Modeling, validation and multi-objective optimization of Al-Si/SiC-ZnO-Graphite composites: integrating neural networks-genetic algorithm with interaction effect modeling for tribological performance

Open Access
02.03.2026
ORIGINAL ARTICLE

Verfasst von: Jimmy Lolu Olajide; Isaac Thlabadira; Tarzaan Ezekiel Ayough
Erschienen in: The International Journal of Advanced Manufacturing Technology

Aktivieren Sie unsere intelligente Suche, um passende Fachinhalte oder Patente zu finden.

search-config

KI-gestützte Suche

Patentsuche

Aus

Abstract

Diese Studie befasst sich mit der Modellierung, Validierung und multiobjektiven Optimierung von Al-Si / SiC-ZnO-Graphitverbundwerkstoffen und konzentriert sich dabei auf deren tribologische Leistungsfähigkeit. Die Forschung integriert neuronale Netzwerke und genetische Algorithmen, um Verschleißfestigkeit vorherzusagen und zu optimieren, was sie zu einem bahnbrechenden Ansatz in der Materialwissenschaft macht. Zu den wichtigsten Erkenntnissen gehört die deutliche Verbesserung der Verschleißfestigkeit und der mechanischen Eigenschaften der Verbundwerkstoffe, insbesondere wenn sie für die Zusammensetzung und den Gehalt der Verstärkung optimiert sind. Die Studie hebt auch den erfolgreichen Einsatz von recyceltem hypereutektischem Al-Si-Kolbenschrott hervor, der im Einklang mit den Grundsätzen der Kreislaufwirtschaft steht. Darüber hinaus bietet die Integration digitaler Bildverarbeitung zur mikrostrukturellen Charakterisierung ein umfassendes Verständnis der Leistungsfähigkeit der Verbundwerkstoffe. Die Ergebnisse zeigen, dass die optimierten Verbundwerkstoffe eine überlegene Verschleißfestigkeit und strukturelle Integrität aufweisen und sich daher ideal für automobile Anwendungen wie Zylinderlaufbuchsen eignen. Diese Forschung bringt nicht nur den Bereich der tribologischen Verbundwerkstoffkonstruktion voran, sondern bietet auch praktische Lösungen für nachhaltige und leistungsstarke Werkstoffe in der Automobilindustrie.

KI-Generiert

Diese Zusammenfassung des Fachinhalts wurde mit Hilfe von KI generiert.

Abstract

While main effects of processing parameters on aluminum matrix composite tribology have been extensively studied, interactive effects of reinforcement parameters—which can dominate performance—remain inadequately understood. This research systematically investigates reinforcement content, composition, and critically, their synergistic interaction, on wear behavior of hypereutectic Al-Si composites reinforced with zinc oxide, silicon carbide, and graphite fabricated from recycled piston scrap. Three predictive models were developed: Two-Factor Interaction from Design of Experiments, and Artificial Neural Networks employing Levenberg-Marquardt and Bayesian Regularization algorithms. Digital Image Processing quantified porosity and surface roughness from microstructures and worn surfaces. Hierarchical effects analysis revealed that reinforcement interaction exhibited dominant influence on wear performance (effect magnitude 156.15), exceeding individual effects of composition (95.0) and content (94.8), providing quantitative evidence that synergistic mechanisms rather than additive contributions govern tribological behavior. All models demonstrated exceptional correlation (R² ≥ 0.999), validated through ANOVA, Kruskal-Wallis, Levene’s, Mood’s Median, and Bonferroni tests. Desirability Function Analysis and Genetic Algorithm optimization converged to identical conditions (RCMP7, 11 wt%), predicting minimum wear of 196.59 and 194.98 mg/min respectively, experimentally verified at 197 mg/min (0.2% deviation). The optimized composite achieved 97.48% wear reduction versus monolithic alloy and 61.3% weight reduction versus cast iron for cylinder liner applications, while maintaining fracture toughness (28.6 MPa·m^1/2) exceeding structural requirements. This work establishes interaction-focused optimization as essential for high-performance tribological composite development, with practical implications for automotive lightweighting and fuel efficiency improvement.

Graphical Abstract

Bild vergrößern

Hinweise

Publisher’s note

Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.

1 Introduction

Aluminum-Silicon Alloys (ASAs) constitute 85–90% of aluminum castings used in automotive and aerospace engine components, valued for their low density, excellent castability, high wear resistance, low thermal expansion coefficient, and favorable mechanical properties across wide operating temperatures [1]. Based on silicon content relative to the eutectic point (11.7% Si), these alloys are classified as hypoeutectic (< 11.7%), eutectic (≈ 11.7%), and hypereutectic (> 11.7%), with hypereutectic alloys exhibiting superior tribological characteristics for wear-critical applications [2].

Despite these advantages, premature wear and failure of critical engine components—including pistons, cylinder liners, and cylinder heads—remain significant challenges, manifesting as damaged piston rings, combustion inefficiency, and loss of engine power [3]. This necessitates efficient and economically viable solutions through systematic modifications of material compositions and processing techniques.

Research on the tribological behavior of Al-Si Matrix Composites (ASAMCs) has primarily focused on wear parameters such as load, sliding speed, sliding time, and sliding distance [3‐6]. However, materials parameters—particularly reinforcement composition (RCMP), reinforcement content (RCNT), and critically, their interaction (RCMP×RCNT)—remain inadequately studied. Understanding these interactive effects is essential because synergistic or antagonistic interactions between composition and content can significantly influence composite performance in ways that cannot be predicted from studying each parameter independently.

Incorporating hard phase particles (HPPs) such as silicon carbide (SiC) alongside solid lubricants (SLs) like graphite (Gr) has been shown to enhance the mechanical and tribological properties of ASAMCs [7, 8]. However, information on ternary reinforcement systems remains limited. This study introduces zinc oxide (ZnO) as a third reinforcement alongside SiC and Gr, representing the first systematic investigation of this ternary combination in hypereutectic Al-Si matrices. Previous studies have demonstrated ZnO’s capability to improve compressive strength, hardness, and wear resistance in aluminum-based systems through mechanisms including grain refinement, strong interfacial adhesion, and homogeneous distribution [9‐13]. The synergistic potential of combining ZnO with SiC and Gr warrants investigation, as each component serves distinct functions: SiC provides load-bearing capacity, Gr offers solid lubrication, and ZnO contributes to matrix strengthening and grain refinement.

The selection of ZnO as the third reinforcement addresses a critical gap in current composite design strategies. While SiC provides exceptional hardness (9.5 Mohs) and load-bearing capacity, and graphite offers lubrication through its lamellar structure, these extreme property contrasts (hard-brittle vs. soft-ductile) can lead to processing challenges and microstructural incompatibilities. ZnO, with intermediate hardness (4.5 Mohs), moderate density (5.6 g/cm³), and demonstrated grain refinement capabilities, serves as a bridging reinforcement facilitating compatible integration of otherwise incompatible phases. During high-temperature casting, ZnO decomposes with liberated oxygen forming Al₂O₃ at solidification interfaces that restricts grain growth, while zinc diffuses to promote heterogeneous nucleation of primary aluminum. Additionally, ZnO offers economic advantages as an industrial by-product, with costs approximately 40% lower than SiC, making it attractive for cost-sensitive automotive applications [13, 14].

Stir casting was selected for composite fabrication due to its proven reliability for particulate-reinforced composites, cost-effectiveness, and ability to achieve homogeneous particle dispersion regardless of particle size [15, 16]. This liquid metallurgy route enables industrial scalability, which is critical for practical applications.

From an economic and environmental sustainability perspective, this research utilizes recycled hypereutectic Al-Si piston scrap as the matrix material, addressing dual objectives of waste valorization and cost reduction. End-of-life automotive pistons, typically discarded after 150,000–200,000 km service life, contain 17–25% silicon and retain excellent metallurgical quality despite thermal cycling exposure [17, 18]. Remelting and reinforcing these scraps enable material circularity while avoiding the substantial energy costs associated with primary Al production (approximately 15 kWh/kg for virgin Al versus 0.5 kWh/kg for secondary processing), supporting circular economy principles where material costs constitute 40–50% of component production expenses [19, 20].

Despite extensive research on binary reinforcement systems, the systematic investigation of interactive effects between reinforcement composition (RCMP) and content (RCNT)—which may exceed individual factor influences—remains critically underexplored. Most optimization studies employ one-factor-at-a-time approaches or consider main effects independently, potentially missing optimal combinations achievable only through synergistic interactions. This study addresses this fundamental gap by explicitly modeling and quantifying two-factor interactions through hierarchical effects analysis, revealing for the first time that interaction terms can dominate tribological performance with effect magnitudes exceeding individual factor contributions by over 60%. The integration of multi-algorithm modeling approaches provides robust validation: Two-Factor Interaction (2FI) modeling from Design of Experiments (DOE) integrated with Desirability Function Analysis (DFA) for statistical modeling and parametric optimization, Artificial Neural Network (ANN) models integrated with Genetic Algorithm (GA) for nonlinear predictive modeling and global optimization, and Digital Image Processing (DIP) for quantitative microstructural characterization. The 2FI factorial model excels at systematically investigating factor interactions through designed experiments while minimizing experimental runs [21]. DFA enables multi-criteria optimization by converting individual responses into composite desirability functions [22]. ANNs capture complex nonlinear relationships through pattern recognition and self-learning capabilities [22], while GAs employ metaheuristic search based on natural selection principles to identify global optima [23]. DIP provides quantitative characterization of porosity and surface roughness from SEM images, enabling microstructure-property correlations [24].

This research comprises three integrated phases. Phase I investigates the influence of ZnO-SiC-Gr reinforcement with varying RCNT and RCMP, and their interaction, on the wear rate of hypereutectic ASAMCs through systematic experimentation. Phase II develops and compares predictive models—one Two-Factor Interaction model from DOE and two ANN models employing Levenberg-Marquardt backpropagation (LMBA) and Bayesian Regularization backpropagation (BRBA) algorithms—followed by wear rate optimization using DFA and GA. This comparative approach provides comprehensive insights into the relationships between wear rate and reinforcement parameters while establishing empirical models for predictive applications. Phase III employs DIP to quantify porosity and surface roughness from SEM images, correlating these microstructural characteristics with wear performance. Complementary fracture toughness testing via the circumferential notched tensile method was conducted to assess mechanical integrity, as high ceramic and graphite content can potentially cause embrittlement in aluminum matrix composites [5].

Experimental techniques include Taber rotary wear testing, circumferential notched tensile testing, SEM, energy-dispersive X-ray spectroscopy (EDX), spark spectroscopy, and X-ray fluorescence spectroscopy (XRF). Model validation employs Fisher’s Least Significant Difference, Kruskal-Wallis-Bonferroni procedure, Levene’s variance test, Mood’s median test, and analysis of variance (ANOVA). Analyses were performed using Design Expert 13 (2FI-DFA), MATLAB 2021a (ANN-GA), Statgraphics 19 (statistical validation), WSxM 5.0 (DIP), and OriginPro 2021 (data visualization).

2 Materials and methods

2.1 Raw materials

Hypereutectic Al-Si piston scraps (density: 2.68 g/cm³) sourced from local vendors in Mechanical Village, Ajaokuta, Kogi State, Nigeria served as the matrix material, enhancing economic viability and sustainability. Reinforcement particulates included zinc oxide (ZnO, 99.9% purity, 30 ± 2 μm, 5.61 g/cm³), silicon carbide (SiC, 99.5% purity, 30 ± 2 μm, 3.16 g/cm³), and graphite (Gr, 99.99% purity, 2.09 g/cm³), supplied by Pascal Scientific Ltd (Akure, Nigeria) and online vendors. Magnesium powder (99.99% purity) was used as a wetting agent to enhance particle-matrix interfacial bonding.

2.2 Experimental design

A multilevel categoric design (MLCD) approach within the Design of Experiments (DOE) framework was employed to systematically investigate the effects of reinforcement composition (RCMP, Factor A) and reinforcement content (RCNT, Factor B) on wear rate. Factor A comprised seven categorical levels representing different mixing ratios of ZnO-SiC-Gr (Table 1), while Factor B consisted of three levels (7, 9, and 11 wt%). This factorial arrangement yielded 21 unique combinations; each replicated three times in randomized order for a total of 63 experimental runs. Randomization and replication enhanced result reliability and minimized systematic errors [21].

Table 1

Experimental design structure, mixing ratios, and design power analysis

Sample Designation	RCMP mixing ratio (wt%)	RCNT (wt%)
	Factor A	Factor B (w7, w9, and w11)
RCMP 1 = Level 1	100 ZnO – 0 SiC – 0 Gr	7 = Level 1, 9 = Level 2, 11 = Level 3
RCMP 2 = Level 2	0 ZnO – 100 SiC – 0 Gr	7 = Level 1, 9 = Level 2, 11 = Level 3
RCMP 3 = Level 3	0 ZnO – 0 SiC – 100 Gr	7 = Level 1, 9 = Level 2, 11 = Level 3
RCMP 4 = Level 4	33.33 ZnO – 33.33 SiC – 33.33 Gr	7 = Level 1, 9 = Level 2, 11 = Level 3
RCMP 5 = Level 5	50 ZnO – 33.33 SiC – 16.67 Gr	7 = Level 1, 9 = Level 2, 11 = Level 3
RCMP 6 = Level 6	16.67 ZnO – 50 SiC – 33.33 Gr	7 = Level 1, 9 = Level 2, 11 = Level 3
RCMP 7 = Level 7	33.33 ZnO – 16.67 SiC – 50 Gr	7 = Level 1, 9 = Level 2, 11 = Level 3
Design Power	Response: Wear Index (mg/min)
Signal/Noise Ratio	Delta (Signal) = 2, Sigma (Noise) = 1	Signal/Noise = 2
Statistical Power	Power for Factor A = 86.9%	Power for Factor B = 99.9%

Factorial design: 7 levels of Factor A × 3 levels of Factor B = 21 unique combinations, replicated three times in randomized order (n = 63 total runs). Design power analysis confirms adequate experimental sensitivity to detect main effects and interactions. Design power analysis indicated detection capabilities of 86.9% for Factor A and 99.9% for Factor B at a signal-to-noise ratio of 2.0, confirming adequate experimental sensitivity for identifying significant effects and interactions

2.3 Composite fabrication

Composites were manufactured using a modified two-step stir casting technique (Fig. 1). The crucible furnace was preheated to 300 °C to minimize oxidation and enhance wettability between the matrix and reinforcements. Al-Si piston scraps were melted and heated to 750 ± 30 °C (above the liquidus temperature), then cooled to 650 °C for dross removal. Pre-blended reinforcement mixtures (preheated to 500 °C) corresponding to predetermined RCMP and RCNT combinations were gradually introduced into the melt along with magnesium powder (wetting agent) and carbon dioxide gas (oxidation minimizer). A particle distributor attachment prevented agglomeration during addition [25].

Fig. 1

Schematic flowchart of the improved two-step stir casting technique showing temperature-time progression, reinforcement addition sequence (preheated ZnO-SiC-Gr + Mg wetting agent + CO₂ oxidation minimizer), manual stirring at 650 °C (Step 1), and automatic stirring profile with progressive speed increase from 300 to 600 rpm (Step 2), culminating in mold pouring at 750 ± 30 °C

Bild vergrößern

In Step 1, manual stirring (4 min) was performed, followed by heating to 800 ± 30 °C above the composite liquidus temperature. Step 2 commenced with automatic mechanical stirring using progressively increasing speeds: 300 rpm (2 min), 450 rpm (4 min), and 600 rpm (5 min), ensuring homogeneous particle distribution. The composite melt was poured into preheated steel molds at 750 ± 30 °C and allowed to solidify under ambient conditions. This procedure was repeated for all 63 experimental combinations according to the randomized run order [26, 27].

2.4 Characterization techniques

2.4.1 Raw material characterization

Chemical composition and phase analysis of raw materials were conducted to verify purity and structural characteristics. The Al-Si alloy composition was determined using spark atomic emission spectroscopy (Spectrolab Jr CCD, SPECTRO). Elemental composition of ZnO and Mg powders was confirmed by X-ray fluorescence spectroscopy (XRF, Rigaku ZSX Primus II). Phase identification of graphite was performed using X-ray diffraction (XRD, PANalytical X’Pert Pro) with Cu-Kα radiation (λ = 1.5406 Å) over a 2θ range of 10–90°. Rietveld refinement was performed using PANalytical HighScore Plus 4 + software to quantify crystalline phases and structural parameters.

2.4.2 Density and porosity measurements

Density measurement is an important parameter for determining the efficiency of composite production techniques. The porosity percentage of the composites was determined through density measurements. Experimental density of each composite grade was determined using two methods: (i) the Archimedes principle, and (ii) gravimetric method by dividing the measured weight by volume. The theoretical density was calculated using the rule of mixtures. Equation (2.1) through (2.3) show the mathematical framework for evaluating experimental density and porosity percentage of the composites [28, 29].

Experimental density:

$$\rho\_Expc=Exp\_Wc/Exp\_Vc$$

(1)

Theoretical density (rule of mixtures):

$$\begin{array}{c}\rho\_TheC=(\rho\_mtx\times W\_mtx)+\\\sum(\rho\_rfc(n)\times W\_rfc(n))\end{array}$$

(2)

Porosity percentage:

$$\begin{array}{c}Porosity(\%)=\lbrack(\rho\_TheC-\rho\_Expc)\\/\rho TheC\rbrack\times100\end{array}$$

(3)

where:

ρ_Expc = experimental density of the composite (g/cm³)
Exp_Wc = actual weight of the composite (g) measured with a precision balance (± 0.00001 g, Adam Equipment, South Africa)
Exp_Vc = actual volume of the composite (cm³)
ρ_TheC = theoretical density of the composite (g/cm³)
ρ_mtx = density of the matrix (g/cm³)
ρ_rfc(n) = density of the nth reinforcement (g/cm³)
W_mtx = weight fraction of the matrix material
W_rfc(n) = weight fraction of the nth reinforcing phase

Three replicate measurements were performed for each composition to ensure statistical reliability.

2.4.3 Wear testing

Wear behavior was evaluated using a Taber Rotary Platform Dual Head Tester (Model 5135, Taber Industries, USA) according to ASTM G195-21 test standard [30]. Cylindrical specimens (diameter: 25 mm, thickness: 10 mm) were ground, polished, cleaned ultrasonically, and dried before testing. Tests were conducted under the following conditions: applied load of 9.8 N, sliding speed of 60 rpm, sliding distance of 1000 m, and CS-10 abrading wheels. Specimens were weighed before and after testing using a precision analytical balance (± 0.0001 g accuracy). Wear rate (mg/min) was calculated from mass loss and testing duration. Each composition was tested in triplicate, and average values were reported.

2.4.4 Microstructural characterization

Microstructural examination was performed using scanning electron microscopy (SEM, JEOL JSM-7600 F) equipped with energy-dispersive X-ray spectroscopy (EDX) for elemental analysis of composites. Specimens were sectioned, mounted in epoxy resin, ground progressively (320–2000 grit SiC paper), and polished to mirror finish using alumina suspension (0.05 μm). Samples were etched with Keller’s reagent prior to SEM observation [31].

2.4.5 Fracture toughness testing

Complementary fracture toughness assessment was conducted to evaluate the mechanical integrity of the composites, as high contents of ceramic reinforcements (SiC) and solid lubricants (Gr) have been reported to potentially cause embrittlement in aluminum matrix composites. Fracture toughness (K₁c) was assessed using the circumferential notched tensile (CNT) method proposed by Suresh et al. [32]. Cylindrical specimens (outer diameter: 5 mm, notch root diameter: 4 mm, d/D ratio: 0.8) were tested to failure on an Instron 5969 Universal Testing System equipped with a 2580 Series load cell at a crosshead speed of 0.5 mm/min. K₁c values were calculated from critical fracture loads using established empirical relationships and validated through complementary methods. This assessment ensures that optimization for wear reduction does not compromise structural integrity for practical applications.

2.5 Predictive modeling

2.5.1 Interaction effect modeling

A Two-Factor Interaction (2FI) model was developed using Design Expert 13 software to correlate wear rate with RCMP, RCNT, and their interaction. The model takes the form:

$$Y=\beta_o+\sum\beta_ix_i+\sum\beta_i{}_jx_ix_j+\varepsilon$$

(4)

where Y is the predicted wear rate, β₀ is the intercept, β_i are linear coefficients, β_iⱼ are interaction coefficients, x_i and xⱼ are coded factor levels, and ε is the random error. Model adequacy was assessed through analysis of variance (ANOVA), coefficient of determination (R²), adjusted R², predicted R², and lack-of-fit tests [21].

2.5.2 Artificial neural network models

Two feedforward artificial neural network (ANN) models were developed using MATLAB 2021a Neural Network Toolbox. The network architecture (Fig. 2) consisted of an input layer (2 neurons: RCMP and RCNT), one hidden layer with 5 neurons, and an output layer (1 neuron: wear rate). Training employed two algorithms:

ANN-LMBA: Levenberg-Marquardt backpropagation algorithm, known for fast convergence and high accuracy.
ANN-BRBA: Bayesian Regularization backpropagation algorithm, which minimizes overfitting through automatic regularization.

Fig. 2

Artificial neural network architecture (2-5-1) employed for wear rate prediction, showing input layer (2 neurons: reinforcement composition and content), hidden layer with five neurons using sigmoid activation function, and output layer (1 neuron: wear rate). Weights (w_iⱼ) and biases (b_i) were optimized during training using Levenberg-Marquardt and Bayesian Regularization backpropagation algorithms

Bild vergrößern

A 2–5-1 network architecture was adopted based on feature-size ratio optimization. The dataset was partitioned as follows: 60% training, 20% validation, and 20% testing. Network performance was evaluated using mean squared error (MSE), correlation coefficient (R), and regression plots. Overfitting was monitored by ensuring testing performance exceeded or matched training performance [33, 34].

2.6 Optimization strategies

2.6.1 Desirability function analysis

Numerical optimization was performed using the desirability function approach implemented in Design Expert 13. For wear rate minimization, an individual desirability function was constructed:

$$d(Y)={\lbrack(U-Y)/(U-L)\rbrack}^\wedge s\;for\;Y<U,\;and\;d(Y)=0\;otherwise$$

(5)

where Y is the predicted wear rate, L and U are the lower and upper bounds, and s is the weight factor. The overall desirability (D) ranges from 0 (completely undesirable) to 1 (ideal), with optimization seeking to maximize D while minimizing wear rate. Constraints were applied based on experimental ranges: RCMP (categorical levels 1–7) and RCNT (7–11 wt%) [22].

2.6.2 Genetic algorithm optimization

GA-based optimization was implemented in MATLAB 2021a to identify optimal RCMP-RCNT combinations yielding minimum wear rate. The ANN-BRBA model served as the fitness function due to its superior predictive accuracy. GA parameters included: population size [50], crossover probability (0.8), mutation probability (0.01), and maximum generations (100). The algorithm employed tournament selection, single-point crossover, and uniform mutation operators. Convergence was monitored through fitness evolution plots, with termination upon reaching predefined tolerance or maximum generations [23].

2.7 Digital image processing

Quantitative microstructural analysis was performed on SEM images using WSxM 5.0 software. Porosity was estimated by counting void features and calculating area fraction relative to total image area. Surface roughness parameters were extracted from grayscale intensity profiles, including arithmetic mean deviation (Ra) and root mean square roughness (Rq). Image preprocessing involved noise reduction, contrast enhancement, and threshold segmentation to isolate features of interest. Multiple regions per specimen were analyzed to ensure statistical reliability [24].

2.8 Statistical validation

Model comparison and validation employed multiple statistical tests:

Analysis of Variance (ANOVA): Assessed model significance through F-values and p-values (α = 0.05), evaluating sum of squares, degrees of freedom, and mean squares for model terms and residuals.
Fisher’s Least Significant Difference (LSD): Determined significant differences between factor levels through pairwise comparisons with 95% confidence intervals.
Kruskal-Wallis Test with Bonferroni Correction: Non-parametric alternative to ANOVA for comparing multiple groups, followed by post-hoc pairwise comparisons with adjusted p-values.
Levene’s Test: Assessed homogeneity of variance across groups, validating ANOVA assumptions.
Mood’s Median Test: Non-parametric test comparing medians across groups, providing robust analysis independent of distribution assumptions.

Model accuracy metrics included correlation coefficient (R), coefficient of determination (R²), adjusted R², predicted R², mean absolute error (MAE), and root mean square error (RMSE). Statistical analyses were conducted using Statgraphics Centurion 19.

3 Results and discussion

3.1 Raw material characterization

Spectroscopic analysis confirmed the composition and purity of all raw materials (Fig. 3). Spark atomic emission spectroscopy of the Al-Si alloy (Fig. 3a) revealed Si content of 14.02 wt%, confirming its hypereutectic classification (> 11.7% Si required) [5]. Al comprised 81.95 wt% of the alloy composition. Minor alloying elements typical of automotive piston alloys were detected, including copper (0.79 wt%), zinc (0.12 wt%), nickel (1.6 wt%), iron (0.61 wt%), magnesium (0.54 wt%), tin (0.71 wt%), and titanium (0.21 wt%). These trace elements are consistent with recycled piston scrap material and do not adversely affect composite fabrication.

Fig. 3

Raw material characterization: (a) Spark AES of hypereutectic Al-Si alloy showing 14.02 wt% Si; (b) XRF of ZnO powder (99.51% purity); (c) XRF of Mg powder (99.9% purity); (d) XRD pattern of graphite with Rietveld analysis showing 44.2% crystalline and 55.8% amorphous phases

Bild vergrößern

X-ray fluorescence analysis verified high purity of reinforcement precursors. The ZnO powder (Fig. 3b) exhibited 99.51 wt% purity, with trace impurities primarily comprising carbonates (0.25% CO₂) and sulfates (0.01% SO₄), well within acceptable limits for composite reinforcement. Similarly, Mg powder (Fig. 3c) demonstrated 99.9 wt% purity with minimal contamination from Zn (0.1%), iron (0.05%), and lead (0.01%).

X-ray diffraction with Rietveld refinement (Fig. 3d) characterized the Gr structure, revealing a dual-phase composition of 44.2% crystalline Gr 2 H and 55.8% amorphous content. The characteristic (002) reflection at 2θ ≈ 26° (Cobalt Kα radiation) confirmed the hexagonal layered structure of graphite. Additional peaks at higher angles correspond to (004), (100), (101), (102), and (110) reflections, validating the graphite phase identification. This crystalline-amorphous combination is beneficial for tribological applications, as crystalline Gr provides structural stability while the amorphous phase facilitates low-friction layer formation during sliding contact [14].

3.2 SEM-EDX elemental distribution analysis

SEM-EDX analysis of representative RCMP6 composites revealed the microstructural evolution with increasing reinforcement content (Fig. 4). Al exhibited the highest spectral peak intensity, followed by silicon, confirming the Al-Si matrix as the dominant phase. Oxygen peaks were detected at low intensity, indicating effective protection from atmospheric oxidation during the fabrication process. Additional elements including carbon, nickel, iron, copper, and magnesium were confirmed, with carbon presence validating Gr incorporation into the composites.

Fig. 4

SEM-EDX microstructural analysis of RCMP6 composites at (a) 7 wt%, (b) 9 wt%, and (c) 11 wt% reinforcement content showing Al-Si-rich regions, tertiary plate structures, micro-voids, and particle clusters. EDX spectra confirm matrix and reinforcement elements. Scale bar: 100 μm

Bild vergrößern

Notably, Zn was not detected by EDX analysis, which can be attributed to two factors: (i) limited equipment sensitivity in the analyzed region, and (ii) thermal decomposition of ZnO at elevated processing temperatures. Previous studies have demonstrated that ZnO addition to hypereutectic Al-Si alloys at high temperatures leads to decomposition into Zn and O, with O preferentially combining with Al to form Al₂O₃ at the liquid-solid interface [11, 35]. This oxygen diffusion restricts primary Al grain growth during solidification, while dissolved Zn at the solid-liquid interface further contributes to growth restriction. Both decomposition products (O and Zn) serve as heterogeneous nucleation sites for primary Al grains, with Al₂O₃ formation in the eutectic region confirming this growth restriction mechanism. The observed high-intensity magnesium peaks correspond to Mg₂Si precipitates, which are characteristic Mg-rich phases in Al-Si-Mg systems [36].

Microstructural analysis revealed several key features (Fig. 4a-c, left panels). The region marked by a green circle represents an Al-Si-rich phase with homogeneous matrix structure. The red ellipse highlights a tertiary side-plate structure within star-shaped silicon rods, characteristic of eutectic morphology. These tertiary plate structures commonly form during melt solidification of Al-Si eutectic binary alloys, developing as regular arrays of vanes emanating from one or both sides of a parent plate. The plates extend from near the core to the leading growth edge, with their extension governed by the local solute field, confirming the eutectic nature of the as-received piston scraps [37].

The purple ellipse identifies one of the few micro-voids observed in the RCMP6_w11 composite. Overall, void content remained minimal across all analyzed samples, indicating high casting quality. The yellow ellipse reveals particle clustering (agglomeration) in the RCMP6_w11 composite. This clustering phenomenon is expected at higher reinforcement contents, as particle-particle interactions intensify and eventually exceed particle-matrix interfacial bonding strength, promoting agglomerate formation [38, 39].

3.3 Porosity characterization via digital image processing

Comprehensive porosity analysis across all seven reinforcement compositions (Fig. 5a) revealed distinct trends between mono-phase inclusions (RCMP1-3) and ternary inclusions (RCMP4-7). Mono-phase composites exhibited porosity levels ranging from 1.32% (RCMP2_w9) to 3.62% (RCMP1_w11), with RCMP1 (100% ZnO) showing the highest susceptibility to void formation. Ternary inclusion composites displayed wider porosity variation (1.54–3.58%), with RCMP7 (33.33% ZnO – 16.67% SiC – 50% Gr) exhibiting the maximum porosity of 3.58% at 11 wt% reinforcement. Notably, RCMP4 (equal ternary blend) demonstrated an anomalous trend, with porosity decreasing from 3.04% at 9 wt% to 1.71% at 11 wt%, suggesting optimal particle packing at higher reinforcement content for this specific composition [39].

Fig. 5

(a): Porosity of composites across reinforcement compositions (RCMP1-7) and contents (w7, w9, w11). Mono-phase and ternary inclusions show different porosity trends, with maximum at RCMP7_w11 (3.58%). (b): WSxM software porosity analysis of RCMP6 composites at (a) 7, (b) 9, and (c) 11 wt% showing increasing void count (22→33→36). Actual SEM images (left) and 3D flooded reconstructions (right) visualize porosity distribution

Bild vergrößern Bild vergrößern

The general trend of increasing porosity with reinforcement content can be attributed to several factors: (i) enhanced particle clustering at higher loading levels (as observed in Fig. 4c), which creates interparticle voids; (ii) increased melt viscosity during casting, hindering gas escape; and (iii) reduced wettability between reinforcement particles and matrix at higher particle volume fractions. The graphite-rich composition (RCMP7) exhibited particularly high porosity, likely due to poor wettability between graphite and aluminum, despite magnesium addition as a wetting agent. Conversely, the ZnO-rich composition (RCMP5) showed relatively low porosity at 7 wt% (1.54%), which may be associated with ZnO decomposition and grain refinement effects discussed previously (Sect. 3.X), leading to improved melt fluidity and reduced gas entrapment [40‐42].

These porosity results correlate with wear performance trends, where compositions exhibiting excessive porosity (> 3.5%) demonstrated diminished wear resistance due to reduced load-bearing capacity and increased subsurface crack propagation during tribological contact [40, 42].

Digital image processing of SEM micrographs using WSxM 5.0 software quantified porosity levels across all composite compositions (Figs. 5-6). Analysis of RCMP6 composites revealed a progressive increase in void count with reinforcement content: 22 voids at 7 wt%, 33 voids at 9 wt%, and 36 voids at 11 wt% (Fig. 5b). The 3D flooded visualization technique effectively identified and quantified microscale porosity, with detected void sizes ranging from 0.00002 μm² minimum to maximum heights of approximately 0.50–0.51 μm.

Fig. 6

Theoretical and experimental density of hypereutectic Al-Si ternary composites with varying SiC-Gr-ZnO reinforcement compositions (RCMP1-RCMP6) and contents (w1-w11: 1–11 wt%). Close agreement indicates effective consolidation with minimal porosity

Bild vergrößern

3.4 Density results

Density plays a major role in fuel consumption of transport vehicles, with weight reduction directly impacting efficiency. Recent studies indicate that every 100 kg weight reduction decreases combined city/highway fuel consumption by approximately 0.4 L/100 km for passenger cars and 0.5 L/100 km for trucks [43, 44]. Given that two of the three reinforcements employed in this study—SiC (3.16 g/cm³) and ZnO (5.61 g/cm³)—possess densities greater than the Al-Si matrix (2.68 g/cm³), meticulous attention to composite density is essential for automotive applications.

Theoretical and experimental density measurements revealed composition-dependent trends influenced by reinforcement density and porosity (Fig. 6). Theoretical densities ranged from 2.45 g/cm³ (RCMP3_w11) to 3.17 g/cm³ (RCMP5_w11), reflecting the density hierarchy of reinforcement phases: ZnO (5.61 g/cm³) > SiC (3.16 g/cm³) > Gr (2.09 g/cm³). The maximum experimental density of 3.17 g/cm³ was observed for RCMP5_w9 (ZnO-rich composition), representing a 16.75% increase relative to the monolithic Al-Si alloy. Notably, most composites exhibiting densities greater than the base alloy (2.68 g/cm³) contained 9 wt% reinforcement, with experimental values ranging from 2.92 to 3.17 g/cm³.

At lower (7 wt%) and higher (11 wt%) reinforcement contents, most composites exhibited densities below the base alloy, despite the higher intrinsic density of SiC and ZnO. This counterintuitive trend can be attributed to competing effects: (i) At 7 wt%, low reinforcement loading provides insufficient density enhancement to overcome porosity-induced density reduction (2.2–3.6% porosity); (ii) At 11 wt%, increased porosity (2.6–3.6%) and graphite content (especially in RCMP3 and RCMP7) dominate, reducing effective density despite higher ceramic content. The intermediate 9 wt% content achieves optimal balance between reinforcement-induced density increase and porosity-related density decrease, explaining the observed peak density values at this loading.

An exception to this trend was RCMP4_w11 (equal ternary blend), which exhibited anomalously high density (2.99 g/cm³ experimental, 3.04 g/cm³ theoretical) with minimal porosity (1.6%). This suggests that the equal ternary composition achieves superior particle packing at 11 wt%, reducing void space and maximizing consolidation efficiency. Conversely, graphite-rich compositions (RCMP7) consistently exhibited lower densities across all reinforcement levels due to graphite’s low intrinsic density (2.09 g/cm³), even at minimal porosity levels.

The observed density trends align with previous studies on ceramic-reinforced aluminum composites. Bandil et al. reported similar SiC-induced density increases in Al-Si composites [45], while Alaneme et al. demonstrated that graphite incorporation reduced Al-Mg-Si composite density compared to steel particle-only reinforcement [46]. The competition between reinforcement density, porosity, and particle packing efficiency explains the absence of a simple monotonic relationship between reinforcement content and composite density. For lightweight automotive applications, compositions such as RCMP7 (Gr-rich) at lower porosity levels offer weight reduction benefits, whereas applications requiring structural integrity may favor denser compositions like RCMP5_w9, accepting the weight penalty for enhanced mechanical properties.

3.5 Mechanical integrity assessment

Fracture toughness (K₁c) is the ability of a material with pre-existing cracks to resist fracture, a critical property since flaws are inevitable during processing and fabrication [47]. Figure 7 presents the K₁c values of the monolithic Al-Si alloy (MASA) and the ASAMCs. All composites outperformed the MASA in resisting crack propagation, with K₁c values ranging from 13.3 to 28.6 MPa·m^1/2 compared to 6.47 MPa·m^1/2 for MASA, representing 106–342% improvement. The most significant enhancement (342%) was achieved by RCMP6_w11, while RCMP1_w7 showed the least improvement (32%).

Fig. 7

Fracture toughness (K₁c) of MASA and ASAMCs with varying reinforcement compositions (RCMP1-RCMP7) and contents (w7, w9, w11). Ternary inclusions show superior performance with 342% improvement at RCMP6_w11

Bild vergrößern

Composites containing ternary inclusions (RCMP4-7) exhibited superior K₁c values compared to those with mono inclusions (RCMP1-RCMP3). For mono inclusions, K₁c increased with reinforcement content for SiC (RCMP1) and ZnO (RCMP2), but decreased for Gr (RCMP3) beyond 7 wt%. This trend persisted in ternary composites, indicating that while high contents of SiC and ZnO are favorable, excessive Gr (> 7 wt%) is detrimental to K₁c whether as mono or ternary inclusion. The enhancement at 7 wt% Gr results from its ability to prevent dimple initiation into internal cracks, while higher contents cause particle agglomeration and clustering [48].

The superior performance of SiC-containing composites can be attributed to interfacial strengthening mechanisms during high-temperature casting. Silicon dissolved from SiC particles acts as a substrate for primary Mg₂Si nucleation, which subsequently serves as a heterogeneous nucleation site for α-Al, lowering interfacial energy. The resulting Al-Mg₂Si binary eutectic structure creates strong bonding between the Al-Si matrix and SiC particles. Additionally, the hard, high-modulus reinforcements act as obstacles to plastic flow and dislocation movement, with diminished inter-particle spacing at higher contents further enhancing resistance [49, 50].

The remarkable K₁c improvements in RCMP5_w11 (28.2 MPa·m^1/2) and RCMP6_w11 (28.6 MPa·m^1/2) demonstrate that incorporating ZnO as a complementary reinforcement does not compromise toughness—contrary to typical MMC behavior where strength gains occur at the expense of ductility [51]. The relative softness of ZnO, combined with optimal reinforcement ratios, enables these ternary composites to achieve both high wear resistance (from the main study) and adequate fracture toughness. The fracture mode observed was mixed ductile-brittle, characterized by matrix dimple formation and interfacial debonding between particles and matrix under applied stress, rather than purely brittle failure [32].

Importantly, the K₁c values of the optimized compositions (RCMP4-6) exceed 20 MPa·m^1/2, meeting the typical requirement for automotive cylinder liner applications [3, 44]. This validates that optimization for minimum wear rate does not compromise structural integrity—a critical consideration where both tribological performance and fracture resistance are simultaneously required.

3.6 Wear performance

Figure 8a furnishes the Taber wear index (TWI) values of the MASA and ASAMCs, revealing that reinforcement additions significantly lowered the wear rate with all composites exhibiting an inverse relationship between RCNT and wear rate. The MASA exhibited the highest wear index (7810 mg/min), while composites demonstrated progressive wear reduction with increasing RCNT. For mono inclusion composites, the greatest reduction was observed for RCMP3_w11 (100% Gr @ 11 wt%, 3833 mg/min, 51% reduction), while for ternary inclusion composites, RCMP7_w11 (33.33% ZnO − 16.67% SiC − 50% Gr @ 11 wt%, 197 mg/min) achieved the most remarkable performance with 97.48% reduction compared to MASA—demonstrating the efficacy of ternary reinforcement strategies.

Fig. 8

(a): Taber wear index of MASA and ASAMCs [52]with varying reinforcement compositions (RCMP1-RCMP7) and contents (w7, w9, w11). Ternary inclusions exhibit superior wear resistance with 97.48% reduction at RCMP7_w11. (b): SEM images and digital roughness analysis of worn surfaces for RCMP6 composites at (a) w7, (b) w9, and (c) w11. RMS roughness values of 0.2885, 0.2880, and 0.2879 μm indicate consistent surface quality with mild abrasive wear characteristics

Bild vergrößern Bild vergrößern

Composites with mono inclusions (RCMP1-3) showed substantial wear reduction but exhibited composition-dependent behavior. SiC mono-inclusion composites (RCMP1) achieved 22–35% reduction (6051 − 5039 mg/min at 7–11 wt%), ZnO mono-inclusions (RCMP2) showed 34–38% reduction (5183 − 4850 mg/min), while Gr mono-inclusions (RCMP3) demonstrated the highest reduction among mono systems at 34–51% (5133 − 3833 mg/min). However, composites with ternary inclusions consistently outperformed their mono-inclusion counterparts across all RCNTs, with RCMP4-RCMP7 achieving 37–97.48.48% wear reduction. This superior performance stems from the hybridization of strengthening mechanisms rather than reliance on individual mechanisms characteristic of mono-inclusion systems [41‐45].

The wear resistance enhancement mechanisms are reinforcement-specific and synergistic in ternary systems. For ZnO-containing composites, ZnO particles prevent dislocation motion and increase both hardness and wear resistance concurrently provided the critical loading content is not exceeded. During abrasive loading, dislodged ZnO particles fracture and enhance the contact area among sliding surfaces, consecutively reducing wear [52]. For SiC-containing composites, SiC particles refine the eutectic Si phase, providing excellent wear resistance and hardness. The hard SiC particles (9.5 Mohs hardness) act as primary load-bearing reinforcements, resisting penetration by abrading asperities and protecting the softer matrix from ploughing and micro-cutting [53, 54]. For Gr-containing composites, Gr acts as a solid lubricant that stabilizes the friction coefficient and reduces the wear process. As Gr content increases with appropriate porosity, the thickness and strength of the tribo-layer increases through participation of more lubricant particulates. The Gr particles smear and form a lubricating film/layer between contacting surfaces, decreasing metal-to-metal interaction and imparting self-lubricating properties to the composites. This smeared Gr layer thickness increases with increasing volume fraction, reducing the coefficient of friction and minimizing contact between wear pin and disc [14, 48, 53].

The transition to ternary inclusions (RCMP4-7) resulted in synergistic wear resistance enhancement far exceeding individual reinforcement contributions. RCMP4 achieved 37–51% reduction (4906 − 3816 mg/min), RCMP5 demonstrated 41–60% reduction (4606 − 3100 mg/min), RCMP6 showed 52–68% reduction (3746 − 2500 mg/min), and RCMP7 exhibited 60–97.48.48% reduction (3103 − 197 mg/min), with wear reduction increasing progressively with RCNT. The synergistic combination of hard SiC particles (load-bearing and eutectic Si refinement), soft Gr flakes (solid lubrication and tribo-layer formation), and intermediate-hardness ZnO particles (dislocation prevention and grain refinement) creates a multifunctional hierarchical microstructure where each phase fulfills a specific tribological function. While only individual strengthening mechanisms are responsible for improved wear resistance in mono-inclusion composites, the hybridization of these mechanisms in ternary composites produces superior performance through concurrent load-bearing, lubrication, and microstructural refinement effects [6, 14, 16, 48, 52].

The progressive improvement with increasing reinforcement content (w7→w9→w11) reflects enhanced reinforcement density providing greater protection. For ternary systems RCMP4-7, the benefits dominate up to 11 wt%, indicating excellent particle dispersion achieved through optimized stir casting parameters. The exceptional 97.48% wear reduction achieved by RCMP7_w11 positions this composition as highly promising for cylinder liner applications, where cast iron (7800–7900 kg/m³) is conventionally used [17]. The RCMP7_w11 composite offers approximately 61% weight reduction (estimated density ~ 3050 kg/m³) while providing superior wear resistance, potentially enabling significant fuel efficiency improvements in automotive engines through reduced reciprocating mass.

3.7 Worn surface morphology and digital roughness analysis

Figure 8b presents the morphologies of worn surfaces and corresponding digital roughness analysis for RCMP6 composites selected as representative samples (optimal formulation in terms OF k_1C and wear resistance) for the ASAMCs. The SEM images reveal characteristic features associated with abrasive wear mechanisms including wear tracks, grooves, ploughs, plastic deformation zones, holes, craters, and material accumulation along wear paths. These features are anticipated due to the relatively low load magnitude and rotational speed employed during Taber wear testing, which promotes abrasive rather than adhesive wear modes [6, 55].

The RCMP6_w7 composite exhibits highly visible grooves and ploughs indicative of both two-body and three-body abrasive mechanisms operating concurrently. Two-body abrasion occurs when hard reinforcement particles protruding from the composite surface directly plough the counter-surface, while three-body abrasion results from loose wear debris trapped between sliding surfaces acting as abrasive media. The SEM images reveal grooves running parallel to the sliding direction with evidence of significant plastic deformation and material removal. Clear evidence of severe abrasive action is shown by the contact zone exhibiting spalling and accumulated materials over the wear track, with holes originating from detachment of small particles from the matrix material. In contrast, RCMP6_w9 displays fairly visible grooves and microvoids with reduced severity compared to w7, while RCMP6_w11 shows visible grooves, ploughs, clusters, and microvoids but with the least severe damage among the three compositions, demonstrating the progressive improvement in wear resistance with increasing reinforcement content [55].

DIP quantification reveals remarkably consistent surface quality across the three reinforcement contents. The RMS roughness values measured were 0.2885 μm for RCMP6_w7, 0.2880 μm for RCMP6_w9, and 0.2879 μm for RCMP6_w11. Despite the visual differences in groove visibility and wear severity observed in SEM images, the quantitative roughness measurements indicate excellent wear stability and uniform material removal characteristics. The consistently low roughness values below 0.3 μm suggest that while localized severe features exist, the overall wear mechanism remains predominantly controlled mild abrasive wear rather than catastrophic adhesive or delamination wear [55‐57]. The roughness profiles show relatively uniform height distributions with peak-to-valley variations of approximately 0–1 μm, confirming consistent surface topography without catastrophic material removal events. The severity of abrasive wear can be qualitatively assessed by the degree of visibility of wear features in the images, with RCMP6_w7 showing the most pronounced features and RCMP6_w11 the least, correlating with the decreasing wear indices observed in Fig. 8b and validating that increased RCNT enhances wear resistance while maintaining excellent surface integrity suitable for demanding cylinder liner applications.

4 Modeling and optimization

4.1 Two-Factor interaction factorial modeling and statistical validation

2FI modeling from DOE was utilized to establish predictive relationships between reinforcement composition (Factor A: RCMP), reinforcement content (Factor B: RCNT), and wear response. The experimental design incorporated 63 experimental runs spanning seven RCMPs (RCMP1-RCMP7) and six RCNTs (7, 9, 11 wt.% for complementary studies plus additional intermediate levels). Due to the wide response range spanning 180–7810 mg/min across the design space, preliminary analysis revealed violation of normality assumptions with non-constant variance (heteroscedasticity) in residuals. Box-Cox transformation analysis recommended an inverse square root transformation (λ = −0.5), yielding the transformed response Y’ = 1/√(Wear Index). This transformation successfully stabilized variance, normalized residual distribution, and linearized factor-response relationships, enabling robust statistical modeling with all subsequent analyses conducted on the transformed scale [58, 59].

Table 2

ANOVA table

Source	Sum of Squares	df	Mean Square	F-value	p-value	Significance
Model	0.0091	20	0.0005	5049.18	< 0.0001	Significant
A-Comp	0.0011	2	0.0005	5933.17	< 0.0001
B-Grade	0.0037	6	0.0006	6839.94	< 0.0001
AB	0.0043	12	0.0004	4006.47	< 0.0001
Pure Error	3.786E-06	42	9.014E-08
Cor Total	0.0091	62

ANOVA results presented in Table 2 demonstrate exceptional model significance. The model F-value of 5049.18 with p < 0.0001 indicates only a 0.01% probability that such statistical significance could occur due to random noise, confirming the model’s validity. Individual factor assessment revealed that Factor A representing RCMP exhibited F-value of 5933.17, Factor B representing RCNT exhibited F-value of 6839.94, and critically, their interaction term AB exhibited F-value of 4006.47, with all p-values below 0.0001. The half-normal probability plot (Fig. 9a) provides visual confirmation of effect magnitudes, showing that AB interaction, A-RCMP, and B-RCNT deviate significantly from the reference line indicating genuine effects rather than random error. Notably, the interaction effect AB exhibits the largest normal effect magnitude of 156.15, exceeding the individual effects of Factor A at 95.0 and Factor B at 94.8. This finding provides quantitative evidence that RCMP and RCNT act synergistically rather than independently, validating the fundamental premise that optimal wear performance requires simultaneous optimization of both compositional ratios and loading levels rather than isolated manipulation of single variables [60].

Fig. 9

2FI model validation and response surface visualization: (a) Half-normal plot showing significant effects: AB interaction (156.15), A-RCMP (95.0), B-RCNT (94.8); (b) Predicted vs actual values (R²=0.9996); (c) Residuals vs run order; (d) Normal probability of residuals; (e) 3D surface showing synergistic interaction with concave curvature; (f) Contour plot with wear index ranging from 6020 mg/min (RCMP1_w7) to 180 mg/min (optimal region RCMP6-7_w11)

Bild vergrößern

Model adequacy was confirmed through multiple statistical metrics. The model shows an excellent fit: the residual standard deviation is 0.0003 and the response mean is 0.0183, giving a coefficient of variation of 1.64%. The coefficient of determination (R²) is 0.9996—i.e., 99.96% of the variance is explained—and the adjusted R² is 0.9994, indicating the fit remains strong after accounting for model complexity. The predicted R² of 0.9991 aligns with strong cross-validated performance. Adequate Precision is 336.9838, far above the recommended minimum of 4, signifying a very high signal-to-noise ratio. The correlation coefficient R² of 0.9996 indicates that 99.96% of total variation in transformed wear response is explained by the model, with only 0.04% attributable to residual error. The adjusted R² of 0.9994 accounts for the number of predictors, and its proximity to R² confirms absence of overfitting with all included terms contributing meaningfully. The predicted R² of 0.9991, obtained through leave-one-out cross-validation, demonstrates excellent predictive stability, with the difference between adjusted and predicted R² of 0.0003 falling well below the recommended threshold of 0.2. Adequate precision of 336.98 far exceeds the minimum acceptable value of 4, confirming the model’s ability to navigate the design space and detect genuine effects against experimental noise. The coefficient of variation of 1.64% indicates excellent experimental reproducibility and precise control of fabrication parameters during composite processing, with such low variation confirming that observed wear responses are predominantly systematic responses to controlled factors rather than random experimental error [60].

Diagnostic plots presented in Fig. 9 validated model assumptions and predictive capability. The predicted versus actual plot (Fig. 9b) demonstrates tight clustering of data points around the 45-degree reference line across the entire wear index range, confirming model accuracy. The normal probability plot of residuals (Fig. 9d) shows residuals aligned reasonably well with theoretical normal distribution, validating the effectiveness of the inverse square root transformation. The residuals versus run order plot (Fig. 9c) exhibits random scatter without time-dependent trends or systematic patterns, confirming experimental stability throughout the study [61].

The final predictive equation in terms of coded factors, with all coefficients statistically significant at p < 0.0001, is expressed as:

$$\begin{array}{c}1/\sqrt(Wear\;Index)=+0.0183-0.0034\times\\A\lbrack1\rbrack-0.0024\times A\lbrack2\rbrack-0.0046\times B\lbrack1\rbrack-0.0043\\\times B\lbrack2\rbrack-0.0035\times B\lbrack3\rbrack-0.0030\times B\lbrack4\rbrack-0.0020\\\times B\lbrack5\rbrack-0.0012\times B\lbrack6\rbrack+0.0026\times A\lbrack1\rbrack B\lbrack1\rbrack+\\0.0028\times A\lbrack2\rbrack B\lbrack1\rbrack+0.0032\times A\lbrack1\rbrack B\lbrack2\rbrack+0.0022\\\times A\lbrack2\rbrack B\lbrack2\rbrack+0.0025\times A\lbrack1\rbrack B\lbrack3\rbrack+0.0019\times A\lbrack2\rbrack B\\\lbrack3\rbrack+0.0023\times A\lbrack1\rbrack B\lbrack4\rbrack+0.0026\times A\lbrack2\rbrack B\lbrack4\rbrack+\\0.0018\times A\lbrack1\rbrack B\lbrack5\rbrack+0.0019\times A\lbrack2\rbrack B\lbrack5\rbrack\\+0.0030\times A\lbrack1\rbrack B\lbrack6\rbrack+0.0017\times A\lbrack2\rbrack B\lbrack6\rbrack\end{array}$$

(7)

where A[i] and B[j] represent coded categorical levels for RCMP composition and RCNT content respectively. The intercept of + 0.0183 represents the transformed wear response at the reference level. Negative coefficients for main effects (A[1] = −0.0034, A[2] = −0.0024, B[1] through B[6] ranging from − 0.0046 to −0.0012) indicate that increasing either RCMP number or RCNT content improves wear resistance by decreasing the transformed response, which corresponds to reduced wear index in original units. Conversely, positive interaction coefficients (ranging from + 0.0017 to + 0.0032) reveal synergistic combinations producing superior performance beyond additive predictions, with the largest interaction coefficient of + 0.0032 for A[1]B[2] indicating particularly favorable synergy at specific RCMP-RCNT combinations consistent with RCMP6-RCMP7 at 11 wt% achieving optimal results experimentally.

Response surface visualization through three-dimensional surface (Fig. 9e) and two-dimensional contour plot (Fig. 9f) illustrates wear index decreasing progressively from 6020 mg/min at RCMP1 with 7 wt% toward approximately 180 mg/min in the optimal region. The concave curvature of the three-dimensional surface provides geometric proof of synergy, with combined effects exceeding the sum of individual contributions. The steepest gradient occurs along the diagonal from lower RCMP and RCNT toward higher values, confirming that simultaneous increases in both factors produce maximum wear reduction. This curvature validates that the two-factor interaction term is not merely statistically significant but physically meaningful, representing genuine synergistic mechanisms where ternary reinforcement architectures at higher loading densities create hierarchical microstructures with emergent tribological properties unavailable to mono-inclusion systems or lower reinforcement contents independently.

4.2 Artificial neural network performance evaluation

Two ANN models employing Levenberg-Marquardt (ANN-LMBA) and Bayesian Regularization (ANN-BRBA) training algorithms were evaluated as alternative predictive frameworks. Training performance results demonstrate that both ANN architectures achieved excellent convergence characteristics. ANN-LMBA reached best validation performance of 1.0837 × 10⁻⁷ at epoch 35 (Fig. 10a), with mean squared error decreasing from initial values exceeding 10⁻² to final training MSE of 5.7597 × 10⁻⁸. The close tracking of training, validation, and test curves throughout the learning process indicates absence of overfitting, with the network generalizing effectively to unseen data. ANN-BRBA demonstrated even more rapid convergence, achieving best training performance of 5.7597 × 10⁻⁸ at epoch 51 (Fig. 10b), with the absence of separate validation curves reflecting the algorithm’s inherent regularization mechanism that eliminates need for explicit validation-based early stopping [33, 34].

Fig. 10

Performance comparison of ANN-LMBA and ANN-BRBA models. (a) MSE convergence for ANN-LMBA (35 epochs, best validation: 1.0837×10⁻⁷). (b) ANN-LMBA regression plots (R=0.99955). (c) MSE convergence for ANN-BRBA (51 epochs, best training: 5.7597×10⁻⁸). (d) ANN-BRBA regression plots showing superior accuracy (R=0.99979). (e): Prediction performance comparison for normalized wear index across 63 experimental runs: The figure shows actual values (red) versus predicted values from 2FI model (green, top left), LMANN model (blue, top right), BRANN model (green, bottom left), and combined comparison of all three models (bottom right). All models accurately capture the peak wear index values occurring around runs 5, 40, and 50

Bild vergrößern Bild vergrößern

Regression analysis reveals exceptional predictive accuracy for both neural network models. ANN-LMBA achieved correlation coefficients of R = 0.99968 for training, R = 0.98857 for validation, R = 0.98955 for testing, and R = 0.99955 overall (Fig. 10c). ANN-BRBA demonstrated superior performance with R = 0.99977 for training, R = 0.99994 for testing, and R = 0.99979 overall (Fig. 10d). The tight clustering of data points around the 45-degree ideal line across all data partitions confirms both networks’ capability to capture complex nonlinear relationships between reinforcement parameters and wear response with minimal deviation. The marginal performance advantage of BRBA over LMBA stems from Bayesian regularization’s automatic weight decay and optimal network complexity selection, though both models achieve prediction errors well within experimental variation, validating their suitability for subsequent optimization applications [23, 33, 34].

4.3 Comparative evaluation of predictive models

Comprehensive comparison of all three modeling approaches—2FI factorial, ANN-LMBA, and ANN-BRBA—presented in Table 3; Fig. 10e reveals that each method provides highly accurate predictions with correlation coefficients exceeding 0.999. The actual versus predicted plots for individual models show excellent agreement across the entire response range, with predictions tracking experimental values from the lowest wear indices near 200 mg/min for optimal ternary compositions through the highest values exceeding 6000 mg/min for mono-inclusion systems. Summary statistics (Table 3) demonstrate remarkable consistency across all 63 experimental runs, with mean predicted values of 0.0182991 (2FI), 0.0183198 (LMBA), and 0.0182298 (BRBA) clustering tightly around the actual mean of 0.0182991 on the transformed scale. Standard deviations show equally tight agreement at 0.0121171 (2FI), 0.0121104 (LMBA), and 0.0121798 (BRBA) compared to actual standard deviation of 0.0121195, with coefficients of variation ranging from 66.10% to 66.56% reflecting the wide dynamic range spanning four orders of magnitude rather than indicating poor precision.

Table 3

Comprehensive model comparison statistics

Model	Count	Mean	Std Dev	COV (%)	Minimum	Maximum	Range
(a) Summary Statistics
Actual	63	0.0182991	0.0121195	66.23	0.0129099	0.0725476	0.0596377
2FI Predicted	63	0.0182991	0.0121171	66.22	0.0129099	0.0725476	0.0596377
LMBA Predicted	63	0.0183198	0.0121104	66.11	0.0128397	0.0714034	0.0585637
BRBA Predicted	63	0.0182298	0.0121798	66.56	0.0129743	0.0712986	0.0583243
(b) ANOVA Comparison
Source	Sum of Squares	DF	Mean Square	F-Ratio	P-Value
Between Groups	2.09939E-8	3	6.99796E-9	0.00	1.0000
Within Groups	0.0365003	248	0.000147179
Total (Corr.)	0.0365004	251
(c) Variance Homogeneity Tests
Test	Statistic	P-Value	Interpretation
Levene’s Test	0.0000811975	1.000	Equal variances confirmed
(d) Pairwise Variance Comparisons (F-Tests):
Comparison	F-Ratio	P-Value
Actual/2FI Predicted	1.00039	0.9988
Actual/LMBA Predicted	1.0015	0.9953
Actual/BRBA Predicted	0.990108	0.9689
2FI/LMBA Predicted	1.00111	0.9965
2FI/BRBA Predicted	0.989723	0.9677
LMBA/BRBA Predicted	0.988625	0.9642
Non-Parametric Tests
Kruskal-Wallis Test:
• Test Statistic: 0.0651275
• P-Value: 0.995665
• Conclusion: No significant differences among groups
(e) Kruskal-Wallis Test
Sample	Sample Size	Average Rank
Actual	63	128.444
2FI Predicted	63	126.365
LMBA Predicted	63	125.429
BRBA Predicted	63	125.762
Mood’s Median Test:
• Total n = 52
• Grand Median = 0.0153423
(e) Mood’s Median Test
Sample	Sample Size	n≤	n>	Median	95.0% Lower CL	95.0% Upper CL
Actual	63	31	32	0.0154303	0.0142857	0.0162221
2FI Predicted	63	31	32	0.0153899	0.0143164	0.0162212
LMBA Predicted	63	33	30	0.0151807	0.0143059	0.0166498
BRBA Predicted	63	33	30	0.0153423	0.014284	0.0162145
Multiple Range Tests (95.0% LSD)
Homogeneous Groups (all marked X):
• BRBA Predicted: X
• 2FI Predicted: X
• Actual: X
• LMBA Predicted: X
(f) Pairwise Contrasts:
Contrast	Difference	+/- Limits
Actual – 2FI Predicted	0	0.00425736
Actual – LMBA Predicted	-0.000020704	0.00425736
Actual – BRBA Predicted	0.00000105414	0.00425736
2FI – LMBA Predicted	-0.000020704	0.00425736
2FI – BRBA Predicted	0.00000105414	0.00425736
LMBA – BRBA Predicted	0.000021758	0.00425736
Critical Threshold: 0.00425736
Maximum Contrast: 0.000021758 (well below threshold)

Summary: All three predictive models (2FI, ANN-LMBA, ANN-BRBA) produce statistically indistinguishable predictions that are equivalent to actual experimental measurements across all statistical tests performed

Range statistics further validate model concordance, with minimum predicted values of 0.0129099 (2FI), 0.0128397 (LMBA), and 0.0129743 (BRBA) closely matching the actual minimum of 0.0129099, while maximum predicted values of 0.0725476 (2FI), 0.0714034 (LMBA), and 0.0712986 (BRBA) approximate the actual maximum of 0.0725476. The resulting ranges of 0.0596377 (2FI), 0.0585637 (LMBA), and 0.0583243 (BRBA) versus actual range of 0.0596377 demonstrate that all models successfully capture the full extent of wear behavior variation across the design space. Standard distance measures (SDS ranging from 13.4924 to 13.4949, SDK ranging from 26.5419 to 26.6053) confirm negligible deviations among model predictions and actual measurements, with differences falling well within measurement uncertainty and experimental repeatability limits.

ANOVA comparing model predictions (Table 3) reveals F-ratio of 0.00 with p-value of 1.0000 for between-groups variation, indicating no statistically significant differences among the three modeling approaches. Levene’s test for variance homogeneity yields p-value of 1.000, confirming equal variances across all groups. Pairwise variance comparisons show F-ratios ranging from 0.98862 to 1.00111 with all p-values exceeding 0.96, further validating variance homogeneity assumptions. Kruskal-Wallis non-parametric test, which does not assume normal distribution, produces test statistic of 0.0651275 with p-value of 0.995665, providing independent confirmation that model predictions are statistically indistinguishable from actual measurements regardless of distributional assumptions. Multiple range tests with 95.0% Least Significant Difference intervals confirm that 2FI Predicted, LMBA Predicted, BRBA Predicted, and Actual values all fall within homogeneous groups (marked as X in Table 3) with no significant contrasts, as all pairwise differences (ranging from − 0.000020704 to 0.000010541) remain well below the critical threshold of 0.00425736. Bonferroni-corrected confidence intervals (± 34.2636) encompass all pairwise contrasts, with differences ranging from − 0.333333 to 3.01587 well within allowable limits. Mood’s Median Test reinforces this conclusion, showing median predictions of 0.0153899 (2FI), 0.0151807 (LMBA), and 0.0153423 (BRBA) statistically indistinguishable from the actual median of 0.0154303, with 95% confidence intervals overlapping extensively across all four groups.

This remarkable concordance across fundamentally different modeling paradigms—parametric polynomial regression versus non-parametric neural networks—provides strong validation that the observed wear behavior genuinely reflects systematic responses to controlled factors rather than artifacts of particular modeling assumptions. The convergence of three independent approaches to statistically indistinguishable predictions enhances confidence in subsequent optimization efforts, as optimal conditions identified by any framework should yield comparable experimental outcomes. From a practical perspective, model selection can be guided by application-specific requirements rather than predictive superiority. The 2FI model offers explicit coefficient values enabling mechanistic interpretation of factor effects and interactions, while neural network models provide superior flexibility for extrapolation without requiring reformulation of functional relationships.

4.4 Multi-Objective optimization and practical validation for automotive application

Numerical optimization was performed using two independent methodologies—Desirability Function Analysis (DFA) and Genetic Algorithm (GA)—to identify factor combinations minimizing wear rate while maintaining feasible processing conditions. Both approaches were constrained to explore the experimental design space with Factor A (RCMP) ranging from Cont7 to Cont11 and Factor B (RCNT) ranging from Grade1 to Grade7, with the objective function specified as maximizing the desirability of minimum wear index between 190 and 6000 mg/min. These constraints ensure that recommended optimal conditions fall within the validated experimental domain where model predictions are reliable, avoiding extrapolation beyond characterized material behavior.

DFA identified optimal conditions at RCNT11 (corresponding to RCMP7) and Grade7 (approximately 11 wt% RCNT), predicting a minimum wear index of 196.59 mg/min with desirability value of 0.979 (Fig. 11a-b). The desirability function, which transforms the objective into a dimensionless scale from 0 (completely undesirable) to 1 (ideal), demonstrates that the identified solution approaches the theoretical optimum with less than 2.1% deviation from perfect desirability [22]. The three-dimensional desirability surface (Fig. 11a) reveals that maximum desirability occurs at the upper boundaries of both factors, with desirability increasing progressively from 0 in the lower-left region (low RCMP, low RCNT) toward 0.98 in the upper-right region (high RCMP, high RCNT). The two-dimensional contour plot (Fig. 11b) illustrates this gradient clearly, showing the optimal region (red zone) concentrated at RCMP7 with 10.5–11 wt% RCNT. The desirability ramps indicate that both factors require maximization to achieve optimal wear resistance, consistent with the experimental observations that ternary reinforcement systems at higher loading densities produce superior tribological performance.

Fig. 11

Optimization results and application validation: (a) DFA 3D desirability surface showing maximum desirability (0.979) at RCMP7 with 11 wt.% reinforcement; (b) DFA 2D contour revealing optimal region (red zone) concentrated at upper boundaries of both factors; (c) GA convergence plots showing fitness stabilization after 400 generations with best fitness of -0.227652; (d) Cylinder liner weight comparison demonstrating optimized Al-Si composite (RCMP6_w11, 461g) achieves 61.3% weight reduction versus cast iron (1192g) while providing superior wear resistance for automotive engine block applications. *(d1-2) fabricated cast iron cylinder sleeve tested (d3-4) AlSi/ZnO/SiC/Gr optimized composite cylinder sleeve tested

Bild vergrößern

Table 4 presents the 20 highest-desirability solutions identified by DFA, arranged in descending order of desirability values. The selected optimal solution (Cont11, Grade7, 196.58 mg/min, D = 0.979) exhibits substantially superior desirability compared to the second-ranked solution (Cont9, Grade7, 2200 mg/min, D = 0.141), demonstrating a clear global optimum rather than multiple local optima of comparable merit. The rapid decline in desirability from 0.979 to 0.141 between ranks 1 and 2 indicates that the optimal region is well-defined and sensitive to compositional variations, with even modest deviations from RCMP7 causing significant performance degradation. Solutions ranked 3–20 show progressively decreasing desirability values ranging from 0.091 to 0.015, corresponding to wear indices from 2966 to 5253 mg/min, confirming that the optimization successfully navigated the complex design space to identify the singular best-performing combination [22].

Table 4

Desirability function Analysis - Top 20 solutions

Rank	RCMP	RCNT	Wear Index (mg/min)	Desirability	Selected
1	Cont11 (RCMP7)	Grade7	196.581	0.979	✓
2	Cont9 (RCMP5)	Grade7	2200.000	0.141
3	Cont11	Grade5	2966.100	0.091
4	Cont11	Grade6	3032.789	0.088
5	Cont7 (RCMP4)	Grade7	3101.712	0.085
6	Cont7	Grade6	3569.478	0.064
7	Cont9	Grade6	3716.555	0.059
8	Cont11	Grade4	3816.558	0.055
9	Cont11	Grade3	3832.902	0.054
10	Cont9	Grade5	4000.000	0.049
11	Cont9	Grade4	4216.568	0.042
12	Cont7	Grade5	4606.652	0.031
13	Cont11	Grade2	4832.990	0.025
14	Cont9	Grade3	4866.322	0.024
15	Cont7	Grade4	4966.329	0.021
16	Cont11	Grade1	5033.004	0.020
17	Cont9	Grade1	5049.263	0.020
18	Cont7	Grade3	5133.010	0.018
19	Cont7	Grade2	5183.253	0.016
20	Cont9	Grade2	5253.047	0.015

Optimal solution

RCMP7 (Cont11) at 11 wt% (Grade7) with predicted wear index of 196.58 mg/min and desirability of 0.979.

Experimental verification

Actual wear index measured at RCMP7_w11 = 197 mg/min (0.2% deviation from prediction).

Genetic Algorithm optimization employed a population size of 70 with constraint-dependent creation and mutation functions, stochastic uniform selection, heuristic crossover (ratio 1.2), and forward migration (fraction 0.2) to evolve candidate solutions over multiple generations. The algorithm converged after approximately 400 generations, as evidenced by fitness scaling reaching steady-state plateaus (Fig. 12). The best fitness achieved was − 0.227652 with mean fitness of −0.221898, where negative values reflect the minimization objective (lower wear index corresponds to better fitness). The current best individual plot shows that Factor A dominated the solution with normalized contribution near 2.0, while Factors 2–4 contributed minimally, confirming that reinforcement composition exerts primary influence on optimal conditions. Expectation plots demonstrate tight convergence, with best, worst, and mean scores collapsing toward identical values beyond generation 100, indicating solution stability. The fitness histogram reveals concentration of individuals near the optimal value, while the selection function plot shows periodic population renewal maintaining genetic diversity [23, 62].

Fig. 12

Genetic algorithm optimization results: Convergence plots showing best fitness ( 0.227652) and mean fitness (-0.221898) over 400 generations, current best individual scores, raw score expectation, selection function, fitness distribution, and individual selection frequency across the population

Bild vergrößern

Comparison of optimization results reveals excellent agreement between DFA (196.59 mg/min) and GA approaches, with both methods converging to the same compositional space of RCMP7 at maximum reinforcement content. The 0.8% difference in predicted wear indices falls well within model prediction uncertainty and experimental variation, validating that the identified optimal region represents a genuine global optimum rather than an artifact of particular optimization methodology. Experimental verification conducted at the recommended optimal conditions (RCMP7_w11) yielded actual wear index of 197 mg/min, demonstrating less than 0.2% deviation from DFA prediction and confirming the practical utility of model-based optimization for reducing experimental screening requirements. The convergence of two fundamentally different optimization paradigms—deterministic desirability maximization versus stochastic evolutionary search—to identical recommendations provides robust confidence that the identified conditions represent the true optimum achievable within the constrained design space, positioning RCMP7_w11 as the superior composition for applications requiring maximum wear resistance.

5 Conclusions

This study systematically addressed the critical knowledge gap in tribological composite design by investigating interaction effects between reinforcement composition and content—parameters conventionally optimized independently despite their potential synergistic influence. A comprehensive three-pronged methodology was employed: Two-Factor Interaction modeling to quantify interaction dominance, Artificial Neural Networks (ANN-LMBA and ANN-BRBA) for non-parametric validation, and dual optimization strategies (Desirability Function Analysis and Genetic Algorithm) to independently identify optimal conditions. This integration of parametric and non-parametric modeling with deterministic and stochastic optimization enables rigorous validation that observed synergistic effects represent genuine material behavior rather than methodological artifacts.

Statistical modeling revealed that reinforcement interaction (AB term, effect magnitude 156.15, F-value = 4006.47, p < 0.0001) dominates wear behavior over individual effects of composition (95.0) and content (94.8), addressing the critical knowledge gap that conventional single-parameter optimization cannot capture synergistic mechanisms governing tribological performance in ternary composite systems.
Two-Factor Interaction modeling achieved exceptional predictive capability (R² = 0.9996, adjusted R² = 0.9994, adequate precision = 336.98, CV = 1.64%), with ANOVA demonstrating model significance (F = 5049.18, p < 0.0001) and diagnostic plots confirming satisfied normality, homoscedasticity, and independence assumptions after inverse square root transformation.
Artificial Neural Networks employing Levenberg-Marquardt (R = 0.99955, best validation MSE = 1.08 × 10⁻⁷ at epoch 35) and Bayesian Regularization (R = 0.99979, best training MSE = 5.76 × 10⁻⁸ at epoch 51) independently corroborated interaction-driven behavior through non-parametric learning, with convergence to parametric 2FI predictions validated by ANOVA (F = 0.00, p = 1.000), Kruskal-Wallis (p = 0.996), Levene’s (p = 1.000), and Mood’s Median tests confirming statistical equivalence across all three modeling approaches.
Dependent optimization using Desirability Function Analysis identified optimal conditions at RCMP7 with 11 wt% reinforcement (predicted wear 196.59 mg/min, desirability = 0.979), while Genetic Algorithm converged to identical composition after 400 generations (predicted wear 194.98 mg/min, best fitness = −0.227652), demonstrating that interaction-informed models enable consistent global optimum identification across deterministic and stochastic methodologies within 0.8% prediction variance.
Experimental validation at optimal conditions (RCMP7_w11) yielded wear index of 197 mg/min, representing 0.2% deviation from DFA and 1.0% from GA predictions, confirming exceptional model accuracy and validating that 33.33% ZnO–16.67% SiC–50% Gr at 11 wt% provides superior synergistic combination compared to alternative reinforcement ratios and contents.
The optimized composite achieved 97.48% wear reduction versus monolithic Al-Si alloy (197 vs. 7810 mg/min), 61.3% weight reduction versus cast iron cylinder liners (461 g vs. 1192 g), fracture toughness of 28.6 MPa·m^1/2 exceeding automotive requirements (> 20 MPa·m^1/2), minimal porosity (1.71%), and consistent worn surface roughness (0.2879 μm), demonstrating that interaction-optimized ternary systems simultaneously achieve tribological excellence, structural integrity, and lightweighting objectives for automotive applications.
Digital image processing quantified porosity evolution (1.71–3.58% across compositions) and surface roughness consistency (0.2879–0.2885 μm for RCMP6 variants), providing auxiliary microstructural validation that optimal compositions maintain superior consolidation quality and uniform material removal characteristics under abrasive wear conditions.
The integration of interaction effect modeling, neural network validation, and dual optimization strategies establishes a robust framework for tribological composite design that reduces experimental screening requirements while ensuring reliable performance predictions, with practical implications for automotive cylinder liner lightweighting enabling potential fuel consumption reduction of 0.4–0.5 L/100 km per 100 kg weight savings and successful utilization of recycled piston scrap validating circular economy principles in high-performance composite manufacturing.

Declarations

Conflicts of interest/Competing interests

The authors have no relevant financial or non-financial interests to disclose.

Open Access This article is licensed under a Creative Commons Attribution 4.0 International License, which permits use, sharing, adaptation, distribution and reproduction in any medium or format, as long as you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons licence, and indicate if changes were made. The images or other third party material in this article are included in the article’s Creative Commons licence, unless indicated otherwise in a credit line to the material. If material is not included in the article’s Creative Commons licence and your intended use is not permitted by statutory regulation or exceeds the permitted use, you will need to obtain permission directly from the copyright holder. To view a copy of this licence, visit http://creativecommons.org/licenses/by/4.0/.

Publisher’s note

Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.

Metadaten
Literatur

Titel: Modeling, validation and multi-objective optimization of Al-Si/SiC-ZnO-Graphite composites: integrating neural networks-genetic algorithm with interaction effect modeling for tribological performance
Verfasst von: Jimmy Lolu Olajide
Isaac Thlabadira
Tarzaan Ezekiel Ayough
Publikationsdatum: 02.03.2026
Verlag: Springer London
Erschienen in: The International Journal of Advanced Manufacturing Technology
Print ISSN: 0268-3768
Elektronische ISSN: 1433-3015
DOI: https://doi.org/10.1007/s00170-026-17577-w

Hernandez FCR, Ramírez JMH, Mackay R (2017) Al-Si alloys: automotive, aeronautical, and aerospace applications. SpringerCrossRef

Khalaf AA (2022) Microstructure evolution of Al-Si hypoeutectic alloys prepared by controlled diffusion solidification. Int J Adv Manuf Technol 120(7):5003–5014CrossRef

Baby AK, Priyaranjan M, Deepak Lawrence K, Rajendrakumar P (2021) Tribological behaviour of hypereutectic Al-Si automotive cylinder liner material under dry sliding wear condition in severe wear regime. Proc Inst Mech Eng Pt J J Eng Tribol 235(7):1450–62CrossRef

Miladinović S, Gajević S, Savić S, Miletić I, Stojanović B, Vencl A (2024) Tribological behaviour of hypereutectic Al-Si composites: a multi-response optimisation approach with ANN and Taguchi Grey method. Lubricants 12(2):61CrossRef

Miladinović S, Stojanović B, Gajević S, Vencl A (2023) Hypereutectic aluminum alloys and composites: a review. Silicon 15(6):2507–2527CrossRef

Singh RK, Telang A (2020) Microstructure, mechanical properties and two-body abrasive wear behaviour of hypereutectic Al—Si—SiC composite. Trans Nonferrous Met Soc China 30(1):65–75CrossRef

Bharathi P, Kumar TS (2023) Mechanical characteristics and wear behaviour of Al/SiC and Al/SiC/B4C hybrid metal matrix composites fabricated through powder metallurgy route. Silicon 15(10):4259–4275CrossRef

Kumar M, Vanmathi M, Sakthivel G (2021) SiC reinforcement in the synthesis and characterization of A356/Al2O3/Sic/Gr reinforced composite-paving a way for the next generation of aircraft applications. Silicon 13(8):2737–2744CrossRef

Kumar Bhoi N, Singh H, Pratap S (2020) Synthesis and characterization of zinc oxide reinforced aluminum metal matrix composite produced by microwave sintering. J Compos Mater 54(24):3625–3636CrossRef

10.

Sadhasivam RS, Ramanathan K, Ravichandran M, Jayaseelan C (2022) Experimental investigations on microstructure, properties and workability behavior of zinc oxide reinforced Al–Si–Mg matrix composites. Silicon 14(5):2175–2187CrossRef

11.

El-Wazery M, Elsad R, Khafagy S, Meiz M (2018) Enhancement of microstructure and mechanical properties of hypereutectic Al–16% Si alloy by ZnO nanocrystallites. Appl Phys A 124(11):736CrossRef

12.

Qasim Z, Jabbar M, Hassan J (2017) Enhancement the mechanical properties of aluminum casting alloys (A356) by adding nanorods structures from zinc oxide. J Mater Sci 6(2):2–5

13.

Liu X, Zhang Z, Le Q, Bao L (2016) The effects of ZnO particles on the grain refinement and mechanical properties of AZ31 magnesium alloys. Trans Indian Inst Met 69(10):1911–1918CrossRef

14.

Kumar M (2021) Mechanical and sliding wear performance of AA356-Al2O3/SiC/graphite alloy composite materials: parametric and ranking optimization using Taguchi DOE and hybrid AHP-GRA method. Silicon 13(8):2461–2477CrossRef

15.

Gowrishankar M, Hiremath P, Shettar M, Sharma S, Rao S (2020) Experimental validity on the casting characteristics of stir cast aluminium composites. J Mater Res Technol 9(3):3340–3347CrossRef

16.

Lakshmikanthan A, Udayagiri SB, Koppad PG, Gupta M, Munishamaiah K, Bontha S (2020) The effect of heat treatment on the mechanical and tribological properties of dual size SiC reinforced A357 matrix composites. J Mater Res Technol 9(3):6434–6452CrossRef

17.

Salonitis K, Jolly M, Pagone E, Papanikolaou M (2019) Life-cycle and energy assessment of automotive component manufacturing: the dilemma between aluminum and cast iron. Energies 12(13):2557CrossRef

18.

Ramachandran M, Agarwal N (eds) (2017) Identification of most affected parameter for design for remanufacturing of scrap piston by Taguchi desirability function analysis. International Conference on Man–Machine Interactions; : Springer

19.

Nunez P, Jones S (2016) Cradle to gate: life cycle impact of primary aluminium production. Int J Life Cycle Assess 21(11):1594–1604CrossRef

20.

Gupta A, Basu B (2019) Sustainable primary aluminium production: technology status and future opportunities. Trans Indian Inst Met 72(8):2135–2150CrossRef

21.

Barati F, Talebi H (2019) Families of search designs for simultaneous detection of the active main and 2-factor interaction effects. J Stat Plann Inference 203:151–159MathSciNetCrossRef

22.

Ganesh M, Arunkumar N, Elavarasan K, Sathish R (2022) Experimental investigation of friction and wear behavior of AL7075-TiO2-B4C using RSM, ANN and desirability function. Int J Veh Struct Syst (IJVSS). ;14(1)

23.

Chiroma H, Noor ASM, Abdulkareem S, Abubakar AI, Hermawan A, Qin H et al (2017) Neural networks optimization through genetic algorithm searches: a review. Appl Math Inf Sci 11(6):1543–1564CrossRef

24.

Elia P, Nativ-Roth E, Zeiri Y, Porat Z (2016) Determination of the average pore-size and total porosity in porous silicon layers by image processing of SEM micrographs. Microporous Mesoporous Mater 225:465–471CrossRef

25.

Kandpal BC, Kumar J, Singh H (2018) Manufacturing and technological challenges in stir casting of metal matrix composites–a review. Mater Today Proc 5(1):5–10CrossRef

26.

Singh AK, Soni S, Rana RS (2020) A critical review on synthesis of aluminum metallic composites through stir casting: challenges and opportunities. Adv Eng Mater 22(10):2000322CrossRef

27.

Sambathkumar M, Navaneethakrishnan P, Ponappa K, Sasikumar K (2017) Mechanical and corrosion behavior of Al7075 (hybrid) metal matrix composites by two step stir casting process. Lat Am J Solids Struct 14:243–255CrossRef

28.

Famodimu OH, Stanford M, Oduoza CF, Zhang L (2018) Effect of process parameters on the density and porosity of laser melted AlSi10Mg/SiC metal matrix composite. Front Mech Eng 13(4):520–527CrossRef

29.

Annigeri UK, Kumar GV (2018) Effect of reinforcement on density, hardness and wear behavior of aluminum metal matrix composites: a review. Mater Today Proc 5(5):11233–7CrossRef

30.

International A (2021) Standard guide for conducting wear tests using a rotary platform abraser. ASTM International, West Conshohocken, PA

31.

Sujith S, Mahapatra MM, Mulik RS (2020) Microstructural characterization and experimental investigations into two body abrasive wear behavior of Al-7079/TiC in-situ metal matrix composites. Proc Inst Mech Eng Pt J J Eng Tribol 234(4):588–607CrossRef

32.

Doddamani S, Kaleemulla M (2015) Review of experimental fracture toughness (KIC) of aluminium alloy and aluminium MMCs. Int J Fract Damage Mech 1(1):1–15

33.

Ahamed KI, Akthar S (2016) A study on neural network architectures. Comput Eng Intell Syst 7:1–7

34.

Abdolrasol MG, Hussain SS, Ustun TS, Sarker MR, Hannan MA, Mohamed R et al (2021) Artificial neural networks based optimization techniques: A review. Electronics 10(21):2689CrossRef

35.

Jeon J, Shin J, Bae D (2019) Si phase modification on the elevated temperature mechanical properties of Al-Si hypereutectic alloys. Mater Sci Eng A 748:367–370CrossRef

36.

Hao J, Yu B, Bian J, Chen B, Wu H, Li W et al (2021) Calculation based on the formation of Mg2Si and its effect on the microstructure and properties of Al–Si alloys. Materials 14(21):6537CrossRef

37.

Barrirero J, Pauly C, Engstler M, Ghanbaja J, Ghafoor N, Li J et al (2019) Eutectic modification by ternary compound cluster formation in Al-Si alloys. Sci Rep 9(1):5506CrossRef

38.

Jahedi M, Ardjmand E, Knezevic M (2017) Microstructure metrics for quantitative assessment of particle size and dispersion: application to metal-matrix composites. Powder Technol 311:226–238CrossRef

39.

Saravana Kumar M, Pruncu C, Harikrishnan P, Rashia Begum S, Vasumathi M (2022) Experimental investigation of in-homogeneity in particle distribution during the processing of metal matrix composites. Silicon 14(2):629–641CrossRef

40.

Alem SAA, Sabzvand MH, Govahi P, Poormehrabi P, Hasanzadeh Azar M, Salehi Siouki S et al (2025) Advancing the next generation of high-performance metal matrix composites through metal particle reinforcement. Adv Compos Hybrid Mater 8(1):3CrossRef

41.

James J, Annamalai AR, Muthuchamy A, Jen C-P (2021) Effect of wettability and uniform distribution of reinforcement particle on mechanical property (tensile) in aluminum metal matrix composite—a review. Nanomaterials 11(9):2230CrossRef

42.

Sowrabh B, Gurumurthy B, Shivaprakash Y (2021) Reinforcements, production techniques and property analysis of AA7075 matrix composites − a critical review. Manuf Rev 8:31

43.

Elsabbagh A (2023) Materials development in transportation industry: fuel consumption, safety and recycling. Ain Shams Eng J 14(6):102096CrossRef

44.

Fentahun MA, Savas MA (2018) Materials used in automotive manufacture and material selection using Ashby charts. Int J Mater Eng 8(3):40–54

45.

Bandil K, Vashisth H, Kumar S, Verma L, Jamwal A, Kumar D et al (2019) Microstructural, mechanical and corrosion behaviour of Al–Si alloy reinforced with SiC metal matrix composite. J Compos Mater 53(28–30):4215–4223CrossRef

46.

Alaneme KK, Fajemisin AV, Maledi NB (2019) Development of aluminium-based composites reinforced with steel and graphite particles: structural, mechanical and wear characterization. J Mater Res Technol 8(1):670–682CrossRef

47.

Hitzler L, Sert E, Schuch E, Öchsner A, Merkel M, Heine B et al (2020) Fracture toughness of L-PBF fabricated aluminium–silicon: a quantitative study on the role of crack growth direction with respect to layering. Prog Addit Manuf 5(3):259–66CrossRef

48.

Sharma P, Sharma S, Kumar Garg R, Paliwal K, Khanduja D, Dabra V (2017) Effect of graphite content on mechanical properties and friction coefficient of reinforced aluminum composites. Powder Metall Met Ceram 56(5):264–272CrossRef

49.

Ye T, Xu Y, Ren J (2019) Effects of SiC particle size on mechanical properties of SiC particle reinforced aluminum metal matrix composite. Mater Sci Eng A 753:146–155CrossRef

50.

Ajagol P, Anjan B, Marigoudar RN, Kumar GP (eds) (2018) Effect of SiC reinforcement on microstructure and mechanical properties of aluminum metal matrix composite. IOP conference series: materials science and engineering; : IOP Publishing

51.

Mao D, Meng X, Xie Y, Yang Y, Xu Y, Qin Z et al (2022) Strength-ductility balance strategy in SiC reinforced aluminum matrix composites via deformation-driven metallurgy. J Alloys Compd 891:162078CrossRef

52.

Jayasathyakawin S, Ravichandran M, Ismail SO, Veerappan G (2023) Effects of ZnO addition on the microstructure/corrosion, wear and mechanical properties of sintered Mg-Al matrix composites. J Alloys Compd 958:170500CrossRef

53.

Singh J (2016) Fabrication characteristics and tribological behavior of Al/SiC/Gr hybrid aluminum matrix composites: a review. Friction 4(3):191–207CrossRef

54.

Kumar J, Singh D, Kalsi NS, Sharma S, Pruncu CI, Pimenov DY et al (2020) Comparative study on the mechanical, tribological, morphological and structural properties of vortex casting processed, Al–SiC–Cr hybrid metal matrix composites for high strength wear-resistant applications: fabrication and characterizations. J Mater Res Technol 9(6):13607–15CrossRef

55.

Valizade N, Farhat Z (2024) A review on abrasive wear of aluminum composites: mechanisms and influencing factors. J Compos Sci 8(4):149CrossRef

56.

Idusuyi N, Olayinka JI (2019) Dry sliding wear characteristics of aluminium metal matrix composites: a brief overview. J Mater Res Technol 8(3):3338–3346CrossRef

57.

Arunagiri K, Radhika N (2016) Studies on adhesive wear characteristics of heat treated aluminium LM25/AlB2 composites. Tribol Ind 38(3):277

58.

Kannan R, Bayraksan G, Luedtke J (2021) Heteroscedasticity-aware residuals-based contextual stochastic optimization. ArXiv Preprint arXiv :210103139

59.

McDonald JB, Nguyen H (2015) Heteroscedasticity and distributional assumptions in the censored regression model. Commun Stat Simul Comput 44(8):2151–2168MathSciNetCrossRef

60.

Das BK, Jha DN, Sahu SK, Yadav AK, Raman RK, Kartikeyan M (2022) Analysis of variance (ANOVA) and design of experiments. Concept Building in Fisheries Data Analysis. Springer, pp 119–36

61.

Antony J (2023) Design of experiments for engineers and scientists. Elsevier

62.

Stojanović B, Gajević S, Kostić N, Miladinović S, Vencl A (2022) Optimization of parameters that affect wear of A356/Al2O3 nanocomposites using RSM, ANN, GA and PSO methods. Ind Lubr Tribol 74(3):350–359CrossRef

Springer Professional

Abstract

Publisher’s note

1 Introduction

2 Materials and methods

2.1 Raw materials

2.2 Experimental design

2.3 Composite fabrication

2.4 Characterization techniques

2.4.1 Raw material characterization

2.4.2 Density and porosity measurements

2.4.3 Wear testing

2.4.4 Microstructural characterization

2.4.5 Fracture toughness testing

2.5 Predictive modeling

2.5.1 Interaction effect modeling

2.5.2 Artificial neural network models

2.6 Optimization strategies

2.6.1 Desirability function analysis

2.6.2 Genetic algorithm optimization

2.7 Digital image processing

2.8 Statistical validation

3 Results and discussion

3.1 Raw material characterization

3.2 SEM-EDX elemental distribution analysis

3.3 Porosity characterization via digital image processing

3.4 Density results

3.5 Mechanical integrity assessment

3.6 Wear performance

3.7 Worn surface morphology and digital roughness analysis

4 Modeling and optimization

4.1 Two-Factor interaction factorial modeling and statistical validation

4.2 Artificial neural network performance evaluation

4.3 Comparative evaluation of predictive models

4.4 Multi-Objective optimization and practical validation for automotive application

5 Conclusions

Declarations

Conflicts of interest/Competing interests

Publisher’s note

Marktübersichten