identified and correlated with the observed tribological behaviors.

3.3. Wear Surface Characterization

3.3.1. Morphological Analysis

SEM examination revealed distinct wear mechanisms across different material combinations (Fig 4). Testing with Si3N4 counterfaces predominantly resulted in abrasive wear patterns, whereas testing with WC counterface conditions showed the dominance of adhesive wear mechanisms. These observations were further supported by quantitative analysis of the friction behavior, where Si3N4 counterfaces demonstrated lower variability in friction coefficients (standard deviation: 0.07–0.09) than WC counterfaces (standard deviation: 0.11–0.19), particularly for Composite B. Transfer films were consistently observed across all the composite coating combinations, indicating complex material interactions during the wear process.

3.3.2. Chemical Analysis

EDS mapping was performed to identify the specific tribochemical interactions in each counterface material. The results revealed distinct oxidation patterns that varied depending on the counterface material used during the wear tests. Thus, the results provide insights into the tribochemical mechanisms at play.

In the tests involving Si3N4 as a counterface material, the EDS analysis revealed the formation of SiO2 through oxidation, as described by the following equation:

(3)
Si 3 N 4 + 3   O 2     3   SiO 2   +   2   N 2 .

This oxidation reaction occurred uniformly across the wear track, which is evident from the consistent distribution of O-rich regions in correlation with the Si signals. Figure 5 illustrates the EDS elemental maps of the wear tracks, showing the elemental distribution of Fe, Cr, Si, and O. Uniform coverage of SiO2 was observed across the wear track, indicating an effective tribochemical interaction between the counterface and the composite material.

When WC was used as the counterface, EDS analysis showed a distinct tribochemical inter-action characterized by the formation of WO2. This was confirmed by the following equation, which describes the oxidation process leading to WO2 formation alongside CO2:

(4)
2   WC   +   5   O 2     2   WO   2   +   2   CO 2 .

The mapping in Fig 6 shows the localized W transfer and oxidation patterns, which are distinct in the wear tracks. In contrast to the Si3N4 counterface, the WC counterface led to the transfer of W to the surface, resulting in areas with high concentrations of W and O. These distinct chemical signatures and morphological features create the visual patterns that form the basis of the proposed deep learning-based classification approach, which are presented in the following section. By training a CNN model on these unique wear patterns, we can establish a direct link between tribochemical mechanisms and automated pattern recognition capabilities.

3.4. Deep Learning-based Classification of Wear Surface Patterns

Although conventional SEM and EDS analyses provide valuable insights into wear mechanisms, the automated classification of complex wear patterns requires a more systematic approach. Consequently, a deep learning-based classification framework was developed and evaluated using a comprehensive dataset of 500 SEM micrographs (Fig 7).

3.4.1. Dataset Organization and Model Training

The dataset was systematically partitioned while maintaining balanced class distributions across all wear categories. The final distribution is as follows:

· Training set: 320 images (64%, 64 images per class)

· Validation set: 80 images (16%, 16 images per class)

· Test set: 100 images (20%, 20 images per class)

The CNN model demonstrated consistent learning behavior throughout the 50-epoch training period (Fig 8), indicating effective knowledge acquisition from the wear pattern images. The training process exhibited three distinct phases of development, each representing a critical stage in the learning progression of the model.

1. Initial rapid learning phase (epochs 1–10): Training loss decreased substantially from approximately 1.3 to 0.6

2. Progressive refinement phase (epochs 11–30): Loss further reduced to approximately 0.3 with increasing stability

3. Fine-tuning phase (epochs 31–50): Optimization of model performance, with loss stabilizing below 0.2

3.4.2. Ensemble Model Performance

This study implemented an ensemble of five independently trained models to enhance classification reliability, particularly for challenging wear patterns. This ensemble approach demonstrated exceptional performance across all evaluation metrics. On the training set, the ensemble achieved accuracy of 0.947, precision of 0.952, recall of 0.947, and F1-score of 0.946. The evaluation metrics were defined as follows: Accuracy represents the proportion of correctly classified samples. Precision measures the accuracy of positive predictions (TP/(TP+FP)). Recall represents the model's ability to identify all positive cases (TP/(TP+FN)). F1-score provides the harmonic mean of precision and recall. All metrics were calculated using macro-averaging to ensure equal treatment across all classes. The validation set exhibited perfect performance across all metrics with values of 1.000, while the test set maintained strong performance with consistent values of 0.960 across all metrics (Table 4).

The ensemble model demonstrated significantly superior performance compared to individual models in the test set evaluation. While individual model performances varied, Model 1 achieved an accuracy of 0.830 and an F1-score of 0.811, Model 2 achieved 0.930 for both metrics, Model 3 attained an accuracy of 0.940 and an F1-score of 0.939, Model 4 achieved 0.920 for both accuracy and F1-score, and Model 5 achieved the highest individual accuracy and F1-score of 0.990. These results highlight the consistency and reliability obtained of the ensemble approach.

3.4.3. Classification Analysis and Error Patterns

Confusion matrix analysis revealed highly accurate classification patterns across most wear categories (Fig 9). Classes 0, 1, and 2 were perfectly classified, with the accurate identification of all 20 test samples. The classification of Class 3 showed strong performance, with the accurate classification of 18 out of 20 samples, exhibiting minimal confusion with Class 4. Similarly, in the identification of Class 4, robust performance was achieved with the accurate classification of 18 out of 20 samples, showing a marginal confusion with Class 3.

The observed misclassifications primarily occurred in cases where the wear patterns shared similar morphological features, particularly between the wear surfaces of the samples under Classes 3 and 4 (Fig 9). This pattern aligns with our tribological findings, where both classes represent WC counterface combinations that produced comparable wear mechanisms, albeit with different intensities due to material-specific properties. Analysis of the feature activations from the CNN model revealed that Class 3 (WC & Composite A) and Class 4 (WC & Composite B) exhibited relatively similar average activation values (0.478 and 0.477, respectively) with comparable standard deviations (0.435 and 0.491), explaining the classification challenge. This pattern suggests that, although the model distinguishes different wear mechanisms, more subtle variations in the surface morphology can present classification challenges (Fig 10).

In our evaluation, the ensemble approach demonstrated strong performance particularly in mitigating classification errors, exhibiting solid performance across all wear pattern categories analyzed. The high level of accuracy (96.0% on the test set) verifies the potential of deep-learning-based approaches for automated wear surface analysis in tribological applications.

4. CONCLUSIONS

This comprehensive study of Fe-Cr-based composite coatings provided significant insights into both material performance and automated analysis capabilities. The study of the mechanical properties revealed that Fe-Cr-based composite coatings achieved superior hardness values ranging from 6.52–7.51 GPa, substantially surpassing that of the GCI substrate of 1.98 GPa. Surface consistency was maintained across all specimens, with roughness values (Ra) in the range of 0.15–0.20 μm, enabling reliable comparative analysis.

The tribological performance analysis demonstrated particularly noteworthy results for Composite B, which exhibited unique characteristics, including elevated friction coefficients against Si3N4 (μ = 1.03), while maintaining significantly low specific wear rates (7.37 × 10-6 mm3 N-1 m-1). This represents a 3.5-fold improvement in the wear resistance compared to that of the conventional GCI. Further analysis of friction behavior indicated that although Composite B exhibited greater stick-slip amplitude and transient events, it maintained superior wear resistance against Si3N4 counterfaces, suggesting a complex relationship between friction instability and wear resistance in these material combinations. This study identified distinct counterface-dependent wear mechanisms, with Si3N4 promoting combined abrasive and oxidative wear. In contrast, WC interfaces primarily exhibited adhesive wear with oxide formation.

A detailed tribochemical analysis revealed specific oxidation patterns that are characteristic of each counterface material. The test considering Si3N4 resulted in SiO2 formation through tribochemical reactions, whereas the WC interfaces produced WO2 via similar mechanisms. These distinct chemical interactions were thoroughly validated by EDS analysis, which confirmed the material-transfer patterns and oxidation mechanisms and provided critical insights into the fundamental wear processes underlying material performance in automotive brake applications.

The implementation of deep learning classification achieved significant results, with the ensemble model demonstrating superior performance across all evaluation metrics. The training, validation, and test accuracies were 0.947, 1.000, and 0.960, respectively. The classification system showed particular strength in distinguishing between most wear patterns, with only minimal confusion between similar wear mechanisms, accounting for 4 out of 100 misclassifications.

These findings show the significant potential of Fe-Cr-based composite coatings for tribological applications and verify the effectiveness of machine learning approaches in wear mechanism analysis. The integrated methodology developed in this study provides a cohesive research framework in which mechanical properties directly influence tribological performance, which in turn produces characteristic wear patterns that can be automatically classified using deep learning techniques. This interconnected approach successfully bridges the traditional tribological characterization with advanced pattern recognition, establishing a comprehensive framework that follows the material life cycle from mechanical properties to wear mechanism identification. Such an integrated approach provides a robust foundation for future wear-related studies in automotive brake applications, particularly those aimed at satisfying Euro 7 emission requirements.

Notes

[1] ACKNOWLEDGEMENT

The authors gratefully acknowledge the support of the KOLON Future Technology Institute Application Development Laboratory for providing the composite coating samples.

REFERENCES

1 
Dornoff J., Rodríguez F., Euro 7: New emission standards for light and heavy-duty vehicles in the European Union, https://theicct.org/publication/euro-7-emission-standard-ldv-hdv-eu-mar24/, (2024)
2 
Woo S. H., Jang H. J., Lee S. B., Lee S. H., Sci. Total Environ,842, 156961 (2022)Google Search
3 
Harrison R. M., Allan J., Carruthers D., Heal M. R., Lewis A. C., Marner B., Murrells T., Williams A., Atmos. Environ,262, 118592 (2021)Google Search
4 
Kukutschová J., Moravec P., Tomášek V., Matějka V., Smolík J., Schwarz J., Šafářová K., Filip P., Environ. Pollut,159, 998 (2011)Google Search
5 
Mulani S. M., Kumar A., Shaikh H. N. E. A., Saurabh A., Saurabh A., Singh P. K., Verma P. C., Mater. Today: Proc,56, 447 (2022)Google Search
6 
European Parliament and Council, Regulation 2024/1257, http://data.europa.eu/eli/reg/2024/1257/oj, (2024)
7 
Dizdar S., Lyu Y., Lampa C., Olofsson U., Atmosphere,11, 621 (2020)Google Search
8 
Verma P. C., Menapace L., Lacaita A. L., Ciudin R., Gialanella S., Straffelini G., Wear,322-323, 251 (2015)Google Search
9 
Chen Y., Gao F., Fu R., Su L., Han X., Yang J., Ind. Lubr. Tribol,74, 975 (2022)Google Search
10 
Zhang Q., Yu Z., Liu H., Mo J., Xiang Z., Zhu S., Jin W., Tribol. Int,196, 109701 (2024)Google Search
11 
Da Q., Kang J. J., Ma G. Z., Zhou Y. K., Fu Z. Q., Zhu L. N., She D. S., Wang H. D., Tribol. Int,200, 110149 (2024)Google Search
12 
Hong S. Y., Nahm S., Oh Y. S., J. Korean Ceram. Soc,59, 465 (2022)Google Search
13 
Federici M., Menapace C., Moscatelli A., Gialanella S., Straffelini G., Tribol. Int,115, 89 (2017)Google Search
14 
Park T. J., Lee G. W., Oh Y. S., J. Korean Powder Metall. Inst,30, 415 (2023)Google Search
15 
Zhang J., Fu B., Jiang Y., Dong T., Li G., Tribol. Int,201, 110254 (2024)Google Search
16 
Wang Y., Xu L., Zhao L., Han Y., Hao K., Wenjing R., Tribol. Int,192, 109245 (2024)Google Search
17 
Cho Y. H., Ham G. S., Park S. Y., Kim C. P., Lee K. A., Arch. Metall. Mater,67, 1521 (2022)Google Search
18 
Jin H. W., Park C. G., Kim M. C., Scr. Mater,41, 589 (1999)Google Search
19 
Cho Y. H., Park S. Y., Ham G. S., Kim C. P., Park S. C., Lee K. A., Powder Metall,66, 702 (2023)Google Search
20 
Herwig N., Peng Z., Borghesani P., Tribol. Int,187, 108670 (2023)Google Search
21 
Sieberg P. M., Kurtulan D., Hanke S., Materials,15, 2358 (2022)Google Search
22 
Wang M., Yang L., Zhao Z., Guo Y., Tribol. Int,169, 107466 (2022)Google Search
23 
ISO 20808:2016, Fine ceramics (advanced ceramics, advanced technical ceramics) — Determination of friction and wear characteristics of monolithic ceramics by ball-on-disc method,International Organization for Standardization (2016)Google Search
24 
Pasha M. B., Rao R. N., Ismail S., Gupta M., Prasad P. S., Tribol. Int,196, 109696 (2024)Google Search
25 
Ni W. Y., Cheng Y. T., Lukitsch M. J., Weiner A. M., Lev L. C., Grummon D. S., Appl. Phys. Lett,85, 4028 (2004)Google Search
26 
Federici M., Menapace C., Moscatelli A., Gialanella S., Straffelini G., Wear,368, 326 (2016)Google Search
27 
Kubiak K. J., Liskiewicz T. W., Mathia T. G., Tribol. Int,44, 1427 (2011)Google Search

Figures and Tables

Fig. 1.

Microstructural characterization of Fe-Cr-based composite coatings. Cross-sectional micrographs showing coating thickness and interface quality (left: Composite Coating A, right: Composite Coating B).

../../Resources/kim/KJMM.2025.63/kjmm-2025-63-12-997f1.jpg
Fig. 2.

Experimental methodology and setup configuration. (a) Schematic representation of the high-velocity oxygen fuel coating process showing key components and processing parameters. (b) Ball-on-disk tribometer configuration indicating load application and measurement points. (c) Systematic wear track measurement positions showing the four equidistant sampling locations (S1, S2, S3, S4) at 90° intervals for wear-volume calculations.

../../Resources/kim/KJMM.2025.63/kjmm-2025-63-12-997f2.jpg
Fig. 3.

Evolution of friction coefficients during tribological testing. The curves represent the real-time measurements of the friction coefficients versus sliding distance for different material combinations (left: Si3N4 counterfaces, right: WC counterfaces). Testing conditions were a normal load of 20 N, sliding speed of 100 RPM, ambient temperature of 33 ± 2 °C, and relative humidity of 50 ± 10%.

../../Resources/kim/KJMM.2025.63/kjmm-2025-63-12-997f3.jpg
Fig. 4.

Wear surface analysis of tested specimens. SEM micrographs showing characteristic wear patterns of: (a) Composite material A and Si3N4 ball, (b) Composite material B and Si3N4 ball, (c) Composite material A and WC ball, and (d) Composite material B and WC ball. Insets show low-magnification overview of wear tracks (scale bar: 500 µm).

../../Resources/kim/KJMM.2025.63/kjmm-2025-63-12-997f4.jpg
Fig. 5.

EDS elemental mapping of wear tracks formed against Si3N4 ball: (a) Composite material A and (b) Composite material B. These mappings show the distribution of the key elements, indicating the formation of a uniform SiO2 layer across the wear track. This result provides evidence of effective tribochemical reactions with the Si3N4 counterface.

../../Resources/kim/KJMM.2025.63/kjmm-2025-63-12-997f5.jpg
Fig. 6.

EDS elemental mapping of wear tracks formed against WC ball: (a) Composite material A and (b) Composite material B. The mapping highlights localized W oxidation and material transfer, as evidenced by W and O distribution patterns.

../../Resources/kim/KJMM.2025.63/kjmm-2025-63-12-997f6.jpg
Fig. 7.

Representative SEM micrographs of characteristic wear patterns for each class (0–4). (a) Untested surface (Class 0), (b) Composite A/Si3N4 interface (Class 1), (c) Composite A/WC interface (Class 2), (d) Composite B/Si3N4 interface (Class 3), and (e) Composite B/WC interface (Class 4).

../../Resources/kim/KJMM.2025.63/kjmm-2025-63-12-997f7.jpg
Fig. 8.

Training and validation curves for the convolutional neural network model over 50 epochs. (left plot) Training loss evolution demonstrating model convergence, (right plot) Validation accuracy progression showing generalization performance.

../../Resources/kim/KJMM.2025.63/kjmm-2025-63-12-997f8.jpg
Fig. 9.

Confusion matrix for the test set showing the classification results across five wear surface categories (n = 100 test images). The matrix highlights the number of correctly and incorrectly classified samples in each category. Color intensity corresponds to the classification frequency, with darker shades indicating higher values. Numerical values represent the absolute counts of classifications.

../../Resources/kim/KJMM.2025.63/kjmm-2025-63-12-997f9.jpg
Fig. 10.

Representative examples of misclassified wear surfaces. (a, b) Cases of Class 3 surfaces misclassified as Class 4, showing similar morphological features. (c, d) Cases of Class 4 surfaces misclassified as Class 3, demonstrating subtle pattern similarities. Scale bar: 50 µm. Labels indicate true (T) and predicted (P) classes.

../../Resources/kim/KJMM.2025.63/kjmm-2025-63-12-997f10.jpg
Table 1.

Dataset composition for wear surface classification analysis.

Class Number of Images Testing Conditions
0 100 No (Composite A)
1 100 Si3N4 & Composite A
2 100 Si3N4 & Composite B
3 100 WC & Composite A
4 100 WC & Composite B
Table 2.

Mechanical properties of test materials: hardness and surface roughness measurements. Values represent the mean of five measurements for hardness (applied load: 1 kgf) and three measurements for surface roughness (Ra), with corresponding standard deviations. Measurements were conducted at room temperature (33 ± 2 ℃) and under controlled humidity (50 ± 10% of relative humidity).

Material Hardness (GPa) Surface Roughness Ra (μm)
Silicon nitride ball 15.33 ± 0.47 < 0.3
Tungsten carbide ball 16.40 ± 0.35 < 0.3
Gray cast iron substrate 1.98 ± 0.15 0.20 ± 0.01
Composite coating A 7.51 ± 0.20 0.19 ± 0.01
Composite coating B 6.52 ± 0.1 0.16 ± 0.01
Table 3.

Summary of tribological test results showing friction coefficients and specific wear rates for all material combinations. Data obtained from ball-on-disk tests conducted under standardized conditions (load: 20 N, sliding distance: 1000 m, speed: 100 RPM). Values represent the mean of three independent tests with corresponding standard deviations.

Material Friction Coefficient (μ)
Specific Wear Rate (1 × 10−6 mm3 N−1 m−1)
Si3N4 WC Si3N4 WC
GCI 0.91 0.36 25.78 2.05
Composite A 0.82 0.90 5.07 13.43
Composite B 1.03 2.00 7.37 387.26
Table 4.

Performance metrics of individual and ensemble models on the test set. The ensemble model demonstrated superior overall performance across all evaluation metrics.

Metric Average of Individual Models Ensemble Model
Accuracy 0.922 0.960
Precision 0.925 0.960
Recall 0.922 0.960
F1 Score 0.918 0.960