Next Article in Journal
Effects of Silica Fume, Perlite, and Polypropylene Fibers on the Mechanical Properties of Lightweight Polystyrene Concrete Composite
Next Article in Special Issue
Review of Predictions of Tensile and Flexural Properties of Fiber-Reinforced Composites Using Machine Learning Models
Previous Article in Journal
Performance Improvement of Paper Dye-Sensitized Solar Cell Using H2/Ar-Treated n-Type Semiconducting Carbon-Nanotube Composite Paper
Previous Article in Special Issue
Leveraging Machine Learning (ML) to Enhance the Structural Properties of a Novel Alkali Activated Bio-Composite
 
 
Search for Articles:
Title / Keyword
Author / Affiliation / Email
Journal
Article Type
 
 
Advanced Search
 
Section
Special Issue
Volume
Issue
Number
Page
 
Journals
J. Compos. Sci.
Volume 10
Issue 3
10.3390/jcs10030123
jcs-logo
Submit to this Journal Review for this Journal Propose a Special Issue
► Article Menu

Article Menu

share Share announcement Help format_quote Cite question_answer Discuss in SciProfiles
first_page
settings
Order Article Reprints
Font Type:
Arial Georgia Verdana
Font Size:
Aa Aa Aa
Line Spacing:
Column Width:
Background:
Article

Machine Learning Approach for Mechanical Property Prediction of a Bio-Epoxy and Glass Fiber Composite Reinforced with Titanium Dioxide Nanoparticles

by
Wilson Navas-Pinto
1,2,3,*,
Pablo Díaz-Leime
4,
Germán Omar Barrionuevo
5,
Jhon Luna-Jaén
1,
Xavier Sánchez-Sánchez
1,
Carlos Navas-Cárdenas
1 and
Duncan E. Cree
6,*
1
Departamento de Ciencias de la Energía y Mecánica, Universidad de las Fuerzas Armadas ESPE, Sangolquí P.O. Box 171-5-231B, Sangolquí, Ecuador
2
Department of Mechanical Engineering, University of Saskatchewan, 57 Campus Drive, Saskatoon, SK S7N 5A9, Canada
3
Centro de Nanociencia y Nanotecnología, Universidad de las Fuerzas Armadas ESPE, Av. General Rumiñahui s/n y Ambato, Sangolquí 171103, Ecuador
4
Laboratory of Composite Materials, Research and Development Center, Ecuadorian Air Force, Ambato 180103, Ecuador
5
Department of Mechanical Engineering, Universidad San Francisco de Quito USFQ, Diego de Robles y Vía Interoceánica, Quito P.O. Box 17-0901, Ecuador
6
Department of Mechanical Engineering, McMaster University, 1280 Main Street West, Hamilton, ON L8S 4L8, Canada
*
Authors to whom correspondence should be addressed.
J. Compos. Sci. 2026, 10(3), 123; https://doi.org/10.3390/jcs10030123
Submission received: 15 January 2026 / Revised: 18 February 2026 / Accepted: 21 February 2026 / Published: 25 February 2026
(This article belongs to the Special Issue Machine Learning Applications in the Design and Analysis of Composite Materials)
Browse Figures
Versions Notes

Abstract

Glass fiber reinforced polymers (GFRPs) have drawn significant attention given their lightweight, mechanical resistance and tunable properties through constituent selection. Due to environmental concerns, research efforts have focused on incorporating sustainable materials, such as bio-epoxy resins, to reduce the ecological impact of GFRPs. This study characterizes a GFRP containing a bio-epoxy resin matrix, various loadings of titanium dioxide (TiO2) nanoparticles, and a stabilized arrangement of glass fiber. The unreinforced composite exhibited a tensile strength and modulus of 214 MPa, and 13 GPa, respectively, and a flexural strength and modulus of 375 MPa and 14.5 GPa, respectively. The addition of TiO2 produced an improvement in mechanical response for all the composites. The formulation with 1 wt.% TiO2 showed the best tensile response with an improvement of 13% and 14% for its tensile strength, and modulus, respectively; meanwhile, the composites with 2 wt.% TiO2 attained an improvement of 19% and 40% for the flexural strength and modulus, respectively. Scanning electron microscopy (SEM) revealed significant changes in the fracture mechanism of the composites, while energy-dispersive spectroscopy (EDS) confirmed an even nanoparticle distribution. Additionally, machine learning (ML) models were developed to predict the mechanical response as a function of the TiO2 content.

1. Introduction

In recent decades, polymer matrix composites have progressively replaced traditional materials in a wide range of applications due to the possibility of tailoring their mechanical resistance, thermal conductivity, wear resistance, electrical properties, among other properties through a careful selection of the composite constituents [1,2,3]. For instance, the addition of aluminum nitride particles to a polymer matrix has shown to significantly increase the composite thermal conductivity when compared to the neat polymer [4,5]. Similarly, the addition of graphene oxide to epoxy resins has demonstrated a substantial improvement in the composite mechanical resistance [2,6]. In the same manner, the incorporation of glass fiber to different polymers has proven an outstanding efficiency in enhancing mechanical properties, corrosion resistance and thermal stability of unreinforced polymers [7,8,9].
Thermoset polymers are widely used in the production of composite materials due to their superior mechanical resistance, durability, chemical resistance, and high glass transition temperature when compared to other polymer families [10,11,12]. Among thermoset polymers, epoxy resins have attracted considerable attention because of their excellent adhesion, good mechanical performance, chemical resistance, and superb electrical insulation properties [13,14]. Despite these advantages, epoxy resins exhibit inherently brittle behavior and relatively low fracture toughness [15,16]. Nevertheless, their processing versatility allows them to be used in the fabrication of particle or fiber reinforced composites with enhanced properties [17,18,19]. In recent years, increasing sustainability concerns have driven the development of bio-based epoxy (i.e., bio-epoxy) resins that are synthesized with reduced petroleum-derived content. Consequently, more environmentally friendly alternatives to conventional epoxy resins have been formulated by partially substituting petroleum-based components with materials originating from renewable sources [20,21,22,23].
Glass fiber reinforced polymers (GFRPs) are manufactured using different techniques by embedding glass fiber with different orientations and arrangements into thermoset polymer matrices, typically epoxy resins, to enhance mechanical strength, wear and corrosion resistance [24,25,26]. In this context, GFRPs can achieve strength to weight ratios comparable to those of steel or aluminum [7,26,27]. Despite these advantages, GFRPs remain vulnerable to matrix-dominated failure mechanisms, such as interlaminar shear and microcracking, which may compromise their long-term structural performance [28,29]. To address these limitations, the incorporation of particle reinforcements has emerged as an effective approach to further increase the mechanical performance and durability of GFRPs by acting as barriers for microcrack propagation and hindering interlaminar shear [30,31,32,33]. For example, Alsaadi et al. [30] investigated the addition of silicon carbide (SiC) particles to a 16-layer glass fiber reinforced epoxy resin laminate fabricated via hot compression molding. The inclusion of 5 wt.% SiC produced an average improvement of approximately 30% in shear strength compared to the unfilled composite. Similarly, Gao et al. [31] incorporated aminosilane-functionalized graphene oxide particles into epoxy-based GFRPs using an ultrasound-assisted impregnation method. The addition of 0.3 g of the reinforcement by square meter of the E-glass fiber improved tensile, flexural, and interlaminar shear strengths by 19%, 22%, and 28, respectively, compared to the unfilled material. Likewise, Megahed et al. [32] developed woven epoxy/E-glass composites reinforced with titanium dioxide (TiO2) nanoparticles through vacuum infusion. The study concluded that the addition of 0.5 wt.% TiO2 increased tensile strength, tensile modulus, flexural strength, and flexural modulus by 9%, 50%, 4%, and 49%, respectively.
Among the various particle reinforcements employed in GFRPs, titanium dioxide (TiO2) has emerged as a promising filler candidate due to its chemical stability [34], cost-effectiveness [35], potential for uniform dispersion [34,36], and high surface area [34,36]. Additionally, TiO2 nanoparticles have been reported to enhance mechanical performance of composites without affecting their glass transition temperature [32,37], while also improving wear and corrosion resistance [32,38,39]. Previous studies have reported that including appropriate loadings of TiO2 can increase interlaminar shear strength of laminated composites by toughening the interfacial regions, thereby delaying crack propagation and suppressing premature delamination [32,40,41,42].
In recent years, machine learning (ML) algorithms have been used to predict the properties of composite materials [2,43,44,45]. For instance, Navas-Pinto et al. [2] employed Gaussian process regression to predict the tensile strength (TS) of bio-epoxy/graphene oxide (GO) composites and identified an optimal filler loading between 0.22 and 0.25 wt.% of GO. Similarly, Barrionuevo and Ramos-Grez [43] applied ML algorithms to validate the mechanical properties of 3D-printed PLA specimens with the addition of 30 wt.% of beech wood fibers, achieving a root mean square error of 3.09 which validated the applicability of the approach. In the same way, Liu et al. [44] applied Gaussian support vector machine (G-SVM) models to estimate the tensile modulus and strength of graphene/aluminum nanocomposites obtaining an R-square value of approximately 0.945. Correspondingly, Kim and Oh [45] utilized artificial neural networks (ANNs) to predict the UTS of basalt fiber-reinforced polymers (BFRPs) and GFRPs, reporting string agreement between experimental and predicted values.
Despite the growing number of studies on TiO2 nanoparticles reinforcement in glass fiber reinforced epoxy resins, studies addressing the influence of varying nanoparticle loadings in bio-epoxy-based GFRPs remain limited. Therefore, the present study investigates the incorporation of TiO2 nanoparticles into a bio-epoxy/glass fiber composite fabricated via vacuum-assisted hand lay-up process, with the objective of enhancing its overall mechanical performance. Furthermore, machine learning models were developed to predict the mechanical properties of the composite which were corroborated through an experimental characterization process.

2. Materials and Methods

2.1. Raw Materials

Super SAP CPM/CPL, a 33% bio-based epoxy resin was acquired from Entropy Resins Inc. (San Antonio, TX, USA). TiO2 nanoparticles, with an average particle size between 15 and 25 nm, were purchased from Jiangsu XFNANO (Nanjing, China). BGF Aerialite 1522, a 4-oz E-glass cloth (glass fiber) with a plain weave, was procured from BGF Industries (Danville, VA, USA).

2.2. Composite Fabrication

The glass fiber reinforced bio-epoxy (GFRBE) resin with TiO2 nanoparticles was manufactured through a vacuum assisted hand layup method using a stabilized arrangement of the fibers, as shown in Figure 1.
Prior to the composite fabrication process, as illustrated in Figure 2, different loadings of the particle reinforcement were added to the bio-epoxy resin at the proportions shown in Table 1, which were selected after performing a literature review [32,46,47,48]. The respective amount of the filler was added to the CPM bio-epoxy and stirred at a constant speed of 750 RPM for a duration of 25 min to achieve a uniform particle dispersion [30,49,50]. Then, the CPL hardener was added to the mixture and mixed thoroughly for 10 min, followed by vacuum degassing [2].
The E-glass cloth was cut into 300 mm × 400 mm rectangular sheets considering the different orientations used for the stabilized arrangement [51]. Subsequently, a sealant tape was placed in a rectangular shape onto an aluminum plate. Then, the area inside of the sealant tape was covered with a release agent to allow the laminate removal after the curing process. Afterwards, every layer, as outlined in Figure 1, was impregnated with the different reinforced bio-epoxy formulations using a spatula and stacked on top of the plate. Later, a peel ply film was placed over the laminate. Next, a perforated releasing film was placed on top, followed by a breather and bleeder fabric and the vacuum bagging film. Once the vacuum sealing of the bagging system was confirmed, the laminate was subjected to a 0.3 MPa pressure, used to uniformly compress the composite laminate. Finally, the sealed vacuum bag containing the laminate was placed in an oven at a temperature of 82 °C during 40 min to allow the curing process of the bio-epoxy resin matrix [52,53]. A schematic vacuum bagging configuration is shown in Figure 3. Once the composites were adequately cured, the vacuum bagging was removed and the composite was cut according to ASTM standards D3039-17R25 [54] and D7264-21 [55] for tensile and flexural testing, respectively, and further described in Table 2.

2.3. Composite Characterization

Tensile tests were performed on rectangular samples for each of the reinforcement loadings (0.0, 0.5, 1.0, and 2.0 wt.%) according to D3039-17R25 ASTM standard [54] with an Instron 600 LX universal testing machine equipped with a video extensometer (Norwood, MA, USA). For the purpose, a constant head displacement of 2 mm/min and an initial gauge length of 90 mm were used.
Three-point flexural tests were performed on rectangular samples for each of the reinforcement loadings (0.0, 0.5, 1.0, and 2.0 wt.%) according to D7264-21 ASTM standard [54] with an Instron 600 LX universal testing machine (Norwood, MA, USA) using a fixed crosshead displacement of 1 mm/min. The flexural stress and strain were calculated according to the mentioned standard using Equations (1) and (2), respectively:
σ = 3 P L 2 b h 2  
ε = 6 δ h L 2  
where σ is the flexural stress at the outer surface in the load span region (MPa), P is the applied force (N), L is the support span (mm), b is the width of the beam, h is the thickness of the beam, ε is the flexural strain at the outer surface at mid-span (mm/mm), and δ is the mid-span deflection (mm).
SEM was used for the fractured surface morphology examination on both tensile and flexural specimens with a Tescan Mira 3 (Brno, Czech Republic) operated with an acceleration voltage of 5 kV. For that purpose, the fractured specimens were cut at a distance of approximately 5 mm from the fracture and mounted in a specimen holder. Prior the SEM observation, the specimens received a gold coating layer by means of Quorum Q150 R gold evaporator (Laughton, United Kingdom) to improve their conductivity and avoid a charging effect on the samples [23,56].
In addition, to corroborate a homogeneous particle distribution of the reinforcement within the bio-epoxy resin matrix, energy-dispersive spectroscopy (EDS) was used to obtain elemental mapping along the samples [57,58,59]. The SEM microscope was equipped with a BrukerXFlash 6-30 EDS detector (Billerica, MA, USA) with an energy resolution of 123 eV at Mn-Kα radiation and operated at an acceleration voltage of 25 kV. Ten measurements were performed at different positions and average values were reported.
Similarly, Fourier-transform infrared (FT-IR) spectroscopy was used to assess the interaction of the reinforcement within the bio-epoxy resin matrix [57] by means of a PerkinElmer Frontier spectrometer (Waltham, MA, USA). In this sense, the different spectra were individually acquired along the range from 500 to 4000 cm−1 using a resolution of 4 cm−1.

2.4. Machine Learning Prediction

The experimental dataset included 49 tensile strength measurements for a bio-epoxy/glass fiber composite reinforced with varying amounts of TiO2 nanoparticles, along with 45 measurements of flexural strength. The raw data were sourced from a GitHub v3.15.12 repository and initially examined using descriptive statistics and scatter plots to verify a normal distribution, confirm the absence of missing values, and identify a clear nonlinear relationship between TiO2 content and both tensile and flexural strength. In order to perform the machine learning modeling, the TiO2 loading for each of the composites was used as an input variable, while the tensile or flexural strengths were employed as the target output. To avoid any biased assessment in the performance of the developed models, the research data were divided into two groups, training and test, which included 80% and 20% of the data by means of a fixed random seed, respectively. Subsequently, the input feature was standardized via Z-score normalization, with the scaling parameters learned only from the training data to avoid information leakage [60].
A broad model screening phase was performed using LazyPredict 0.2.16, which included the training process for over 40 regression algorithms with the default hyperparameters. This allowed rapid identification of the most suitable model families for the dataset. Several characteristic performance metrics, such as R2, adjusted R2, root mean square error (RMSE), and computation time, were employed to find the robust algorithm for mechanical property prediction of the developed composite. The top performers in the screening phase were then selected for detailed evaluation. For the prediction of the properties obtained from the flexural test, the same procedure was followed.
Four regressors, (1) random forest (2000 trees), (2) XGBoost (1000 estimators), (3) Multi-layer perceptron (4-neuron hidden layer), and (4) Gaussian process regressor using an RBF kernel with constant scaling, were trained. Afterwards, every model was evaluated using mean absolute error (MAE), mean square error (MSE), and R2 for training and test datasets. Parity plots were developed to compare the experimental and predicted values of the tensile and flexural strength in order to assess the accuracy obtained from each model.
A 5-fold cross-validation process was also developed as a supplementary level of validation to compare how the developed models functioned when evaluated with the same data subset. In addition, Leave-One-Out Cross-Validation (LOOCV) was also employed, as it provides a nearly unbiased performance estimate, ideal for small datasets. For this purpose, the predicted values for every model were generated and displayed in overlaid parity plots, which allowed an easy and direct comparison between the different model families and provided a clear assessment of prediction stability and agreement under identical conditions. All data processing, model training, and visualization tasks were performed in Python 3.12.12 using scikit-learn libraries within a Google Colab environment.

3. Results and Discussion

3.1. Tensile Tests

Tensile stress vs. strain curves for the GFRBE composites with different TiO2 loadings are presented in Figure 4. The main tensile properties (tensile strength, modulus of elasticity, and elongation percent at break) are summarized in Table 3, while Figure 5 illustrates the variation in tensile strength and modulus as a function of TiO2 content. For the unreinforced GFRBE, the tensile strength, modulus of elasticity, and elongation at break were 214.43 ± 7.35 MPa, 13.11 ± 0.36 GPa, and 1.99 ± 0.10%, respectively. The incorporation of TiO2 nanoparticles enhanced the tensile performance of the composites. It can be observed that the tensile strength improved by 9.0, 12.8, and 8.6% with the addition of 0.5, 1.0, and 2.0 wt.% of the reinforcement, respectively. Accordingly, tensile modulus improved by 12, 13.6, and 5.3% for the same TiO2 loadings. The more effective improvement at low TiO2 loadings can be attributed to an enhanced stress transfer mechanism caused by a uniform nanoparticle dispersion [2], improved interfacial interaction between the bio-epoxy matrix and glass fibers [41], and interfacial strengthening induced by the nanoparticle addition [61]. However, although the composite including 2.0 wt.% TIO2 still exhibited an improved mechanical resistance compared to the unreinforced GFRBE, this improvement was less pronounced than the observed at lower reinforcement contents. This behavior may be ascribed to a possible nanoparticle agglomeration at higher contents, which can reduce the homogeneity of the bio-epoxy matrix and promote localized stress concentrations [62,63,64]. Previous studies have reported that increasing TiO2 nanoparticle loadings, enhanced the possibility of agglomeration due to greater van del Waals forces which might compromise the mechanical performance of the manufactured composites [46,47,48].

3.2. Flexural Tests

Flexural stress vs. deflection percentage curves for the GFRBE composites with various TiO2 reinforcement loadings are portrayed in Figure 6. Additionally, Table 4 depicts the flexural strength, and flexural modulus for the composites developed in the present study. Moreover, Figure 7 visually presents how both mechanical properties changed with increasing additions of the nanoparticle reinforcement. As a baseline, the flexural properties of the GFRBE were determined to be 374.97 ± 5.40 MPa, and 14.54 ± 0.17 GPa for the flexural strength and flexural modulus, respectively. After the addition of different TiO2 nanoparticle loadings, a distinguishable flexural strength reinforcement effect can be observed for composites containing 0.5, 1.0, and 2.0 wt.% of the nanoparticle reinforcement due to improvements of 11.5, 17.5, and 19.3%, respectively. Correspondingly, the flexural modulus increased by 17.6, 23.3, and 39.6% for the same filler loadings. This composite flexural property behavior produced can be attributed to the high surface area of the TiO2 nanoparticles which produced an increased interface, allowing an effective energy absorption process during the composite deformation [65]. In addition, the strengthening mechanism caused by the addition of the TiO2 reinforcement can be ascribed to a favorable interlocking effect between the glass fiber, the bio-epoxy resin matrix and the reinforcement [65,66]. Similarly, it has been previously observed that the addition of nanoparticle reinforcement prevented slip mechanisms between the fiber and the matrix which delayed crack propagation in the composites [67].

3.3. SEM of Fractured Specimens

Figure 8 depicts SEM micrographs (1300× magnification) of tensile fractured surfaces for the GFRBE at increasing TiO2 loadings. For the unreinforced GFRBE (Figure 8a), the fractured surface appears relatively smooth, indicating a near brittle fracture, which eases fiber debonding [68]. With the incorporation of TiO2 nanoparticle (Figure 8b–d), different fractured zones show an increased surface roughness, which is indicative of crack pinning mechanisms that enhance energy dissipation while improving the mechanical performance of the composite [2,32,47,68]. Different failure features are observed in the SEM micrographs, including fiber breakage, fiber delamination, and fiber pullout, which reveals an adequate load transfer from the bio epoxy resin matrix to the glass fiber [69,70,71]. Similar observations were reported by Abebe et al. [72], who concluded that nanoparticle reinforcement delayed the polymer matrix cracking and reduced premature fiber debonding. However, other studies have observed that higher reinforcement concentrations can lead to nanoparticle clustering or agglomeration which can result in uneven stress distribution resulting in a reduced reinforcing effect of the nanoparticles [69,73]. Figure 9 illustrates the SEM micrographs (1300× magnification) for the flexural fractured specimens with increasing nanoparticle reinforcement loadings. Comparable to the tensile fractured specimens, flexural fractured specimens depict the same modified surface pattern when the nanoparticle reinforcements were added which act as pinning sites to suppress rapid crack propagation when a nanoparticle is reached. In addition, the fractured fibers’ orientations slightly differ from the tensile specimens as the applied load configurations are different, as observed in Figure 9. This effect might be attributed to the compressive and tensile stresses that are applied during a flexural test and the stabilized arrangement used for the composite fabrication [73,74,75].

3.4. Nanoparticle Distribution in the Matrix

Figure 10 depicts the elemental mapping images and EDS plots for the GFRBE composites with various TiO2 nanoparticle loadings, while Table 5 presents the elemental compositions by weight percentage within the polymer matrix. The elemental distribution of the neat bio-epoxy resin, shown in the left column, exhibited a homogeneous distribution of carbon and oxygen, which are the main elemental components of the bio-epoxy resin matrix, and the absence of TiO2 particles. In contrast, the columns corresponding to the composites containing 0.5, 1.0, and 2.0 wt.% of the reinforcement showed a consistent distribution of the TiO2 nanoparticles, corroborated through their respective titanium mapping. These observations suggest high effectiveness in the process used to disperse the reinforcement within the polymer matrix prior to the vacuum-assisted lamination process.

3.5. FT-IR Spectroscopy

Cured bio-epoxy resin spectrum, illustrated in Figure 11.a, exhibits a broad absorption band at circa 3360 cm−1 which is indicative of the presence of hydroxyl group stretching (-OH) [6,76]. In addition, two characteristic peaks of C-H stretching from epoxy groups can be observed between 3000 and 2800 cm−1 [77]. Another absorption band can be distinguished at around 1730 cm−1 which can be attributed to carbonyl and carboxyl group vibration [78]. A group of peaks can be found between 1650 and 1500 cm−1 which are attributed to C=C bond vibration present in aromatic groups in epoxy resin [79]. The absorption band found at around 1230 cm−1 and 935 cm−1 are indicative of the symmetric vibration of epoxy ring and the C-O-R bond of ether [6,79]. A relatively high peak can be observed at around 820 cm−1 which is assigned to the vibration of oxirane rings [80,81]. After the addition of the TiO2 nanoparticle reinforcement, a slight attenuation can be observed in the broad band around 3360 cm−1 due to the presence of O-H vibration from moisture usually found in TiO2 [69,82]. In the same way, the peak found at around 1024 cm−1 is modified due to the OH out-of-plane vibration. Finally, a minor peak can be observed for the FT-IR spectra of the bio-epoxy resin after the addition of the TiO2 reinforcement (Figure 11b–d) at a wavenumber of around 660 cm−1 which is indicative of the vibration of O-Ti-O bond [69,82].

3.6. Machine Learning Predictions

An initial model screening was performed using LazyPredict which revealed clear differences in how well the various regression families captured the nonlinear relationship between TiO2 loading and the ultimate strength measured in tensile and flexural tests. In addition, the initial screening revealed that the models that rely on linear or strongly constrained functional forms failed to successfully follow the material behavior in both tensile and flexural datasets. On the other hand, it was observed that nonlinear approaches, particularly ensemble-based and kernel-based regressors, displayed a consistent ability to explain the obtained data. The analysis of the R2 values applied to the evaluated algorithms revealed that most of the best performing models relayed on tree-based ensembles, gradient boosting techniques, and Gaussian process regressors.
This trend indicates that the material’s mechanical response is dominated by nonlinear effects, which are more effectively captured by flexible, non-parametric learning methods (Figure 12a). The automated comparison also helped identify regressors that were unstable or incompatible with single-feature datasets, such as Orthogonal Matching Pursuit and LightGBM under default configurations, further reinforcing the importance of algorithm selection when modeling low-dimensional materials data.
For the flexural dataset, the initial screening allowed to rank the regression families in terms of their ability to describe the flexural behavior of the developed composites. It could be differentiated that linear and regularized models exhibited a limited predictive performance. Conversely, ensemble approaches, such as ExtraTreesRegressor, BaggingRegressor, RandomForestRegressor, and GradientBoostingRegressor, achieved the highest R2 values and minimized prediction errors. These results suggest that the flexural strengths of the developed composites are commanded through nonlinear reinforcement mechanisms, such as nanoparticle-driven matrix stiffening and changes in interfacial shear transfer, which are more effectively captured by tree-based and other non-parametric models that adaptively partition the feature space. In addition, Poisson-Regressor, DummyRegressor, and certain spline-based methods obtained a poor prediction ability given their limited suitability for a dataset with a single input and a monotonic response (Figure 12b).
Random forest and gradient boosting are often more accurate for predicting mechanical properties because they can naturally capture the nonlinear and interacting mechanisms that control material behavior [83,84]. Properties such as strength or stiffness do not follow simple linear trends, but instead result from the combined effects of microstructure, reinforcement efficiency, defects, and processing variability. The developed process outlined that tree-based models possess a high adaptability to the underlying complexity through avoiding the assumption of predefined equations into different data regions. Moreover, random forest models achieved a strong prediction ability by averaging many independent trees, which reduced noise and overfitting, while gradient boosting improved their accuracy by a gradual correction of the precision error and capturing subtle trends. These advantages allowed such methods to become robust with suitable approaches to predict complex responses observed in composite materials.
Following the initial screening, the four selected regressors—random forest, XGBoost, Multi-layer perceptron (MLP), and Gaussian process regression (GPR)—were trained in detail and evaluated using identical training–testing partitions and consistent error metrics. All models successfully learned the monotonic increase in tensile strength of the GFRBE composites with increasing TiO2 content but differed in their ability to reproduce finer variations within the 200–260 MPa range. The parity plots generated for each regressor illustrated these differences clearly: tree-based ensembles provided robust trend capture, the neural network yielded smoother but less precise predictions, and the Gaussian process model provided tightly clustered predictions along the ideal fit line (Figure 13). Across all regressors, the results confirmed that the one-dimensional feature space still contained sufficient information to generate accurate predictions of mechanical performance when paired with an appropriate nonlinear model.
Table 6 listed the metrics obtained for the training, testing, and cross-validation (CV) procedures. On the one hand, the metrics associated with errors (MAE, MSE) show similar values, which indicates the consistency of the predicted results, while the R2 values show a significant decrease. LOOCV evaluation shows consistent error magnitudes and stable R2 behavior comparable to the reported k-fold validation results, confirming that the learned relationships are not artifacts of data partitioning. The algorithm that best fits the data is GPR.
Tree-based ensembles (RFR and XGBoost) obtained adequate R2 values around 0.71–0.72, while the MLP exhibited a minor accuracy due to the limited size of the dataset. The Gaussian process regressor obtained the best predictive response with an R2 of 0.78 and the lowest error metrics among all models, which confirmed their suitability low-dimensional, nonlinear mechanical property prediction.
The implemented cross-validation process revealed coherence among the three non-linear regressors and a consistent similarity between predicted and experimental tensile strength values. Although each model approach converses, the obtained predictions converged for all the TiO2 reinforcement loadings, which indicated stability of the learned relationships and accuracy to reproduce the modeling pattern. The GPR models obtained the most consistent and smooth prediction profile, while RFR and XGBR provided more discrete responses given their tree structure.
Regarding the results obtained in the flexural test dataset, RFR and XGBR further confirmed the advantages that ensemble methods provide for this type of problem. Both random forest and XGBoost produced accurate predictions that followed the experimental flexural strength values, with a strong alignment along the prediction plots (Figure 14).
Their acceptable performance indicates their ability to predict abrupt and smooth nonlinearities produced by microstructural changes given the increasing TiO2 concentration. In contrast, the shallow MLP, configured with a single hidden layer of four neurons, showed higher dispersion and weaker generalization on the test set, consistent with the limited dataset size and the sensitivity of neural networks to small sample regimes.
The cross-validation analysis (Table 7), in which all trained models were evaluated on the same held-out test set, confirmed both the robustness of the ensemble-based approaches and the stability of the learned structure–property relationship. Overall, the results show that tree-based ensemble regressors delivered the most reliable and physically meaningful predictions of flexural strength in TiO2-modified composites, as they are able to capture the nonlinear mechanical response introduced by nanoparticle reinforcement.
The metrics obtained in different datasets for predicting flexural strength show more accurate results compared to tensile testing. Predicting flexural response is often easier than predicting tensile response in composite materials because flexural behavior is typically more stable, less sensitive to defects, and governed by a more localized set of deformation mechanisms [85,86]. In flexural testing, the stress state is non-uniform and dominated by a combination of tension and compression, with failure often initiating gradually in the most highly stressed region. This tends to smooth out the influence of local flaws, voids, or weak interfaces, leading to more repeatable and less noisy data [85,87].
In contrast, tensile response is highly sensitive to the weakest link in the material, such as micro voids, imperfect bonding, or slight variations in fiber or particle distribution [88]. Therefore, small defects or voids can lead to an early failure which may possibly produce a certain scattering in the experimental dataset which complicates the learning stage of the model development. From a modeling perspective, the lower noise and more monotonic trends typical of flexural tests eases the property prediction, when using algorithms that depend on structure–property relationships. Collectively, the results demonstrate that machine learning regressors can capture both the global trend and local variations in tensile and flexural strength behavior in nanoparticle-reinforced composites, confirming the suitability of data-driven approaches for mechanical property prediction even in modest datasets.

4. Conclusions

In this study, the characterization of a GFRBE composite reinforced with TiO2 nanoparticles was developed. The main conclusions can be summarized as follows:
•
The addition of TiO2 nanoparticles to the bio-epoxy resin allowed to improve the mechanical resistance, both tensile and flexural, for the totality of composite formulations developed. The composite containing 1 wt.% of the reinforcement displayed the best tensile property results with an improvement of 13.6% for tensile strength and 12.8% for tensile modulus when compared to the unreinforced GFRBE. Similarly, the composite with 2 wt.% reinforcement obtained the best flexural response with an improvement of 19.3% and 39.6% for the flexural strength and modulus, respectively. The obtained results confirm that controlled nanoparticle addition can significantly improve load transfer and improve the overall mechanical resistance on GFRPs.
•
SEM micrographs demonstrate that the improvement in mechanical performance of the composites were associated with the modification of the fracture surfaces, which allowed crack pinning mechanisms to develop.
•
EDS mapping determined the manufacturing process used in this study was adequate to obtain a homogenous TiO2 nanoparticle dispersion in composite matrix.
•
Across both mechanical tests—tension and flexure—the ML models consistently showed that the mechanical response of TiO2-reinforced composites is governed by nonlinear structure–property relationships. In both datasets, ensemble regressors such as Random Forest, ExtraTrees, and XGBoost demonstrated the highest predictive stability, confirming that nanoparticle-induced mechanisms (matrix stiffening, enhanced interfacial bonding, and restricted deformation) produced trends that are not adequately captured by linear or low-complexity models.
•
The agreement among the top-performing regressors in both tensile and flexural predictions demonstrates that supervised machine learning can reliably characterize and generalize the mechanical behavior of TiO2-modified composites, even with modest dataset sizes. The cross-validated consistency observed in both studies indicated that the learned relationships are robust rather than model-dependent, providing strong evidence that data-driven methods are suitable for predicting and optimizing multiple mechanical properties in nanoparticle-reinforced polymer composites.
•
Overall, this study demonstrates that the integration of a thorough characterization process of materials accompanied by machine learning prediction models constitutes a powerful tool to provide an optimization for composite materials’ formulations. Undoubtedly, the observed interaction between machine learning prediction and experimental validation can become the foundation for keen design strategies, which may optimize conventional trial and error experimentation.

Author Contributions

Conceptualization, W.N.-P., G.O.B. and X.S.-S.; methodology, W.N.-P. and P.D.-L.; software, W.N.-P., P.D.-L., G.O.B., and J.L.-J.; validation, W.N.-P., G.O.B. and D.E.C.; formal analysis, W.N.-P., G.O.B. and C.N.-C.; investigation, W.N.-P., P.D.-L., J.L.-J. and C.N.-C.; resources, P.D.-L., C.N.-C. and D.E.C.; data curation, W.N.-P., G.O.B., J.L.-J. and X.S.-S.; writing—original draft preparation, W.N.-P., G.O.B. and J.L.-J.; writing—review and editing, X.S.-S., C.N.-C. and D.E.C.; visualization, W.N.-P.; supervision, D.E.C.; project administration, W.N.-P., X.S.-S. and D.E.C.; funding acquisition, W.N.-P., X.S.-S. and D.E.C. All authors have read and agreed to the published version of the manuscript.

Funding

Duncan E. Cree reports financial support was provided by Natural Sciences and Engineering Research Council of Canada (NSERC), Discovery Grant (RGPIN-2025-05989). Wilson Navas-Pinto reports support from Universidad de las Fuerzas Armadas ESPE through the project 2025-PIC-IND-05.

Data Availability Statement

The data presented in this study are available on request from the corresponding authors. The data are not publicly available because the raw and processed data required to reproduce these findings is part of an ongoing study. The ML dataset is available at https://github.com/GermanOmar/Composite2 (accessed on 20 February 2026).

Acknowledgments

Duncan E. Cree acknowledges the financial support of the Natural Sciences and Engineering Research Council of Canada (NSERC), Discovery Grant (RGPIN-2025-05989). W. Navas-Pinto would acknowledge the support given by Universidad de las Fuerzas Armadas ESPE through the project 2025-PIC-IND-05.

Conflicts of Interest

The authors declare no conflicts of interest. The funders had no role in the design of the study; in the collection, analyses, or interpretation of data; in the writing of the manuscript; or in the decision to publish the results.

Abbreviations

The following abbreviations are used in this manuscript:
GFRPGlass fiber reinforced polymer
SEMScanning electron microscopy
EDSEnergy-dispersive spectroscopy
MLMachine learning
TSTensile strength
GOGraphene oxide
G-SVMGaussian support vector
ANNArtificial neural networks
BFRPBasalt fiber reinforced polymer
GFRBEGlass fiber reinforced bio-epoxy
ASTMAmerican Society for Testing and Materials 
FT-IRFourier-transform infrared spectroscopy
FSFlexural strength
RMSERoot mean square error
GPRGaussian process regression
RFRRandom forest regression
MSEMean square error
MAEMean absolute error
MLPMulti-layer perceptron

References

  1. Hsissou, R.; Seghiri, R.; Benzekri, Z.; Hilali, M.; Rafik, M.; Elharfi, A. Polymer composite materials: A comprehensive review. Compos. Struct. 2021, 262, 113640. [Google Scholar] [CrossRef]
  2. Navas-Pinto, W.; Cree, D.E.; Wilson, L.D.; Barrionuevo, G.O.; Sánchez-Sánchez, X. Mechanical response of bio-epoxy resin composites reinforced with graphene oxide: Machine learning approach for property prediction. Funct. Compos. Struct. 2025, 7, 035002. [Google Scholar] [CrossRef]
  3. Radzi, F.S.M.; Ali, A.; Anwar, U.K.; Khalina, A.; Mansor, M.R.; Sani, M.S.M. Effect of reinforcement of Alkaline-treated sugar palm/bamboo/kenaf and fibreglass/Kevlar with polyester hybrid biocomposites: Mechanical, morphological, and water absorption properties. J. Mater. Res. Technol. 2023, 24, 4190–4202. [Google Scholar] [CrossRef]
  4. Xu, Y.; Chung, D.D.L.; Mroz, C. Thermally conducting aluminum nitride polymer-matrix composites. Compos. Part A Appl. Sci. Manuf. 2001, 32, 1749–1757. [Google Scholar] [CrossRef]
  5. Zhou, Y.; He, J.; Fu, H.; Zhou, X. Fabrication and characterization of aluminum nitride polymer matrix composites with high thermal conductivity and low dielectric constant for electronic packaging. Mater. Sci. Eng. B 2012, 177, 892–896. [Google Scholar] [CrossRef]
  6. Loeffen, A.; Cree, D.E.; Sabzevari, M.; Wilson, L.D. Effect of Graphene Oxide as a Reinforcement in a Bio-Epoxy Composite. J. Compos. Sci. 2021, 5, 91. [Google Scholar] [CrossRef]
  7. Chen, R.S.; Muhammad, Y.H.; Ahmad, S. Physical, mechanical and environmental stress cracking characteristics of epoxy/glass fiber composites: Effect of matrix/fiber modification and fiber loading. Polym. Test. 2021, 96, 107088. [Google Scholar] [CrossRef]
  8. Morampudi, P.; Namala, K.K.; Gajjela, Y.K.; Barath, M.; Prudhvi, G. Review on glass fiber reinforced polymer composites. Mater. Today Proc. 2021, 43, 314–319. [Google Scholar] [CrossRef]
  9. Yang, G.; Park, M.; Park, S.J. Recent progresses of fabrication and characterization of fibers-reinforced composites: A review. Compos. Commun. 2019, 14, 34–42. [Google Scholar] [CrossRef]
  10. Patel, A.; Kravchenko, O.; Manas-Zloczower, I. Effect of Curing Rate on the Microstructure and Macroscopic Properties of Epoxy Fiberglass Composites. Polymers 2018, 10, 125. [Google Scholar] [CrossRef]
  11. Megahed, M.; Fathy, A.; Morsy, D.; Shehata, F. Mechanical Performance of glass/epoxy composites enhanced by micro- and nanosized aluminum particles. J. Ind. Text. 2021, 51, 68–92. [Google Scholar] [CrossRef]
  12. Sajan, S.; Selvaraj, D.P. A review on polymer matrix composite materials and their applications. Mater. Today Proc. 2021, 47, 5493–5498. [Google Scholar] [CrossRef]
  13. Haeri, Z.; Ramezanzadeh, B.; Ramezanzadeh, M. Recent progress on the metal-organic frameworks decorated graphene oxide (MOFs-GO) nano-building application for epoxy coating mechanical-thermal/flame-retardant and anti-corrosion features improvement. Prog. Org. Coat. 2022, 163, 106645. [Google Scholar] [CrossRef]
  14. Jin, K.; Luo, H.; Wang, Z.; Wang, H.; Tao, J. Composition optimization of a high-performance epoxy resin based on molecular dynamics and machine learning. Mater. Des. 2020, 194, 108932. [Google Scholar] [CrossRef]
  15. Zhou, S.; Zhang, W.; Wang, C.; Zhang, J. Highly improving the mechanical and thermal properties of epoxy resin via blending with polyetherketone cardo. Compos. Commun. 2019, 13, 80–84. [Google Scholar] [CrossRef]
  16. Ke, J.; Li, X.; Liao, H.; Wang, J. Critical transition of epoxy resin from brittleness to toughness by incorporating CO2-sourced cyclic carbonate. J. CO2 Util. 2018, 26, 302–313. [Google Scholar] [CrossRef]
  17. Luo, J.; Zhang, Y.; Wang, Y.; Li, Y.; Zhao, Y. Simultaneous optimization of the thermal conductivity and mechanical properties of epoxy resin composites through PES and AgNP functionalized BNs. Compos. B Eng. 2023, 248, 110373. [Google Scholar] [CrossRef]
  18. Wang, H.; Lu, R.; Li, L. Strong, tough, and thermally conductive nacre-inspired boron nitride nanosheet/epoxy layered nanocomposites. Nano Res. 2024, 17, 820–828. [Google Scholar] [CrossRef]
  19. Azman, M.A.; Asyraf, M.R.M.; Khalina, A.; Petrů, M.; Ruzaidi, M.S.; Sapuan, S.M.; Wan Aman, V.G.; Ishak, M.R. Natural Fiber Reinforced Composite Material for Product Design: A Short Review. Polymers 2021, 13, 1917. [Google Scholar] [CrossRef]
  20. Pappa, C.; Feghali, E.; Vanbroekhoven, K.; Triantafyllidis, K.S. Recent advances in epoxy resins and composites derived from lignin and related bio-oils. Curr. Opin. Green Sustain. Chem. 2022, 38, 100687. [Google Scholar] [CrossRef]
  21. Ebrahimnezhad-Khaljiri, H.; Ghadi, A. Recent advancement in synthesizing bio-epoxy nanocomposites using lignin, plant oils, saccharides, polyphenols, and natural rubbers: A review. Int. J. Biol. Macromol. 2024, 256, 128041. [Google Scholar] [CrossRef] [PubMed]
  22. Hidalgo, P.; Salgado, L.; Ibacache, N.; Hunter, R. Influence of Biochar and Bio-Oil Loading on the Properties of Epoxy Resin Composites. Polymers 2023, 15, 1895. [Google Scholar] [CrossRef] [PubMed]
  23. Navas-Pinto, W.; Pila-Quishpe, M.; Velasquez, S.; Valadez-Gonzalez, A.; Espinoza-Santos, C. Comprehensive characterization of Yanchama (Poulsenia Armata) fiber cloth from the Ecuadorian Amazon region: Towards sustainable reinforcement materials. J. Mater. Res. Technol. 2025, 38, 6053–6062. [Google Scholar] [CrossRef]
  24. Birleanu, C.; Udroiu, R.; Cioaza, M.; Pustan, M.; Paul, B.; Vilau, C. The Effect of Fiber Weight Fraction on Tribological Behavior for Glass Fiber Reinforced Polymer. Polymers 2025, 17, 720. [Google Scholar] [CrossRef]
  25. Matykiewicz, D. Hybrid Epoxy Composites with Both Powder and Fiber Filler: A Review of Mechanical and Thermomechanical Properties. Materials 2020, 13, 1802. [Google Scholar] [CrossRef]
  26. Chinnasamy, V.; Subramani, S.P.; Palaniappan, S.K.; Mylsamy, B.; Aruchamy, K. Characterization on thermal properties of glass fiber and kevlar fiber with modified epoxy hybrid composites. J. Mater. Res. Technol. 2020, 9, 3158–3167. [Google Scholar] [CrossRef]
  27. Colombo, M.; Pinho, S.T.; Greenhalgh, E.S. Interfacial Chemistry Behind Damage Monitoring in Glass Fiber-Reinforced Composites: Attempts and Perspectives. Polym. Compos. 2026, 47, 2925–2954. [Google Scholar] [CrossRef]
  28. Wang, Z.; Zhao, X.L.; Xian, G.; Wu, G.; Raman, R.K.S.; Al-Saadi, S. Durability study on interlaminar shear behaviour of basalt-, glass- and carbon-fibre reinforced polymer (B/G/CFRP) bars in seawater sea sand concrete environment. Constr. Build. Mater. 2017, 156, 985–1004. [Google Scholar] [CrossRef]
  29. Januš, O.; Girgle, F.; Rozsypalová, I.; Kostiha, V.; Štěpánek, P. Influence of test parameters on the interlaminar shear strength of FRP bars. Compos. Struct. 2022, 299, 116061. [Google Scholar] [CrossRef]
  30. Alsaadi, M.; Ugla, A.A.; Erklig, A. A comparative study on the interlaminar shear strength of carbon, glass, and Kevlar fabric/epoxy laminates filled with SiC particles. J. Compos. Mater. 2017, 51, 2835–2844. [Google Scholar] [CrossRef]
  31. Gao, H.; Zhang, J.; Li, Y.; Zhao, Y. Enhanced interfacial adhesion in glass fiber fabric/epoxy composites employing fiber surface treatment with aminosilane-functionalized graphene oxide. Text. Res. J. 2021, 91, 790–801. [Google Scholar] [CrossRef]
  32. Megahed, M.; Youssef, S.M.; Ali-Eldin, S.S.; Agwa, M.A. Upgraded Mechanical Properties of Diluent Nano-filled Glass/Epoxy Composites Fabricated by Vacuum Assisted Resin Infusion. Fibers Polym. 2021, 22, 1063–1081. [Google Scholar] [CrossRef]
  33. Kose, H.; Bayar, I.; Ergün, R.K. Experimental optimization of CuO and MgO hybrid nanoparticle reinforcement ratios to enhance fatigue life of GFRP composites. Polym. Compos. 2024, 45, 11125–11137. [Google Scholar] [CrossRef]
  34. Nawaz, H.; Abbas, N.; Arshad, M.A.; Ahmad, S.; Shakoor, A. Polymer Nanocomposites Based on TiO2 as a Reinforcing Agent: An Overview. Adv. Eng. Mater. 2022, 24, 2200844. [Google Scholar] [CrossRef]
  35. Singh, S.; Kumar, V. Investigation on the titanium dioxide nanoparticles infusion on tensile behavior of optimized stacking pattern of hybrid fiber-reinforced polymer composite. Adv. Compos. Mater. 2025, 34, 901–927. [Google Scholar] [CrossRef]
  36. Wang, N.; Zhang, J.; Li, Y.; Zhao, Y. Preparation and Characterization of Polyimide/TiO2 Nano-composites for Non-lethal Weapon. J. Phys. Conf. Ser. 2021, 1759, 012005. [Google Scholar] [CrossRef]
  37. Liu, Z.; Bobrinetskiy, I.; Morozov, R.; Lisenkov, S. Reinforcement of polymer composite materials by titanium dioxide nanoparticles synthesized in plasma discharge under ultrasonic cavitation. J. Phys. Conf. Ser. 2022, 2231, 012012. [Google Scholar] [CrossRef]
  38. Radhakrishnan, S.; Siju, C.R.; Mahanta, D.; Patil, S.; Madras, G. Conducting polyaniline–nano-TiO2 composites for smart corrosion resistant coatings. Electrochim. Acta 2009, 54, 1249–1254. [Google Scholar] [CrossRef]
  39. Díez-Pascual, A.M.; Díez-Vicente, A.L. Nano-TiO2 reinforced PEEK/PEI blends as biomaterials for load-bearing implant applications. ACS Appl. Mater. Interfaces 2015, 7, 5561–5573. [Google Scholar] [CrossRef]
  40. Prasad, V.; Sekar, K.; Varghese, S.; Joseph, M.A. Evaluation of interlaminar fracture toughness and dynamic mechanical properties of nano TiO2 coated flax fibre epoxy composites. Polym. Test. 2020, 91, 106784. [Google Scholar] [CrossRef]
  41. Perumal, K.P.S.; Boopathi, R.; Saravanan, P.; Selvarajan, L. Effect of zircon and anatase titanium dioxide nanoparticles on glass fibre reinforced epoxy with mechanical and morphological studies. Ceram. Int. 2023, 49, 21667–21677. [Google Scholar] [CrossRef]
  42. Elamvazhudi, B.; Gopalakannan, S. Effect of nanoclay and nanoscale TiO2 on carbon/glass fibrereinforced polymer composites. Mater. Res. Express 2020, 7, 075303. [Google Scholar] [CrossRef]
  43. Barrrionuevo, G.O.; Ramos-Grez, J.A. Machine Learning for Optimizing Technological Properties of Wood Composite Filament-Timberfill Fabricated by Fused Deposition Modeling. Commun. Comput. Inf. Sci. 2019, 1194, 119–132. [Google Scholar] [CrossRef]
  44. Liu, J.; Zhang, Y.; Zhang, Y.; Kitipornchai, S.; Yang, J. Machine learning assisted prediction of mechanical properties of graphene/aluminium nanocomposite based on molecular dynamics simulation. Mater. Des. 2022, 213, 110334. [Google Scholar] [CrossRef]
  45. Kim, Y.; Oh, H. Comparison between Multiple Regression Analysis, Polynomial Regression Analysis, and an Artificial Neural Network for Tensile Strength Prediction of BFRP and GFRP. Materials 2021, 14, 4861. [Google Scholar] [CrossRef] [PubMed]
  46. Hirani, H.; Verma, M. Effect of titanium dioxide (TiO2) nanoparticles on the wear rate of marine journal bearings made of composite materials. Results Eng. 2025, 27, 105949. [Google Scholar] [CrossRef]
  47. Khan, S.B.; li, N.; Chen, S.; Liang, J.; Xiao, C.; Sun, X. Influence of nanoparticle size on the mechanical and tribological characteristics of TiO2 reinforced epoxy composites. J. Mater. Res. Technol. 2023, 26, 6001–6015. [Google Scholar] [CrossRef]
  48. Saffar Shamshirgar, A.; Toghyani, A.; Ghasemi-Ghalebahman, A.; Ahmadikhah, M. Raman spectroscopy study of the role of dispersion and interfaces in the improvement of the mechanical properties of epoxy-TiO2 composites. J. Mater. Res. Technol. 2025, 34, 1744–1758. [Google Scholar] [CrossRef]
  49. Zeng, X.L.; Tan, L.S.; He, S.Y.; Zhang, J. Effect of graphene oxide size on interlaminar shear strength of glass fabric/epoxy composites. Mater. Res. Express 2019, 6, 105306. [Google Scholar] [CrossRef]
  50. Parveen, S.; Pichandi, S.; Goswami, P.; Rana, S. Novel glass fibre reinforced hierarchical composites with improved interfacial, mechanical and dynamic mechanical properties developed using cellulose microcrystals. Mater. Des. 2020, 188, 108448. [Google Scholar] [CrossRef]
  51. Harsha, M.S.; Shivamurthy, B.; Anandkumar, R. Synergistic effects of titanium dioxide and graphene nanofillers on delamination and thrust forces in machining glass fiber reinforced nanocomposites. Sci. Rep. 2025, 15, 1–18. [Google Scholar] [CrossRef] [PubMed]
  52. Varghese, A.S.; Sreekanth, M.S. Investigation on the effect of multiwalled carbon nanotubes in carbon fiber-reinforced epoxy composites manufactured using vacuum infusion process and hand layup process followed by vacuum bagging. Polym. Compos. 2024, 45, 1600–1618. [Google Scholar] [CrossRef]
  53. Amar, T.A.; Kumar, A.; Yadav, D.K. A Comparative Study on the Properties of Carbon Fiber-Reinforced Polymer Composites Developed by Hand Layup and Vacuum Bagging Molding Techniques. J. Inst. Eng. (India) Ser. D 2024, 106, 263–275. [Google Scholar] [CrossRef]
  54. Standard Test Method for Tensile Properties of Polymer Matrix Composite Materials. Available online: https://store.astm.org/d3039_d3039m-08.html (accessed on 25 October 2025).
  55. Standard Test Method for Flexural Properties of Polymer Matrix Composite Materials. Available online: https://store.astm.org/d7264_d7264m-21.html (accessed on 25 October 2025).
  56. Navas-Pinto, W.; Cree, D.E.; Wilson, L.D.; Barrionuevo, G.O.; Sánchez-Sánchez, X.; Calvopiña, H. Characterization of graphene oxide synthesized through a modified Hummers method. Ingenius 2025, 34, 31–42. [Google Scholar] [CrossRef]
  57. Abdelhamied, M.M.; Atta, A.; Alotaibi, B.M.; Al-Harbi, N.; Henaish, A.M.A.; Rabia, M. Synthesis of flexible polymer nanocomposites based on methyl cellulose/copper oxide with desired dielectric properties for electrical applications. Inorg. Chem. Commun. 2023, 157, 111245. [Google Scholar] [CrossRef]
  58. Çağlar, S.; Gaylan, Y. Optimization of the microstructural, mechanical, and radiation shielding properties of Al-30B4C-25 W hybrid composites with Gd2O3 reinforcement. Sci. Rep. 2025, 15, 29867. [Google Scholar] [CrossRef]
  59. Noman, M.T.; Amor, N.; Petru, M.; Mahmood, A.; Kejzlar, P. Photocatalytic Behaviour of Zinc Oxide Nanostructures on Surface Activation of Polymeric Fibres. Polymers 2021, 13, 1227. [Google Scholar] [CrossRef]
  60. La Fé-Perdomo, I.; Loyola, O.; Barrionuevo, G.O.; Ramos-Grez, J.A. Adiabatic efficiency and mechanical properties optimization for the laser-based powder bed fusion process of 316L stainless steel alloy. Mater. Today Commun. 2025, 49, 113616. [Google Scholar] [CrossRef]
  61. Bukichev, Y.S.; Bogdanova, L.M.; Lesnichaya, V.A.; Chukanov, N.V.; Golubeva, N.D.; Dzhardimalieva, G.I. Mechanical and Thermophysical Properties of Epoxy Nanocomposites with Titanium Dioxide Nanoparticles. Appl. Sci. 2023, 13, 4488. [Google Scholar] [CrossRef]
  62. Yang, J.; Dai, J.; Liu, X.; Fu, S.; Zong, E.; Song, P. A lignin-based epoxy/TiO2 hybrid nanoparticle for multifunctional bio-based epoxy with improved mechanical, UV absorption and antibacterial properties. Int. J. Biol. Macromol. 2022, 210, 85–93. [Google Scholar] [CrossRef] [PubMed]
  63. Xia, R.; Xi, J.; Zhang, Z.; He, Y.; Yu, Z. Curing Behavior of UV-Initiated Surface-Modified Nano-TiO2/Epoxy Resin Prepolymers and the Properties of Cured Composites. Polymers 2023, 15, 1756. [Google Scholar] [CrossRef] [PubMed]
  64. Goyat, M.S.; Rana, S.; Halder, S.; Ghosh, P.K. Facile fabrication of epoxy-TiO2 nanocomposites: A critical analysis of TiO2 impact on mechanical properties and toughening mechanisms. Ultrason. Sonochem. 2018, 40, 861–873. [Google Scholar] [CrossRef] [PubMed]
  65. Yadhav, B.R.L.; Govindaraju, H.K.; Kiran, M.D.; Suresha, B. Three-point bending and impact behaviour of carbon/epoxy composites modified with titanium dioxide nanoparticles. Mater. Today Proc. 2021, 43, 1755–1761. [Google Scholar] [CrossRef]
  66. Hatata, S.M.; Salman, H.H. Study of Impact Resistant and Flexural Strength for Epoxy Composite Reinforced by Nano and Micro Materials of Silica and Titanium Oxide. AIP Conf. Proc. 2023, 2845, 070036. [Google Scholar] [CrossRef]
  67. Kamesh, B.; Singh, L.K.; Kassa, M.K.; Arumugam, A.B. Synergetic effect of incorporating graphene, CNT and hybrid nanoparticles on the mechanical properties of glass fiber reinforced epoxy laminated composites. Cogent Eng. 2023, 10, 2232604. [Google Scholar] [CrossRef]
  68. Pinto, D.; Bernardo, L.; Amaro, A.; Lopes, S. Mechanical properties of epoxy nanocomposites using titanium dioxide as reinforcement—A review. Constr. Build. Mater. 2015, 95, 506–524. [Google Scholar] [CrossRef]
  69. Kumar, A.; Saha, A.; Kumar, S. Structural analysis of sol-gel derived TiO2 nanoparticles: A critical impact of TiO2 nanoparticles on thermo-mechanical mechanism of glass fiber polymer composites. J. Polym. Res. 2021, 28, 441. [Google Scholar] [CrossRef]
  70. Perumal, K.P.S.; Selvarajan, L.; Kumar, P.M.; Shriguppikar, S. Enhancing mechanical and morphological properties of glass fiber reinforced epoxy polymer composites through rutile nanoparticle incorporation. Prog. Addit. Manuf. 2024, 10, 831–848. [Google Scholar] [CrossRef]
  71. Dasari, S.; Lohani, S.; Dash, S.S.; Fulmali, A.O.; Prusty, R.K.; Ray, B.C. A novel study of flexural behavior of short glass fibers as secondary reinforcements in GFRP composite. Mater. Today Proc. 2021, 47, 3370–3374. [Google Scholar] [CrossRef]
  72. Abebe, Y.; Palani, S.; Sirahbizu, B. Effect of Al2O3 nanoparticle on mechanical properties of polyester/ glass-wool fiber reinforced polymer composites. Hybrid Adv. 2025, 10, 100472. [Google Scholar] [CrossRef]
  73. Barbaz-Isfahani, R.; Shariati, M.; Jalali Azizpour, M. Synergistic Effects of Incorporating Various Types of Nanoparticles on Tensile, Flexural, and Quasi-static Behaviors of GFRP Composites. Fibers Polym. 2022, 23, 2003–2016. [Google Scholar] [CrossRef]
  74. Turaka, S.; Chintalapudi, R.; Geetha, N.K.; Pappula, B.; Makgato, S. Experimental and numerical analysis of the Microstructure and mechanical properties of unidirectional glass fiber reinforced epoxy composites. Compos. Struct. 2024, 331, 117887. [Google Scholar] [CrossRef]
  75. Ojha, S.; Bisaria, H.; Mohanty, S.; Kanny, K. Fractographic analysis of fiber-reinforced polymer laminate composites for marine applications: A comprehensive review. Polym. Compos. 2024, 45, 6771–6787. [Google Scholar] [CrossRef]
  76. Howlader, M.I.; Islam, M.S.; Haque, M.M. Mechanical and morphological analysis of natural fiber-reinforced epoxy and polyester hybrid composites. Results Eng. 2025, 26, 105010. [Google Scholar] [CrossRef]
  77. Zhang, W.; Yin, L.; Zhao, M.; Tan, Z.; Li, G. Rapid and non-destructive quality verification of epoxy resin product using ATR-FTIR spectroscopy coupled with chemometric methods. Microchem. J. 2021, 168, 106397. [Google Scholar] [CrossRef]
  78. Pannico, M.; Mensitieri, G.; Musto, P. In-situ FTIR spectroscopy of epoxy resin degradation: Kinetics and mechanisms. Front. Chem. 2024, 12, 1476965. [Google Scholar] [CrossRef]
  79. Mamudu, U.; Dufresne, A.; Lim, R.C. Synthesis and characterization of phosphorylated cellulose nanocrystals for enhanced UV stability of epoxy nanocomposite films. Compos. Part C Open Access 2023, 12, 100393. [Google Scholar] [CrossRef]
  80. Nikolic, G.; Zlatković, S.J.; Živković, M.Ž.; Nikolić, A.S.; Kakurinov, V.B.; Živković, D.T. Fast Fourier Transform IR Characterization of Epoxy GY Systems Crosslinked with Aliphatic and Cycloaliphatic EH Polyamine Adducts. Sensors 2010, 10, 684–696. [Google Scholar] [CrossRef]
  81. Gonçalves, F.A.M.M.; Ferreira, P.; Alves, P. Synthesis and characterization of itaconic-based epoxy resin: Chemical and thermal properties of partially biobased epoxy resins. Polymer 2021, 235, 124285. [Google Scholar] [CrossRef]
  82. Girigoswami, A.; Deepika, B.; Pandurangan, A.K.; Girigoswami, K. Preparation of titanium dioxide nanoparticles from Solanum Tuberosum peel extract and its applications. Artif. Cells Nanomed. Biotechnol. 2024, 52, 59–68. [Google Scholar] [CrossRef] [PubMed]
  83. Barrionuevo, G.O.; Walczak, M.; Ramos-Grez, J.; Sánchez-Sánchez, X. Microhardness and wear resistance in materials manufactured by laser powder bed fusion: Machine learning approach for property prediction. CIRP J. Manuf. Sci. Technol. 2023, 43, 106–114. [Google Scholar] [CrossRef]
  84. Barrionuevo, G.O.; Ramos-Grez, J.A.; Walczak, M.; Betancourt, C.A. Comparative evaluation of supervised machine learning algorithms in the prediction of the relative density of 316L stainless steel fabricated by selective laser melting. Int. J. Adv. Manuf. Technol. 2021, 113, 419–433. [Google Scholar] [CrossRef]
  85. Ramezani, M.; Choe, D.E.; Rasheed, A. Prediction of the flexural strength and elastic modulus of cementitious materials reinforced with carbon nanotubes: An approach with artificial intelligence. Eng. Appl. Artif. Intell. 2025, 150, 110544. [Google Scholar] [CrossRef]
  86. Mirindi, D.; Sinkhonde, D.; Bezabih, T.; Mirindi, F.; Oshineye, O.; Mirindi, P. Prediction of flexural and split tensile strength of waste glass-concrete composite using machine learning algorithms. Green Technol. Sustain. 2026, 4, 100275. [Google Scholar] [CrossRef]
  87. Sarmiento-Pupo, K.D.; Escalante-Tovar, J.D.; Altamiranda, J.E.; Abellan-Garcia, J.; Vélez, J.I. Predicting flexural strength in fiber-reinforced UHPC via random forest. Sci. Rep. 2025, 15, 42045. [Google Scholar] [CrossRef]
  88. Fu, Y.; Yao, X. A review on manufacturing defects and their detection of fiber reinforced resin matrix composites. Compos. Part C Open Access 2022, 8, 100276. [Google Scholar] [CrossRef]
Jcs 10 00123 g001
Figure 1. Orientation of the different layers in the composite.
Figure 1. Orientation of the different layers in the composite.
Jcs 10 00123 g001
Jcs 10 00123 g002
Figure 2. Composite fabrication schematic diagram.
Figure 2. Composite fabrication schematic diagram.
Jcs 10 00123 g002
Jcs 10 00123 g003
Figure 3. Vacuum bagging schematic configuration.
Figure 3. Vacuum bagging schematic configuration.
Jcs 10 00123 g003
Jcs 10 00123 g004
Figure 4. Tensile stress vs. elongation percentage for the GFRBE composites containing different TiO2 loadings.
Figure 4. Tensile stress vs. elongation percentage for the GFRBE composites containing different TiO2 loadings.
Jcs 10 00123 g004
Jcs 10 00123 g005
Figure 5. Tensile strength and modulus variation for GFRBE composites.
Figure 5. Tensile strength and modulus variation for GFRBE composites.
Jcs 10 00123 g005
Jcs 10 00123 g006
Figure 6. Flexural stress vs. deflection percentage for the GFRBE composites containing different TiO2 loadings.
Figure 6. Flexural stress vs. deflection percentage for the GFRBE composites containing different TiO2 loadings.
Jcs 10 00123 g006
Jcs 10 00123 g007
Figure 7. Flexural strength and modulus variation for GFRBE composites.
Figure 7. Flexural strength and modulus variation for GFRBE composites.
Jcs 10 00123 g007
Jcs 10 00123 g008
Figure 8. SEM micrographs of tensile fractured specimens of (a) GFRBE, (b) GFRBE including 0.5 wt.% of TiO2, (c) GFRBE including 1.0 wt.% of TiO2, and (d) GFRBE including 2.0 wt.% of TiO2.
Figure 8. SEM micrographs of tensile fractured specimens of (a) GFRBE, (b) GFRBE including 0.5 wt.% of TiO2, (c) GFRBE including 1.0 wt.% of TiO2, and (d) GFRBE including 2.0 wt.% of TiO2.
Jcs 10 00123 g008
Jcs 10 00123 g009
Figure 9. SEM micrographs of flexural fractured specimens of (a) GFRBE, (b) GFRBE including 0.5 wt.% of TiO2, (c) GFRBE including 1.0 wt.% of TiO2, and (d) GFRBE including 2.0 wt.% of TiO2.
Figure 9. SEM micrographs of flexural fractured specimens of (a) GFRBE, (b) GFRBE including 0.5 wt.% of TiO2, (c) GFRBE including 1.0 wt.% of TiO2, and (d) GFRBE including 2.0 wt.% of TiO2.
Jcs 10 00123 g009
Jcs 10 00123 g010
Figure 10. Elemental mapping images and EDS plots for GFRBE at different nanoparticle formulations.
Figure 10. Elemental mapping images and EDS plots for GFRBE at different nanoparticle formulations.
Jcs 10 00123 g010
Jcs 10 00123 g011
Figure 11. FT-IR spectra of (a) GFRBE, (b) GFRBE including 0.5 wt.% of TiO2, (c) GFRBE including 1.0 wt.% of TiO2, and (d) GFRBE including 2.0 wt.% of TiO2.
Figure 11. FT-IR spectra of (a) GFRBE, (b) GFRBE including 0.5 wt.% of TiO2, (c) GFRBE including 1.0 wt.% of TiO2, and (d) GFRBE including 2.0 wt.% of TiO2.
Jcs 10 00123 g011
Jcs 10 00123 g012
Figure 12. Comparison of the performance of different machine learning models using LazyPredict for ultimate strength prediction in (a) tensile tests and (b) flexural tests.
Figure 12. Comparison of the performance of different machine learning models using LazyPredict for ultimate strength prediction in (a) tensile tests and (b) flexural tests.
Jcs 10 00123 g012
Jcs 10 00123 g013
Figure 13. Model performance evaluation for predicting tensile strength of (a) random forest regressor (RFR), (b) Extreme gradient boosting regressor (XGBR), (c) Multi-layer perceptron (MLP), and (d) Gaussian process regressor (GPR).
Figure 13. Model performance evaluation for predicting tensile strength of (a) random forest regressor (RFR), (b) Extreme gradient boosting regressor (XGBR), (c) Multi-layer perceptron (MLP), and (d) Gaussian process regressor (GPR).
Jcs 10 00123 g013
Jcs 10 00123 g014
Figure 14. Model performance evaluation for predicting the ultimate strength in the flexural test of (a) random forest regressor (RFR), (b) Extreme gradient boosting regressor (XGBR), (c) Multi-layer perceptron (MLP), and (d) Gaussian process regressor (GPR).
Figure 14. Model performance evaluation for predicting the ultimate strength in the flexural test of (a) random forest regressor (RFR), (b) Extreme gradient boosting regressor (XGBR), (c) Multi-layer perceptron (MLP), and (d) Gaussian process regressor (GPR).
Jcs 10 00123 g014
Table 1. Weight percentage of each of the components for the reinforced matrix preparation.
Table 1. Weight percentage of each of the components for the reinforced matrix preparation.
FormulationTitanium
Dioxide
(wt.%)		CPM
Bio-Epoxy
(wt.%)		CPL
Hardener
(wt.%)
10.070.4029.60
20.570.0529.45
31.069.7029.30
42.069.0029.00
Table 2. Specimen dimensions for composite characterization.
Table 2. Specimen dimensions for composite characterization.
Characterization
Test		ASTM
Standard		Specimen Dimensions
Tensile testD3039-17R25 [54]Length: 250 mm
Width: 25 mm
Thickness: 1.5 mm
Flexural testD7264-21 [55]Length: 125 mm
Width: 12.5 mm
Thickness: 1.5 mm
Support span to depth ratio: 40 to 1
Table 3. Tensile properties of the GFRBE composites at different reinforcement loadings.
Table 3. Tensile properties of the GFRBE composites at different reinforcement loadings.
Composite
Formulation		Tensile Strength
(MPa)		Percent of
Increment
(%)		Tensile Modulus
(GPa)		Percent of
Increment
(%)		Elongation
Percent at
Break
(%)		Percent of
Increment
(%)
GFRBE214.43 ± 7.35--13.11 ± 0.36--1.99 ± 0.10--
GFRBE + 0.5 wt.% TiO2233.65 ± 2.228.9614.67 ± 0.1911.902.00 ± 0.060.50
GFRBE + 1.0 wt.% TiO2241.80 ± 3.7212.7614.89 ± 0.1313.582.01 ± 0.031.01
GFRBE + 2.0 wt.% TiO2232.93 ± 6.608.6313.80 ± 0.295.262.05 ± 0.073.02
Table 4. Flexural properties of the GFRBE composites at different reinforcement loadings.
Table 4. Flexural properties of the GFRBE composites at different reinforcement loadings.
Composite FormulationFlexural
Modulus
(GPa)		Percent of
Increment
(%)		Flexural
Strength
(MPa)		Percent of
Increment
(%)
GFRBE14.54 ± 0.17--374.97 ± 5.40--
GFRBE + 0.5 wt.% TiO217.10 ± 0.1017.61418.24 ± 4.8011.54
GFRBE + 1.0 wt.% TiO217.93 ± 0.0523.31440.54 ± 2.8017.49
GFRBE + 2.0 wt.% TiO220.30 ± 0.1239.62447.32 ± 3.6919.29
Table 5. Elemental distribution for the different matrix formulations.
Table 5. Elemental distribution for the different matrix formulations.
Composite FormulationC
(wt.%)		O
(wt.%)		Ti
(wt.%)
GFRBE80.4019.60–
GFRBE + 0.5 wt.% TiO279.7019.900.40
GFRBE + 1.0 wt.% TiO278.9020.300.80
GFRBE + 2.0 wt.% TiO280.3018.101.60
Table 6. Comparative evaluation of the performance of the ML algorithms employed to predict the tensile strength for the training, testing, and cross-validation dataset.
Table 6. Comparative evaluation of the performance of the ML algorithms employed to predict the tensile strength for the training, testing, and cross-validation dataset.
Model Training Testing
MAEMSER2MAEMSER2
RFR 2.96322.0870.8474.12232.1630.723
XGBR2.96322.0470.8464.38733.2110.714
MLP4.08229.7010.7935.27437.7810.674
GPR3.03622.0800.8463.45625.8080.777
Cross-validationLOOCV
MAEMSER2MAEMSER2
RFR3.58029.3750.4823.84031.4960.773
XGBR4.02731.7540.4654.11932.4070.766
MLP4.44836.4710.2844.70537.2720.731
GPR3.41427.6350.5043.52629.5190.787
Table 7. Comparative evaluation of the performance of the ML algorithms employed to predict the flexural strength for the training, testing, and cross-validation dataset.
Table 7. Comparative evaluation of the performance of the ML algorithms employed to predict the flexural strength for the training, testing, and cross-validation dataset.
ModelTrainingTesting
MAEMSER2MAEMSER2
RFR 4.91559.0430.9028.863107.4540.880
XGBR4.39357.8220.9049.429109.1100.878
MLP6.53872.7720.8799.235154.5050.827
GPR4.44257.8330.9049.447109.9230.877
Cross-validationLOOCV
MAEMSER2MAEMSER2
RFR7.933127.3210.6257.29695.4410.856
XGBR8.716136.9470.6067.74397.7290.852
MLP8.293118.9570.6847.58898.2510.852
GPR10.033219.5300.5447.816101.3620.847
Disclaimer/Publisher’s Note: The statements, opinions and data contained in all publications are solely those of the individual author(s) and contributor(s) and not of MDPI and/or the editor(s). MDPI and/or the editor(s) disclaim responsibility for any injury to people or property resulting from any ideas, methods, instructions or products referred to in the content.

Share and Cite

MDPI and ACS Style

Navas-Pinto, W.; Díaz-Leime, P.; Barrionuevo, G.O.; Luna-Jaén, J.; Sánchez-Sánchez, X.; Navas-Cárdenas, C.; Cree, D.E. Machine Learning Approach for Mechanical Property Prediction of a Bio-Epoxy and Glass Fiber Composite Reinforced with Titanium Dioxide Nanoparticles. J. Compos. Sci. 2026, 10, 123. https://doi.org/10.3390/jcs10030123

AMA Style

Navas-Pinto W, Díaz-Leime P, Barrionuevo GO, Luna-Jaén J, Sánchez-Sánchez X, Navas-Cárdenas C, Cree DE. Machine Learning Approach for Mechanical Property Prediction of a Bio-Epoxy and Glass Fiber Composite Reinforced with Titanium Dioxide Nanoparticles. Journal of Composites Science. 2026; 10(3):123. https://doi.org/10.3390/jcs10030123

Chicago/Turabian Style

Navas-Pinto, Wilson, Pablo Díaz-Leime, Germán Omar Barrionuevo, Jhon Luna-Jaén, Xavier Sánchez-Sánchez, Carlos Navas-Cárdenas, and Duncan E. Cree. 2026. "Machine Learning Approach for Mechanical Property Prediction of a Bio-Epoxy and Glass Fiber Composite Reinforced with Titanium Dioxide Nanoparticles" Journal of Composites Science 10, no. 3: 123. https://doi.org/10.3390/jcs10030123

APA Style

Navas-Pinto, W., Díaz-Leime, P., Barrionuevo, G. O., Luna-Jaén, J., Sánchez-Sánchez, X., Navas-Cárdenas, C., & Cree, D. E. (2026). Machine Learning Approach for Mechanical Property Prediction of a Bio-Epoxy and Glass Fiber Composite Reinforced with Titanium Dioxide Nanoparticles. Journal of Composites Science, 10(3), 123. https://doi.org/10.3390/jcs10030123

Article Metrics

Back to TopTop