Optimizing the drilling performance of Syagrus romanzoffiana fiber biocomposites: minimizing delamination with RSM and ANN modeling

Abstract

This study evaluates the drilling performance of Syagrus romanzoffiana fiber-reinforced bio-epoxy biocomposites, focusing on reducing delamination for better industrial use. An experimental investigation was conducted on drilling composite laminates made by hand lay-up with 30% fiber content. The research examined how various factors—drill bit type (high-speed steel (HSS) and HSS coated with titanium nitride (HSS-TiN)), drill diameter (d, 5–10 mm), spindle speed (N, 800–1600 rpm), and feed rate (f, 50–150 mm/min)—influenced delamination damage, measured by the delamination factor (F_d) through digital image analysis. Predictive models for F_d were created using Response Surface Methodology (RSM) and an Artificial Neural Network (ANN). Results showed that the ANN model had higher predictive accuracy (R² > 0.966, root mean square error (RMSE) < 0.032) than the quadratic RSM model. The analysis identified f as the most influential factor on delamination, followed by d. Additionally, HSS-TiN tools outperformed standard HSS bits. Optimization using a desirability function produced minimum F_d values of 1.02319 for HSS-TiN and 1.03199 for HSS at an f of 50 mm/min, N of 1419.49 rpm, and d of 10 mm. The RSM model was confirmed to be statistically significant through analysis of variance (p < 0.0001), which also revealed a notable interaction between f and N. These results indicate that the hole quality achievable in this biocomposite matches or surpasses that of carbon fiber-reinforced polymer and glass fiber-reinforced polymer under comparable dry drilling conditions. This evidence supports its potential for use in industries such as automotive, sporting goods, and non-critical aerospace components.

Introduction

Natural fibers (NFs) as sustainable reinforcements in polymer composites

In recent years, biocomposite materials have attracted significant interest as sustainable alternatives to synthetic composites, with natural fiber-reinforced polymers (NFRPs) being a primary focus^1–5. Their popularity is mostly driven by advantages such as environmental friendliness, biodegradability, and robust mechanical properties, including high specific strength^6–8. By using plant-derived fibers—such as flax, jute, yucca, and Washingtonia—these materials meet strict ecological standards and avoid the economic challenges associated with recycling traditional composites^9–13. The low density, affordability, and wide availability of natural fibers (NFs) make them particularly suitable for lightweight, eco-friendly applications^14,15.

Challenges in machining natural fiber-reinforced polymers (NRFPs)

Despite these advantages, challenges like poor adhesion between fiber and matrix and difficulties during machining operations, particularly drilling, emphasize the need for further research on how processing parameters impact the material^16–19. Therefore, this study focuses on optimizing the drilling of NFRPs by examining key factors such as tool shape, spindle speed (N), and feed rate (f) to improve mechanical strength and surface quality^20–22.

Drilling is a vital machining process for applying NFRPs in industries that require high precision, such as automotive, aerospace, and construction, where part assembly and performance depend on tight tolerances and high-quality surfaces^23–27. The natural anisotropy of these biocomposites, caused by the different properties of NFs and polymer matrices, makes drilling especially challenging²⁸. This often results in machining defects like delamination, fiber pull-out, matrix cracking, and increased surface roughness^29–32. The high strength and absorbency of NFs can worsen these issues, which can be severe enough to lead to rejection of up to 60% of machined parts, causing significant economic losses^33–35. Therefore, a detailed analysis of machining parameters, including N, f, tool material, and shape, is crucial to understanding their effects on damage mechanisms. This understanding is essential for optimizing processes to minimize defects and tool wear^36–41.

State of the art in drilling optimization of biocomposites

Conventional parametric studies and response surface methodology (RSM)

Early studies on jute-polyester⁴², neem-banyan-sawdust hybrid epoxy⁴³, and jute polymer composites⁴⁴ consistently utilized Taguchi L27 orthogonal arrays to examine the effects of spindle speed (N), feed rate (f), and drill diameter (d) or drill geometry. Variance Analysis (ANOVA) and Response Surface Methodology (RSM) based desirability functions (DFNs) indicated that delamination is minimized with high N, low f, and smaller d. Raja et al.⁴³ reported minimal peel-up (F_d = 0.87) and push-out (F_d = 0.91) delamination using a 6 mm High-Speed Steel (HSS) tool at 1500 rpm and 10 mm/rev, while thrust force and torque remained below 24 N and 5.2 N·m, respectively. Three-dimensional (3D) response surface plots further emphasized strong interactions between parameters, confirming that excessive f is the main contributor to both thrust force and delamination.

Emergence of artificial intelligence predictive and optimization models

Recent studies have shifted from RSM toward hybrid artificial intelligence methods because of the highly nonlinear behavior of natural fiber composites (NFCs). Adda et al.⁴² and Belaadi et al.⁴⁵ compared RSM with artificial neural networks (ANNs) and ANN/genetic algorithm (ANN/GA) hybrids for jute-polyester systems. In both cases, models based on ANN showed significantly higher correlation coefficients (R² > 0.98 across training, validation, and testing) and lower prediction errors than quadratic RSM models. Belaadi et al.⁴⁵ also showed that fiber length is an important but previously overlooked factor that significantly affects fiber-matrix adhesion and the risk of delamination.

Advanced defect detection with deep learning

Hrechuk et al.⁴⁶ introduced a U-Net convolutional neural network for automated semantic segmentation of drilling-induced delamination and uncut fibers in flax-polyactic acid biocomposites. The model achieved high Intersection over Union scores and accurately measured multiple delamination factor (F_d) and hole circularity metrics, providing a robust, operator-independent quality-control tool for industrial uses.

Practical damage-mitigation strategies

Simple yet highly effective process modifications have also been validated. Váňa et al.⁴⁷ showed that introducing a Sololite backing plate during drilling of flax-epoxy kayak laminates reduced exit delamination by up to 80% (from approximately 3500 μm to around 693 μm after 50 holes). Combining the backing plate with carbide tooling and a low f (0.05 mm/rev) further minimized tool wear and damage, significantly improving the machinability of these otherwise difficult NFCs.

Research gap and objectives of the present study

Syagrus romanzoffiana rachis fiber (SrF) composites exhibit notable mechanical properties, such as high tensile strength and stiffness, which compare favorably with traditional NFs^8,25,27. These qualities suggest they are well-suited for eco-friendly, lightweight applications across various industries. However, a crucial aspect for practical use remains unexamined: their machinability. Specifically, there is a complete lack of systematic research on drilling-induced delamination, a common and critical defect that impacts the quality and cost-effectiveness of composite parts. Without this knowledge, industrial adoption of SrF composites is greatly limited.

Unlike previously reported machining investigations on natural-fiber-reinforced polymers, which have focused on fibers such as flax, jute, sisal, and hemp, the present work specifically examines Syagrus romanzoffiana fibers (SrF). These fibers possess a distinct morphology, stiffness, and lumen structure that differ significantly from more conventional plant fibers, directly influencing stress transfer, fiber-matrix debonding, and the delamination damage mechanisms during drilling. Although SrF is recognized within the broader category of NFRPs for its advantageous specific properties, its relative performance and drilling behavior, including damage mechanisms, optimal processing conditions, and final hole quality, remain poorly understood and absent from the scientific literature. This study addresses these gaps by conducting the first focused investigation into the drilling performance of SrF/bio-epoxy composites using combined experimental and modeling approaches. It thereby extends the current understanding of drilling-induced damage in bio-composites by identifying SrF-specific interactions between feed rate, spindle speed, drill diameter, and TiN coating, while also deriving machining guidelines that are explicitly grounded in the experimentally observed SrF damage behavior and are transferable to other natural-fiber systems.

This work shifts focus to a new, unexplored biocomposite: Syagrus romanzoffiana fiber–reinforced bio-epoxy (Bio-EP-SrF). This fiber’s unique morphology, including its aspect ratio, cellulose content, and interfacial properties with Bio-EP, leads to untested machining behavior. The main contributions of this study are threefold: (i) it presents the first experimental investigation of drilling-induced delamination mechanisms in Bio-EP-SrF, a material not previously studied in machining literature; (ii) it offers the first direct, quantitative evaluation of how a TiN coating affects drill performance for this specific fiber-resin system, revealing interactions that cannot be generalized from other NFRPs; and (iii) it introduces the first dedicated RSM and ANN predictive models trained on empirical data from SrF composites, capturing unique non-linear parameter interactions specific to this material.

Therefore, the contribution goes beyond applying established methodologies; it creates fundamental knowledge, material-specific predictive models, and optimized machining guidelines for a new sustainable composite, thus bridging a crucial gap between this novel material’s development and its practical industrial processing.

Materials and methods

Preparation of biocomposites

This study illustrates the process of extracting SrFs in Fig. 1, showing the steps involved in water retting of Syagrus romanzoffiana and rachis palms to produce the final fiber. The Syagrus romanzoffiana rachises used were collected in Skikda Province, northeastern Algeria (36.9°N, 6.9°E), during routine municipal pruning. The collection did not involve any protected or endangered species and did not require special permission or licensing. All experimental procedures adhered to institutional, national, and international regulations, and the species is not listed as threatened under the IUCN Red List or CITES conventions. Samples from the Botanical Circular Pole of the University of August 20, 1955 - Skikda (Algeria), are also available for public reference and identification confirmation.

Fig. 1.

Extraction process of Syagrus romanzoffiana fibers (SrFs).

For biocomposite preparation, locally sourced SrFs and a Bio-EP system were used to create biocomposite samples, as shown in Fig. 2. This figure provides a schematic of the fabrication process, including mixing and pouring Bio-EP, hardener, and SrFs into a silicone mold, resulting in the biocomposite. Samples were made from a rectangular plate measuring 150 mm x 150 mm x 4 ± 0.2 mm with a 30% fiber weight fraction. The SrFs, cut to short lengths of about 5–10 mm, have a tensile strength (σ) of 671 MPa, a strain at break (ɛ) of 1.84%, a density of 1.23 g/cm³, and a Young’s modulus (E) of 41.5 GPa⁴⁸.

Fig. 2.

Schematic overview of the fabrication of Syagrus romanzoffiana fiber-reinforced bio-epoxy resin (Bio-EP-SrF) biocomposite.

The Bio-EP system consisted of a resin with a density of 1.13 g/cm³ and a hardener with a density of 0.95 g/cm³, having a σ of 54.5 MPa, a flexural strength of 76.5 MPa, a flexural modulus of 2.8 GPa, and an elongation at break of 6.5%. These components were mixed in a silicone mold.

Impregnation of the biocomposites was carried out at 25 °C using a Bio-EP system mixed with its corresponding hardener in a 2:1 weight ratio. The short SrFs (5–10 mm) were gradually added to ensure uniform dispersion and proper wetting before curing.

After curing, the molded plates were kept to polymerize for 24 h at room temperature. To ensure full polymerization, the plates were then air-dried for an additional 20 days.

The resulting biocomposite plates containing 30% by weight of SrFs exhibited the following mechanical properties: a σ of 83.21 MPa, an ɛ of 6.65%, an E of 2.15 GPa, a flexural strength of 47.57 MPa, and a flexural modulus of 2.04 GPa⁴⁹.

Fourier transform infrared (FTIR) spectroscopy

The chemical composition and functional groups of the Bio-EP matrix, SrF reinforcement, and their biocomposites were analyzed using Fourier Transform Infrared (FTIR) spectroscopy. A Bruker Alpha II spectrometer (Bruker Optik GmbH, Germany) equipped with an attenuated total reflectance accessory was used. Spectra were collected at the National Polytechnic School of Constantine, Algeria. Measurements were taken in transmission mode over the wavenumber range 500–4000 cm^− 1, with a spectral resolution of 4 cm^− 1. Each spectrum is an average of 32 scans to ensure a high signal-to-noise ratio. This analysis confirmed the completeness of the epoxy curing reaction and identified interactions within the composite system.

Experimental drilling techniques

Drilling tests were performed using an EMCO CONCEPT MILL 55 milling machine with an f of 2–650 mm/min and a maximum N of 3500 rpm. A wooden backing support was placed beneath the biocomposite plates to minimize bending during drilling and to prevent enlargement of the exit-side damage. The operation was conducted dry—without any cooling lubricant—on 150 × 150 × 4 mm³ samples in a single pass (Fig. 3).

Fig. 3.

(a) Experimental setup with biocomposite part being machined; (b) high-speed steel (HSS) and HSS coated with titanium nitride (HSS-TiN) drills; (c) chip obtained after machining; (d) entrance of the hole; (e) exit of the hole after drilling; and (f) final image obtained with ImageJ software.

To ensure stable processing and maintain the integrity of the laminate during dry drilling, specific control protocols were used. Excess heat generation was managed by operating at controlled cutting speeds (V_c) from 12.6 to 50.3 m/min, depending on d (V_c = πdN/1000), with corresponding spindle speeds (N, 800, 1200, and 1600 rpm) and feed rates (f, 50–150 mm/min), which were selected to minimize frictional heating in NRFPs^25,27. A wooden backing plate was also employed to absorb and disperse localized heat, reducing the risk of thermal damage.

Each HSS and HSS-TiN drill bit was limited to a maximum of six cumulative holes, with most bits used for only five holes. Given the 27-run experimental campaign, this required five new bits of each type. This replacement protocol, conservatively based on the reported HSS tool life of 5–10 holes in more abrasive carbon fiber composites, ensured minimal progressive wear during testing. Because the SrFs used here are less abrasive than synthetic reinforcements, thrust and torque remained steady throughout all runs. No measurable impact from tool dulling was observed, confirming that drilling responses and delamination were assessed under consistent cutting conditions.

The goal of this experimental plan was to develop directly usable predictive models and optimized machining parameters for final hole quality (F_d) using only controllable process inputs routinely available in industrial settings (f, N, d, and tool coating). Thrust force and drilling torque were not recorded in this study, as numerous prior investigations on natural fiber composites have consistently demonstrated a strong correlation between these forces and delamination onset^50,51. The observed damage patterns in the current work—dominating influence of f, synergistic f × N interaction, and significant effect of d—align closely with established thrust-force-driven mechanisms, validating the inferred physical interpretations without additional instrumentation. However, we acknowledge that direct force/torque measurements could offer finer resolution of transient cutting phenomena specific to SrF morphology and would be valuable in complementary future studies to further refine mechanistic understanding⁵².

To maintain hole dimensional accuracy and minimize tool wear, drill bits were replaced after every five to six drilling cycles.

The drill bits used in this study were sourced from Sandvik Coromant. The tools included two types: an uncoated HSS model and a physically vapor deposited (PVD) titanium nitride (TiN)-coated HSS version. Both drills had the same core geometry, featuring a 118° point angle, a 30° helix angle, a two-flute design, a 45 mm flute length, and a total length of 93 mm. The cylindrical shank had a consistent diameter of 10 mm. For the coated model, the PVD TiN layer had a nominal thickness of 2–4 μm and a hardness of approximately 2300 HV. The identical base geometry of both tools allowed for a direct, isolated evaluation of how the TiN coating affected delamination during drilling.

Table 1 summarizes additional variables, including N (800, 1200, and 1600 rpm) and f (50, 100, and 150 mm/min).

Table 1.

Drilling process parameters reported in the literature on natural fiber-reinforced epoxy biocomposites. (HSS: high-speed steel, HSS-TiN: high-speed steel coated with titanium nitride, SrF: Syagrus Romanzoffiana fiber, WC: tungsten carbide, SC: solid carbide, f: feed rate, N: spindle speed, d: drill diameter, Bio-EP: Bio-epoxy).

Matrix	Fiber	Fiber content %	Cutting parameter				Ref.
			Tool material	d (mm)	f (mm/min)	N (rpm)
Bio-EP	SrFs	30	HSS, HSS-TiN	5, 8, 10	50, 100, 150	800, 1200, 1600	This study
Epoxy	Jute-cork	40	HSS, HSS-TiN	5, 7, 10	50, 100, 200	450, 900, 1800	70
Epoxy	treated and untreated Washingtonia filifera fibers	32	Twist drills (HSSTiN)	5, 7, 10	50, 150, 200	500, 1500, 2500	102
Epoxy	Jute fabric and cork	30	HSS-TiN twist drill	5, 7, 10	50, 108, 190	355, 710, 1400	103
Epoxy	Jute and flax fabric	-	HSS, HSS-TiN, WC	4	0,01, 0,015, 0,020 (mm/rev)	2500, 5000, 7500	104
Epoxy	Sisal-jute	30	HSS twist drill	6	0.01, 0.06, 0.12 (mm/rev)	500, 1000, 1500	105
Epoxy	Jute-carbon	-	HSS, SC twist drill	6, 8, 10	0.03, 0.06, 0.12 (mm/rev)	750, 1250, 1750	106
Epoxy	Jute fabric	-	HSS twist drill	6, 8, 10	50, 150, 250	1000, 2000, 3000	107
Epoxy	Jute fabric	43	HSS, CoroDrill 854, N2OC, CoroDrill 856, N2OC	8	0.05, 0.10, 0.15 (mm/rev)	750, 1250, 1750	108

Following the drilling process, the samples were imaged using a Canon CanoScan LiDE 400 high-resolution scanner (2400 × 4800 dpi, 48-bit color depth). The resulting digital images were analyzed with the open-source ImageJ software (version 1.47, National Institutes of Health, USA)^53–55 to quantify the damage around the bore, expressed as the delamination factor (F_d). According to Davim et al.⁵³, F_d = D_max/D. All samples were scanned under identical conditions, and ImageJ was calibrated using a fixed pixel-to-millimeter ratio. A consistent image-processing protocol—applying the same grayscale threshold, edge-detection method, and region-of-interest settings—was used for every image to ensure measurements were consistent and directly comparable. The damaged area identified through this process was interpreted as the delamination zone.

Figure 3 shows the method for locating different damage areas around a drilled hole. The size of this damage was assessed using the delamination factor (F_d).

The exit F_d was determined using Eq. (1):

1

where: D is the original, nominal diameter of the hole; D_max indicates the largest diameter surrounding the delaminated zone (in this work, D_max was measured at the hole exit surface, corresponding to push-out delamination); and F_d is the calculated exit delamination factor.

This study quantified and optimized drilling-induced delamination as the primary quality metric in Bio-EP-SrF biocomposites. The physical mechanisms behind delamination are inferred from statistically significant effects of machining parameters, consistent monotonic damage trends, and established principles in composite machining literature, rather than from direct measurement of thrust force, torque, or cutting temperature^56,57.

Response surface methodology (RSM)

RSM is a computational technique used to build mathematical models, design experiments, evaluate factor effects, and identify optimal parameter sets. During model development, RSM employs statistically planned experiments along with least squares regression^58,59 to establish a strong connection between drilling input variables—such as feed rate (f), spindle speed (N), and drill diameter (d)—and process responses like F_dHSS and F_dHSS−TiN, which represent delamination factors (i.e., F_d), enabling accurate prediction of these outcomes.

To investigate the input parameters in the drilling tests of Bio-EP-SrF biocomposite samples, a quadratic RSM model using a Central Composite Design (CCD) with three independent variables was employed in Design-Expert 13. As shown in Table 2, each variable was tested at three levels: feed rate (f) at 50, 100, and 150 mm/min; spindle speed (N) at 800, 1200, and 1600 rpm; and drill diameter (d) at 5, 8, and 10 mm.

Table 2.

Experimental design.

No.	Factor	Notation	Units	Level
				−1	0	1
1	Spindle speed	N	rpm	800	1200	1600
2	Feed rate	f	mm/min	50	100	150
3	Drill diameter	d	mm	5	8	10

The CCD consisted of 27 experimental runs, including 22 unique factorial and axial points, each performed once, along with 5 replicated center points to estimate pure error and model curvature. The pooled standard deviation (SD), derived from these five center-point replicates, was 0.0127 for F_dHSS and 0.0109 for F_dHSS−TiN. These pooled SD values were used in the ANOVA to measure experimental variability and evaluate model adequacy.

After coding the variables and finalizing the experimental setup, the 27-trial sequence was completely randomized in Design-Expert 13 to prevent systematic bias.

Each of the 27 combinations was tested once (one hole per run and tool type), resulting in a total of 54 drilled holes. The design’s statistical validity was confirmed, as the lack-of-fit was not significant (p > 0.10), and the coefficient of variation (CV) remained very low (5.2–6.3%), indicating high reliability in experiments and measurements despite the absence of additional full replications.

The responses from these tests were recorded to model the relationships between the drilling parameters and the resulting outcomes⁶⁰. A second-order polynomial model was developed for the response surface analysis using Eq. (2)⁶¹:

2

where: Y represents the response variables F_dHSS and F_dHSS−TiN, while B₀ indicates the constant term. The coefficients B_i and B_ii relate to the linear and quadratic terms, respectively, and B_ij reflects the interaction effects between variables⁴². It should be noted that RSM relies on the assumption that the underlying relationship between the process parameters and the response can be adequately approximated by a second-order polynomial. While this approach is powerful for screening and optimization, it may under-represent highly nonlinear or multi-regime behaviors that may exist in anisotropic bio-composites such as SrF-reinforced laminates. This structural limitation may partly explain the reduced predictive capability of RSM compared with ANN when strong nonlinear interactions are present^62,63.

Artificial neural network (ANN) modeling

The multilayer perceptron (MLP), a type of ANN, has three main layers: an input layer with neurons representing the independent variables, at least one hidden layer, and an output layer that provides the network’s predictions. Specifically, the MLP was designed with an input layer of three neurons (representing feed rate (f), spindle speed (N), and drill diameter (d)), at least one hidden layer, and an output neuron for the delamination factor (F_d). An exhaustive architecture search was conducted by systematically adjusting the hidden-layer structure, testing configurations with one hidden layer (4–15 neurons) and two hidden layers⁶⁴. The architectures that best reduced RMSE and increased R² in both training and independent validation/test sets were identified as 3–10-1 for F_d predictions with HSS tools and 3–9-1 for F_d predictions with HSS-TiN tools.

To ensure full reproducibility, all simulations were conducted using a fixed random seed (seed = 42). The data were partitioned as follows: for the HSS model, the dataset was randomly divided into 80% for training, 10% for validation, and 10% for testing, while for the HSS-TiN model, a 70%–15%–15% split was used to ensure enough validation samples, given the lower variance in the coated-tool data. This randomization was performed once using the fixed seed to ensure consistent data partitioning across repeated runs⁶⁵.

Before training, the input variables (f, N, and d) were normalized with min–max scaling to the range [0,1]. Network weights and biases were initialized using the Xavier (Glorot) uniform method. Model training was carried out with the Adam optimizer at a learning rate of 0.001 and a batch size of 8 samples. Each network was trained up to a maximum of 1000 epochs; however, an early-stopping mechanism with a patience of 10 epochs based on validation loss was used to stop training once the validation error ceased to improve. This occurred at epoch 5 for the HSS model and epoch 4 for the HSS-TiN model.

The hidden layer used a hyperbolic tangent activation function, while a linear function was applied to the output layer. This approach, combined with the models’ strong performance on unseen test data (R² > 0.95, RMSE < 0.032), shows that overfitting was effectively prevented and confirms that the models have excellent generalization ability⁶⁴.

Given the limited yet statistically structured nature of the experimental dataset (27 CCD-based trials), the use of nested or repeated k-fold cross-validation was intentionally avoided because such techniques are mainly suitable for large, unstructured datasets and can produce unstable or biased estimates when applied to design-of-experiments-driven experiments. Instead, model robustness was maintained through fixed-seed data partitioning, independent test sets, early stopping, and consistency checks against RSM predictions.

The hidden layers used hyperbolic tangent activation functions, while the output layer employed a linear function.

A nonlinear hyperbolic transfer function, called f_f, was commonly used and described by Eq. (3):

3

where: W_i refers to the adjustable synaptic weights, X_i represents the input signals received by the neurons, and n denotes the total number of neuronal units.

Predictions from the ANN model were generated by conducting simulation trials where both the number of hidden layers and the neurons per layer were systematically varied. The model’s performance was assessed using the R² and the RMSE related to the transfer functions, calculated through Eqs. (4) and (5)⁶⁶, respectively. Configurations with the lowest RMSE and highest R² values were identified as optimal, guiding the selection of input variables that enhance predictive performance.

4

5

where: y_{i, e} is the experimental or actual value, y_{i, p} is the predicted value, and n is the number of data points.

Results and discussion

Fourier transform infrared (FTIR) spectroscopy

FTIR spectroscopy was employed to analyze chemical interactions in the Bio-EP-SrF composites and to confirm the full curing of the epoxy matrix. Figure 4 displays the spectra for all materials.

Fig. 4.

Comparative fourier transform infrared (FTIR) spectroscopy of Syagrus romanzoffiana fiber (SrF), neat Bio-EP, and Bio-EP-SrF.

The SrF spectrum displays characteristic peaks of lignocellulosic material: a broad –OH stretch (3600–3000 cm^{− 1}), C–H stretches (~ 2918, 2850 cm^{− 1}), a weak C = O band (~ 1730 cm^{− 1}), signals from aromatic lignin (~ 1605, 1508 cm^{− 1}), and C–O–C vibrations (~ 1032 cm^{− 1})⁴⁸. The spectrum of the neat Bio-EP matrix shows the expected aliphatic C–H stretches and a prominent broad –OH band from the crosslinked network, confirming successful polymerization⁶⁷.

Critically, the composite spectrum shows no new absorption bands, especially in the 910–920 cm^{− 1} range that indicate unreacted epoxy rings, confirming that the curing process reached completion. The main change after fiber addition is a decrease in the intensity of the broad –OH band. This reduction is due to hydrogen bonding and physical interactions between hydroxyl groups on the fiber surface and the cured epoxy network, which limit the vibrational freedom of these groups⁶⁸. No new peaks are observed in the carbonyl region (1680–1760 cm^{− 1}), indicating that the epoxy network structure remains intact.

The decreasing intensity of the –OH band and lignin-related peaks with higher fiber content indicates enhanced fiber-matrix interfacial adhesion and better dispersion, which further restricts hydrophilic groups. The lack of reaction-specific bands confirms these are purely physical interfacial effects within a fully cured matrix, not changes in curing chemistry.

In summary, the FTIR analysis confirms consistent, complete curing across all composites. The observed spectroscopic changes align with improved physical interaction at the fiber-matrix interface, supporting the reported mechanical enhancements and reduced delamination⁶⁹.

Influence of drilling parameters on the delamination factor (F_d)

All reported F_d values correspond to exit delamination. Feed rate (f) emerged as the primary influencing factor, with Fd consistently increasing at higher f across both tool types due to elevated thrust forces, contributing up to ~ 66% to damage (similar to trends in jute/cork composites⁷⁰. Drill diameter (d) ranked second, with larger diameters yielding 12–18% higher Fd owing to quadratically scaled thrust and torque. Spindle speed (N) exhibited a non-monotonic effect, with an optimal ~ 1200 rpm balancing matrix softening (reducing thrust) against excessive thermal interfacial degradation. HSS-TiN tools consistently produced lower F_d than HSS, attributed to reduced friction and improved heat dissipation, minimizing fiber pull-out⁷¹.

Also, d is a major contributor. Larger drills (10 mm) consistently produced higher F_d values compared to 5 mm under the same conditions.

Statistical analysis identified d as the second most important primary factor influencing delamination, after f (ANOVA: F-value = 18.14, p = 0.0004). Its effect was mostly linear, with no significant quadratic trend or interaction with spindle speed (N × d). On average, increasing d from 5 mm to 10 mm resulted in a 12–18% rise in F_d across all tested conditions. This increase is due to the roughly quadratic relationship between thrust force and torque with diameter, which worsens both peel-up and push-out damage mechanisms at the hole exit. This pattern aligns with established theoretical predictions and is supported by extensive experimental data from studies on both natural-fiber and synthetic composites.

For example, at f = 150 mm/min and N = 1600 rpm, F_d increased from 1.8725 (5 mm) to 1.9506 (10 mm) with HSS. This matches previous reports, in which d accounted for up to 88.19% of delamination⁴⁵. The larger contact area of bigger drills increases thrust force and damage.

Comparing tools reveals a consistent advantage for the HSS-TiN drill, which produced slightly lower F_d values under various conditions. For example, at f = 150 mm/min, N = 1200 rpm, and d = 5 mm, F_d decreased from 1.8303 (HSS) to 1.3588 (HSS-TiN). The coating probably reduces friction and improves heat resistance, helping to decrease fiber pull-out, in line with the results observed with other coated tools⁷⁰.

The comparison in Tables 3, 4; Fig. 5 reveals that machining parameters, especially f, N, d, and tool material, significantly affect F_d in NF-epoxy biocomposites. This study’s lowest F_d of 1.0212, achieved with Bio-EP-SrF and an HSS-TiN tool at 50 mm/min and 800 rpm, falls within the range (1.0004–1.2101) reported in other researches. Coated HSS-TiN tools outperform uncoated HSS, generating a smaller F_d due to superior heat resistance and reduced friction.

Table 3.

Experimental results for the delamination factor (F_d) of drilled holes at the exit. (HSS: high-speed steel, HSS-TiN: high-speed steel coated with titanium nitride). (f: feed rate, N: spindle speed, d: drill diameter).

Experiment number	Randomization order	Input variables			Output variables
		f (mm/min)	N (rpm)	d (mm)	FdHSS	FdHSS−TiN
1	17	50	800	5	1.0641	1.0886
2	13	100	800	5	1.2939	1.1969
3	10	150	800	5	1.5936	1.4212
4	27	50	1200	5	1.0552	1.0641
5	20	100	1200	5	1.3815	1.1577
6	7	150	1200	5	1.8303	1.3588
7	6	50	1600	5	1.1106	1.1315
8	11	100	1600	5	1.5746	1.4351
9	16	150	1600	5	1.8725	1.7735
10	23	50	800	8	1.0714	1.0944
11	12	100	800	8	1.2296	1.2168
12	19	150	800	8	1.2958	1.3624
13	2	50	1200	8	1.0462	1.0713
14	15	100	1200	8	1.3582	1.2538
15	25	150	1200	8	1.5746	1.5143
16	14	50	1600	8	1.0358	1.0554
17	3	100	1600	8	1.2287	1.4740
18	18	150	1600	8	1.8083	1.6818
19	26	50	800	10	1.0476	1.0212
20	5	100	800	10	1.1487	1.2839
21	22	150	800	10	1.3250	1.472
22	8	50	1200	10	1.0386	1.0379
23	9	100	1200	10	1.2371	1.3094
24	24	150	1200	10	1.3614	1.5738
25	1	50	1600	10	1.0306	1.0449
26	4	100	1600	10	1.4210	1.2419
27	21	150	1600	10	1.9506	1.6808

Table 4.

Comparative analysis of the minimum delamination factor (F_d) across various studies on natural fiber-reinforced epoxy biocomposites. (HSS: high-speed steel, HSS-TiN: high-speed steel coated with titanium nitride, SrF: Syagrus Romanzoffiana fiber, WC: tungsten carbide, f: feed rate, N: spindle speed, d: drill diameter, SC: solid carbide, Bio-EP: Bio-epoxy).

Matrix	Fiber	Cutting parameters				Fd	Ref.
		Tool material	d (mm)	f (mm/min)	N (rpm)
Bio-EP	SrFs	HSS-TiN	10	50	800	1.0212	This study
Epoxy	Jute/Cork	HSS-TiN	5	50	1800	1.2101	70
Epoxy	Treated and Untreated WF fibers	HSS-TiN	10	50	1500	1.020	102
Epoxy	Jute fabric and cork	HSS-TiN	5	50	710	1.025	103
Epoxy	Jute and flax fabric	WC	4	0.01 mm/rev	2500	1.010	104
Epoxy	Sisal/jute	HSS	6	0.06 mm/rev	1500	1.0857	105
Epoxy	Jute/carbon	SC	6	0.04 mm/rev	1750	1.160	106
Epoxy	Jute fabric	HSS	6	150	3000	1.007	107
Epoxy	Jute fabric	HSS	8	0.15 mm/rev	1250	1.004	108

Fig. 5.

Bar graph illustrating the comparative delamination factor (F_d).

Response surface methodology (RSM) and analysis of variance (ANOVA) for delamination factor (F_d) data

Quadratic models for the F_d were developed using RSM based on experimental data from the CCD detailed in Table 2 and the results in Tables 3, 4. This method allows for the analysis of the individual and combined effects of drilling parameters (f, N, and d) on F_d for both tool types. This approach is common in recent optimization studies of machining NF composites⁷². Table 5 displays the resulting second-order quadratic models, expressed in actual factors.

Table 5.

Mathematical models for various estimator delamination factors (F_d) using response surface methodology (RSM). (F_d: delamination factor, HSS: high-speed steel, HSS-TiN: high-speed steel coated with titanium nitride).

RSM response
F dHSS
F dHSS−TiN

The statistical significance of the developed models and their components was evaluated using ANOVA, with results presented in Table 6. The ANOVA for the F_dHSS model confirms its high significance, indicated by an F-value of 46.69 and a p-value less than 0.0001. This indicates only a 0.01% chance that such a large F-value is due to random noise, a common benchmark in modern machining model validation⁷³. Among the linear factors, feed rate (f) was the most influential (F-value = 212.42, p-value < 0.0001), followed by spindle speed (N) (F-value = 25.62, p-value < 0.0001), and drill diameter (d) (F-value = 18.14, p-value = 0.0004). Significant interaction effects were observed for f × N (F-value = 20.65, p-value = 0.0002) and f × d (F-value = 7.83, p-value = 0.0111), while the N × d interaction was not significant.

Table 6.

Analysis of variance (ANOVA) for the quadratic response surface model of various delamination variables. (SD: standard Deviation, R²: correlation coefficient, CV: coefficient of Variation, DF: degree of freedom, F_d: delamination factor, HSS: High-speed steel, MS: mean square, HSS-TiN: High-speed steel coated with titanium nitride, f: feed rate, N: spindle speed, d: drill diameter, SS: sum of squares, AP: adequate precision).

Source	DF	SS	MS	F-value	p-value	Partial η2	Effect size	Remark
a) ANOVA for Fd of HSS
Model	9	1.98	0.3297	46.69	< 0.0001	0.943	Large	Significant
f	1	1.50	1.50	212.42	< 0.0001	0.928	Large	Significant
N	1	0.1809	0.1809	25.62	< 0.0001	0.561	Large	Significant
d	1	0.1281	0.1281	18.14	0.0004	0.476	Large	Significant
f x N	1	0.1458	0.1458	20.65	0.0002	0.508	Large	Significant
f x d	1	0.0553	0.0553	7.83	0.0111	0.281	Moderate	Significant
N x d	1	0.0013	0.0013	0.1796	0.6762	0.009	Small
f x f	1	0.0483	0.0483	6.84	0.0181	0.255	Moderate	Significant
N x N	1	0.0553	0.0553	7.83	0.0123	0.281	Moderate	Significant
d x d	1	0.0013	0.0013	0.184	0.673	0.009	Small
Residual	17	0.1412	0.0071
Cor Total	26	2.12
SD = 0.0840 Mean = 1.33 CV = 6.31%				R2 = 0.9434 Adjusted R2 = 0.9243 Predicted R2 = 0.8762 AP = 22.1561
b) ANOVA for Fd of HSS-TiN
Model	9	1.18	0.1968	42.76	< 0.0001	0.928	Large	Significant
f	1	0.9708	0.9708	210.91	< 0.0001	0.913	Large	Significant
N	1	0.1084	0.1084	23.56	< 0.0001	0.541	Large	Significant
d	1	0.0015	0.0015	0.0305	0.8631	0.016	Small
f x N	1	0.0609	0.0609	13.23	0.0016	0.398	Moderate	Significant
f x d	1	0.0093	0.0093	2.03	0.1698	0.092	Small
N x d	1	0.0151	0.0151	3.27	0.0855	0.141	Moderate
f x f	1	0.5456	0.5456	118.61	< 0.0001	0.856	Large	Significant
N x N	1	0.0609	0.0609	13.21	0.0020	0.398	Moderate	Significant
d x d	1	0.0001	0.0001	0.02	0.8846	0.001	Negligible
Residual	17	0.0921	0.0046
Cor Total	26	1.27
SD = 0.0678 Mean = 1.30 CV = 5.23%				R2 = 0.9277 Adjusted R2 = 0.9060 Predicted R2 = 0.8516 AP = 19.4914

The interaction between f and spindle speed (f × N) exhibited the most significant statistical significance among the two-factor interactions (p = 0.0002). This aligns with well-known principles in drilling fiber-reinforced polymers. When f is low (50 mm/min), variations in N have little effect on F_d because the thrust forces stay minor and thermal effects are not significant. In contrast, at a high f (150 mm/min), increasing N greatly worsens delamination. This is due to the combined influence of a larger uncut chip thickness (f/N) and higher cutting temperature, which together greatly weaken the matrix’s shear strength near the hole exit. This makes peel-up and push-out damage more extensive. The harmful combined effect of high f and N is well-documented not only in NFCs^42,45,70, but also in carbon fiber-reinforced polymer (CFRP) and glass fiber-reinforced polymer (GFRP), emphasizing its key role in the drilling process.

The significance of quadratic terms like f² and N² shows a nonlinear relationship among the parameters, which matches findings in studies on drilling biocomposites where such complex interactions affect damage⁷⁴. The model’s validity is also supported by a non-significant lack of fit (p-value > 0.05).

Their fit statistics confirm the strength of the RSM models. For the F_dHSS model, an R² of 0.9434 shows that the model accounts for 94.34% of the variation in the response. The adjusted R² (0.9243) and the predicted R² (0.8762) are well aligned, indicating strong predictive ability. An adequate precision (AP) ratio of 22.1561, well above the threshold of 4, signifies a solid signal-to-noise ratio and excellent model performance for exploring the design space—which is crucial for industrial applications⁷⁵. A low CV (6.31%) demonstrates high reliability. The F_dHSS−TiN model exhibits similar performance, with R² = 0.9377, adjusted R² = 0.9060, predicted R² = 0.8516, AP = 19.4914, and a CV = 5.23%, meeting standard criteria for RSM model validity.

As shown in Fig. 6, analysis of the parameter interactions reveals that F_d in these biocomposites is strongly affected by the interaction between f and N. Specifically, the positive relationship between f and F_d is much stronger at higher N (Figs. 6a, d), confirming that these parameters do not act independently but together negatively impact hole quality. Additionally, the influence of d is significant, as larger diameters cause a steeper increase in F_d across different N levels compared to smaller diameters (Figs. 6b, e). This trend is consistent across various tool types, although the HSS-TiN tool generally produces lower F_d values. Similarly, the interaction between f and d shows that F_d increases more noticeably at higher f when using larger diameter tools (Figs. 6c, f), highlighting that specific parameter combinations are more important than individual ones in controlling drill-induced damage.

Fig. 6.

Interaction plots of the delamination factor (F_d) of Syagrus romanzoffiana fiber-reinforced bio-epoxy resin (Bio-EP-SrF) biocomposites.

The influence of drilling parameters on F_d is visually shown with scatter plots (Fig. 7) and box plots (Fig. 8). For F_dHSS (Figs. 7a-c), scatter plots reveal a strong positive relationship between f and F_d, which increases significantly as f rises from 50 to 150 mm/min. N displays a more complex, non-linear relationship, while d shows a direct positive correlation with F_d. Similar trends are seen for F_dHSS−TiN (Figs. 7d-f), although the lower F_d values suggest that the coating helps reduce delamination. The box plots for F_dHSS (Figs. 8a-c) support these findings, showing higher medians and wider data ranges at higher levels of f and d, indicating increased variability and damage⁷⁰. In contrast, the box plots for F_dHSS−TiN (Figs. 8d-f) are narrower, indicating more consistent results and less spread, especially at lower N values.

Fig. 7.

Scatter plots showing the effect of feed rate (f), spindle speed (N), and drill diameter (d) on the delamination factor (F_d) of Syagrus romanzoffiana fiber-reinforced bio-epoxy resin (Bio-EP-SrF) biocomposites: (a-c) F_dHSS and (d-f) F_dHSS−TiN. (HSS: high-speed steel, HSS-TiN: high-speed steel coated with titanium nitride).

Fig. 8.

Box plots showing the effect of feed rate (f), spindle speed (N), and drill diameter (d) on the delamination factor (F_d) of Syagrus romanzoffiana fiber-reinforced bio-epoxy resin (Bio-EP-SrF) biocomposites: (a–c) F_dHSS and (d–f) F_dHSS−TiN. (HSS: high-speed steel, HSS-TiN: high-speed steel coated with titanium nitride).

As shown in Fig. 9, the diagnostic plots confirm the core assumptions of the RSM models. The normal probability plots of the residuals (Figs. 9a for HSS and 9b for HSS-TiN) closely follow a straight line, validating a normal distribution. The strong alignment of data points along the 45-degree line in the predicted vs. actual value plots (Fig. 9c and d) indicates a high correlation between the experimental results and the model’s predictions. Additionally, the random, pattern-free scatter of residuals in the residuals vs. predicted values plots (Fig. 9e and f) confirms homoscedasticity and independence of the residuals. These diagnostics demonstrate that the models are statistically valid and suitable for optimization, consistent with standard RSM practices in composite drilling research^76,77.

Fig. 9.

Diagnostic plots of F_dHSS and F_dHSS−TiN delamination factors (F_d): (a, b) Normal probability plots; (c, d) predicted vs. actual values; and (e, f) residuals vs. predicted values. (HSS: high-speed steel, HSS-TiN: high-speed steel coated with titanium nitride).

Surface contour and three-dimensional (3D) response surface plots for the delamination factor (F_d)

To better understand how drilling parameters interact with F_d, contour and 3D response surface plots were generated from the quadratic RSM models presented in Table 5. These visual tools help identify optimal machining settings by showing how F_d changes across the experimental range for both HSS and HSS-TiN drill types.

As shown in Fig. 10, the contour plots display iso-response lines with color gradients indicating F_d values; blue areas signify minimal damage, while yellow and red areas indicate higher delamination. For the HSS tool, the interaction between f and N at a fixed d (Fig. 10a) reveals a significant increase in F_d as f rises. The lowest F_d (~ 1.05) occurs at the slowest f (50 mm/min), while the highest (~ 1.80) appears at 150 mm/min. The curvature of the contours suggests an optimal N of about 1200 rpm to minimize damage. At a fixed speed (Fig. 10b), the combined effects of f and d cause F_d to increase together, reaching over 1.80 at the highest feed and diameter settings. Conversely, the interaction between N and d at a constant f (Fig. 10c) shows a much weaker effect, with F_d remaining relatively stable across most of the parameter range.

Fig. 10.

Contour plots for the predicted data of the different delamination factors (F_d): (a–c) F_dHSS and (d–f) F_dHSS−TiN, evaluated based on the cutting parameters of the Syagrus romanzoffiana fiber-reinforced bio-epoxy resin (Bio-EP-SrF) biocomposites produced (feed rate (f), spindle speed (N), and drill diameter (d)). (HSS: high-speed steel, HSS-TiN: high-speed steel coated with titanium nitride).

A similar analysis for the HSS-TiN tool (Figs. 10d-f) shows comparable trends but with consistently lower F_d values, emphasizing the protective effect of the coating. The F_d response to f and N (Fig. 10d) is diminished, with the lowest damage (F_d ~1.00) occurring at low f. The combined effect of feed and diameter (Fig. 10e) leads to a less significant increase in delamination than with the uncoated tool. The smoothest response again appears in the speed-diameter interaction (Fig. 10f), confirming that d is a more influential factor than N in this pairing.

As shown in Figs. 3D and 11 surface plots provide a complementary perspective on the non-linear relationships described by the models. The HSS tool’s surface for f and N (Fig. 11a) exhibits a saddle-shaped curvature, with a distinct valley at low feed and medium speed. The surface for f and d (Fig. 11b) forms a steep ridge, indicating a strong quadratic interaction. The speed-diameter surface (Fig. 11c) is nearly flat, with a slight incline toward larger diameters. The surfaces for the HSS-TiN tool (Figs. 11d-f) are similar in shape but generally lower, with a reduced maximum F_d, visually confirming the coating’s effectiveness in decreasing delamination across all parameter interactions.

Fig. 11.

Three-dimensional (3D) surface plots showing predicted data from Response Surface Methodology (RSM) for different delamination factors (F_d): (a–c) F_dHSS and (d–f) F_dHSS−TiN, evaluated based on Syagrus romanzoffiana fiber-reinforced bio-epoxy resin (Bio-EP-SrF) biocomposites produced. The parameters include feed rate (f), spindle speed (N), and drill diameter (d). The response surface plots were generated using Design-Expert software (Version 13). (HSS: high-speed steel, HSS-TiN: high-speed steel coated with titanium nitride).

Using graphical optimization is a well-established method in RSM studies for composite machining, where these plots are essential for identifying parameter sets that minimize undesirable outcomes⁷⁸. For the Bio-EP-SrF biocomposites analyzed, the blue regions on these plots clearly indicate the optimal conditions for minimal delamination: low f (50–70 mm/min), moderate N (1000–1200 rpm), and smaller d (5–6 mm).

Artificial neural network (ANN) modeling of the delamination factor (F_d)

As shown in Fig. 12, an ANN model, specifically an MLP architecture, was developed to predict F_d using the experimental dataset from Table 3. As depicted in the schematic in Fig. 12a, the network structure includes an input layer with three nodes representing the input variables: feed rate (f), spindle speed (N), and drill diameter (d). The optimal configuration for predicting F_d with the HSS tool was a 3–10-1 architecture with 10 neurons in a single hidden layer (Fig. 12b). In contrast, a 3–9-1 architecture with nine hidden neurons was found to be most effective for the HSS-TiN tool (Fig. 12c). These topologies were finalized after an extensive, iterative process that tested different hidden-layer and neuron counts. A hyperbolic tangent (tanh) activation function was used in the hidden layer, along with a linear function in the output layer, which is a standard approach for regression-based modeling in machining process optimization⁷⁹.

Fig. 12.

(a) Artificial Neural Network (ANN) architecture used for Syagrus romanzoffiana fiber-reinforced bio-epoxy resin (Bio-EP-SrF) biocomposites, (b) delamination factor F_dHSS data, and (c) F_dHSS−TiN data. (HSS: high-speed steel; HSS-TiN: high-speed steel coated with titanium nitride).

As detailed in Table 7, the performance evaluation of the ANN models was conducted using RMSE and R² metrics across the training, validation, and testing phases. The optimized 3–10-1 architecture for predicting FdHSS demonstrated high accuracy, registering a low overall RMSE of 1.18845E-2 and a high R² value of 9.84582E-1; this model was developed with an data split of 80% for training and 10% each for validation and testing. Similarly, the 3–9-1 model for F_dHSS−TiN also performed strongly, achieving an RMSE of 3.22276E-2 and an R² of 9.66791E-1, using 70%, 15%, and 15% of the data for training, validation, and testing, respectively. The consistently high R² values, all exceeding 0.96, confirm the models’ exceptional ability to learn the complex, nonlinear relationships within the machining data, surpassing traditional linear regression approaches⁸⁰.

Table 7.

Artificial neural network (ANN) architectures were assessed using root mean square error (RMSE) and correlation coefficient (R²) values for training, validation, and testing of delamination factors (F_d): F_dHSS and F_dHSS−TiN. (HSS: high-speed steel, HSS-TiN: high-speed steel coated with titanium nitride).

Model	Network structure		Percentage	Samples	RMSE	R 2
F dHSS	3–10-1	Training	80	21	1.92658 E-4	9.98418 E-1
		Test	10	3	1.30480 E-2	9.98591 E-1
		All	10	3	1.18845 E-2	9.84582 E-1
F dHSS−TiN	3–9-1	Training	70	19	3.17106 E-3	9.96412 E-1
		Test	15	4	5.52006 E-2	9.50287 E-1
		All	15	4	3.22276 E-2	9.66791 E-1

As shown in Fig. 13, the validation and error analysis of the ANN models confirm their high predictive accuracy. The error histograms (Figs. 13a for F_dHSS and 13c for F_dHSS−TiN) indicate that most prediction residuals are tightly clustered around zero across all data subsets, reflecting minimal error. The convergence curves (Fig. 13b and d) show a rapid decline, with the best validation performance reached at epoch 5 (0.013072) for F_dHSS and epoch 4 (0.0086853) for F_dHSS−TiN. Early stopping was employed at these points to prevent overfitting and ensure strong generalization. As depicted in Fig. 14, this robustness is further supported by the regression plots, which reveal an excellent match between predicted and actual values. For F_dHSS, the data points in the training (0.99842), test (0.9986), and overall (0.99846) sets are very close to the ideal fit line. Similarly, the F_dHSS−TiN model achieved strong correlations in training (0.99641), test (0.95029), and overall (0.96679) results, confirming its accuracy in mapping input parameters to the F_d⁸¹.

Fig. 13.

Distribution of errors and validation results from an Artificial Neural Network (ANN) for (a, b) delamination factor F_dHSS, and (c, d) F_dHSS−TiN data. (HSS: high-speed steel, HSS-TiN: high-speed steel coated with titanium nitride).

Fig. 14.

Validation scheme for regression of (a–c) delamination factor F_dHSS data and (d–f) F_dHSS−TiN data. (HSS: high-speed steel, HSS-TiN: high-speed steel coated with titanium nitride).

Table 8 presents the finalized mathematical equations of the ANN models. These equations describe F_d as a function of the normalized drilling parameters, utilizing weighted sums and hyperbolic tangent (tanh) activation functions of the hidden neurons (labeled H₁ to H₁₀ for F_dHSS and H₁ to H₉ for F_dHSS−TiN).

Table 8.

Mathematical models for various delamination factors (F_d) in drilling Syagrus Romanzoffiana fiber-reinforced bio-epoxy resin (Bio-EP-SrF) biocomposite using the artificial neural network (ANN) method. (HSS: high-speed steel, HSS-TiN: high-speed steel coated with titanium nitride).

ANN response
F dHSS
F dHSS−TiN

By providing closed-form equations (Table 8), which allow direct calculation of F_d from input parameters⁴⁵.

Figure 15 shows the 3D response surfaces created by the finalized ANN models, illustrating how F_d depends on the input parameters. For F_dHSS (Figs. 15a-c), the surfaces highlight a sharp increase in F_d with higher f and larger d, while a gentler curve shows the effect of N. In contrast, the surfaces for F_dHSS−TiN (Figs. 15d-f) are noticeably smoother and generally lower, visually underlining the superior ability of the TiN-coated tool to reduce damage. These visualizations assist in finding optimal machining conditions, such as low f and small d, to minimize delamination when drilling Bio-EP-SrF biocomposites⁸².

Fig. 15.

Three-dimensional (3D) surface plots of delamination factor (F_d) used Artificial Neural Network (ANN) for F_d vs. feed rate (f), spindle speed (N), and drill diameter (d) of Syagrus romanzoffiana fiber-reinforced bio-epoxy resin (Bio-EP-SrF) biocomposites elaborated: (a–c) F_dHSS and (d–f) F_dHSS−TiN, using MATLAB (Neural Network Toolbox). (HSS: high-speed steel, HSS-TiN: high-speed steel coated with titanium nitride).

Evaluation of response surface methodology (RSM) and artificial neural network (ANN) predictive models

While both RSM and ANN have been used in studies on drilling NFCs, rigorous comparative analyses under identical experimental conditions are notably rare. This is a significant methodological gap, as RSM is limited by its assumed polynomial structure, which may not effectively capture the complex, non-linear damage mechanics inherent to anisotropic biocomposites⁵². Conversely, ANN’s model-free, data-driven approach is naturally suited to such complexity. Addressing this gap, the present analysis provides the first direct comparison, showing that the ANN model reduces prediction error by 60–75% compared to the best quadratic RSM model derived from the same data set. This result strongly supports the superior predictive ability of ANNs, indicating they should be the preferred approach for future drilling optimization of advanced biocomposites^83,84. RSM, while transparent and statistically interpretable, is inherently limited by its quadratic polynomial assumption, which may inadequately capture the highly nonlinear and anisotropic damage mechanics typical of natural fiber biocomposites^85–87. In contrast, the data-driven ANN approach excels at modeling such complexity but carries a potential risk of overfitting, particularly with smaller datasets. This risk was effectively mitigated here through early stopping, independent test sets, fixed-seed partitioning, and strong performance on unseen data (R² > 0.95, RMSE < 0.032), confirming robust generalization^88,89.

Table 9 shows a comparison of the forecast accuracy for the RSM and ANN methods. The ANN model achieved higher R² values (0.967–0.985) compared to RSM (0.944–0.957). Its better predictive accuracy and fit demonstrate its superior ability to model the complex, non-linear relationships between the drilling parameters—f, N, and d—and the resulting F_d in Bio-EP-SrF biocomposites⁹⁰.

Table 9.

Comparison of experimental (EXP), response surface methodology (RSM), and artificial neural network (ANN) approaches for various delamination factors (F_d). (HSS: high-speed steel, HSS-TiN: high-speed steel coated with titanium nitride).

Order	FdHSS			FdHSS−TiN
	EXP	RSM	ANN	EXP	RSM	ANN
1	1.0641	1.0843	1.0415	1.0886	1.0469	1.1473
2	1.2939	1.3311	1.2974	1.1969	1.1810	1.1729
3	1.5936	1.5779	1.5857	1.4212	1.3150	1.4151
4	1.0552	1.0644	1.0880	1.0641	1.0888	1.0877
5	1.3815	1.4214	1.4195	1.1577	1.2941	1.1536
6	1.8303	1.7785	1.9171	1.3588	1.4994	1.3640
7	1.1106	1.0445	1.0797	1.1315	1.1306	1.2182
8	1.5746	1.5118	1.5698	1.4351	1.4071	1.3958
9	1.8725	1.9791	1.8704	1.7735	1.6837	1.7624
10	1.0714	1.0524	1.0821	1.0944	1.0593	1.0746
11	1.2296	1.2183	1.2412	1.2168	1.2266	1.2421
12	1.2958	1.3842	1.3005	1.3624	1.3939	1.3663
13	1.0462	1.0448	1.0472	1.0713	1.0588	1.0758
14	1.3582	1.3209	1.3555	1.2538	1.2974	1.2359
15	1.5746	1.5970	1.7426	1.5143	1.5359	1.5061
16	1.0358	1.0372	1.0311	1.0554	1.0584	1.0670
17	1.2287	1.4235	1.4121	1.4740	1.3682	1.4519
18	1.8083	1.8099	1.8430	1.6818	1.6780	1.6612
19	1.0476	1.0311	1.1189	1.0212	1.0675	1.0495
20	1.1487	1.1431	1.1603	1.2839	1.2569	1.2757
21	1.3250	1.2550	1.3346	1.472	1.4464	1.4122
22	1.0386	1.0317	1.0392	1.0379	1.0389	1.1620
23	1.2371	1.2539	1.2446	1.3094	1.2996	1.3895
24	1.3614	1.4760	1.3566	1.5738	1.5603	1.5597
25	1.0306	1.0322	1.0319	1.0449	1.0103	1.0886
26	1.4210	1.3647	1.4178	1.2419	1.3423	1.2548
27	1.9506	1.6971	1.8184	1.6808	1.6742	1.6754

Figure 16 presents a graphical comparison of the predictive abilities of both modeling techniques, plotting the estimated F_d values against the actual experimental results. For F_dHSS (Fig. 16a), the ANN model data points are more closely clustered along the ideal Y = T line, indicating better agreement with the measured values. In contrast, the RSM predictions show more scatter. A similar trend appears for F_dHSS−TiN (Fig. 16b), where the ANN’s output displays a more concentrated distribution of points along the central fit line compared to the RSM. This visual data strongly supports the numerical results in Table 9, clearly demonstrating the ANN’s superior ability to capture the complex, non-linear relationships between drilling parameters and their effect on delamination.

Fig. 16.

Comparison of experimental and predicted delamination factor (F_d) using Response Surface Methodology (RSM) and Artificial Neural Network (ANN) models: (a) F_dHSS data, (b) F_dHSS−TiN data. (HSS: high-speed steel, HSS-TiN: high-speed steel coated with titanium nitride).

Figure 17 presents a residual analysis that provides further insight into the performance differences between the models. For both F_dHSS (Fig. 17a) and F_dHSS−TiN (Fig. 17b), the residuals from the RSM model show wider, more random scatter, with several deviations exceeding 0.05, indicating it captures the data patterns less effectively. In contrast, the ANN model residuals are closely clustered around zero, with very few surpassing 0.02, demonstrating its greater predictive stability and lower error. This much tighter distribution of ANN residuals confirms its superior ability to adapt to the data’s inherent variability, a well-known advantage of such advanced computational techniques in machining composite materials⁹¹.

Fig. 17.

Delamination factor (F_d) residuals for Response Surface Methodology (RSM) and Artificial Neural Network (ANN): (a) F_dHSS data, (b) F_dHSS−TiN data. (HSS: high-speed steel, HSS-TiN: high-speed steel coated with titanium nitride).

In summary, the RSM provides a transparent and efficient way to optimize parameters. However, for predicting F_d in Bio-EP-SrF biocomposites, the ANN model shows higher accuracy and flexibility. Its improved ability to model complex, non-linear interactions between drilling parameters makes it a more suitable and reliable choice for achieving precise F_d predictions across various machining conditions.

Optimization of responses

To minimize F_d in drilled Bio-EP-SrF biocomposites, a multi-response optimization was carried out using a DFN methodology. Figure 18 displays the desirability plot, which considers F_d for both HSS and HSS-TiN tool types to reduce damage for each. This plot uses a normalized scale from 0 to 1, where values near 1 indicate the most favorable parameter combinations that effectively lessen delamination for both tools. The highest desirability regions are clearly indicated by specific contour patterns and a color gradient, offering clear guidance for selecting the best drilling parameters⁹².

Fig. 18.

Desirability function (DFN) plot showing the combination of delamination factors (F_d): F_dHSS and F_dHSS−TiN, generated using Design-Expert software (Version 13). (HSS: high-speed steel, HSS-TiN: high-speed steel coated with titanium nitride).

Figure 19 summarizes the distinct optimal drilling configurations identified for each tool type at a desirability of 0.991. The conventional HSS tool achieved a minimal predicted F_d of 1.03199 using a 10 mm d, an f of 50 mm/min, and an N of 1419.49 rpm. The HSS-TiN tool showed better performance, producing a lower predicted F_d of 1.02319 under the same f, N, and d conditions.

Fig. 19.

Optimal solutions for Syagrus romanzoffiana fiber-reinforced bio-epoxy resin (Bio-EP-SrF) biocomposite developed for delamination factors (F_d): F_dHSS and F_dHSS−TiN. (HSS: high-speed steel, HSS-TiN: high-speed steel coated with titanium nitride).

The analysis shows that the best machining parameters do not pinpoint a single exact point. Both desirability plots and response surfaces highlight a broad plateau of minimal delamination, especially within an f range of 50–70 mm/min and an N range of approximately 1000–1200 rpm. This plateau indicates that small, unavoidable changes in machining conditions during production will not significantly impact hole quality. As a result, this “flat” optimum creates a robust processing window suitable for practical shop-floor use^93,94.

Figure 19, which shows overlaid contours and data points, confirms that these optimal solutions are stable and well within the boundaries of the tested experimental region. The differences in the parameters highlight the coating’s performance advantage, enabling a higher material removal rate without a significant increase in drilling-induced damage.

The optimized drilling conditions reported in this study are based on model-based predictions within a statistically validated CCD domain. Since the optimal solutions are well within the experimental region and supported by replicated center-point error estimates, no additional confirmatory drilling trials at the optimum were performed. The strong agreement between RSM- and ANN-predicted responses further supports the robustness and stability of the identified optima⁹⁵.

The results indicate that hole quality in fully bio-based composites reinforced with SrF can be optimized to reduce delamination, with F_d values ranging from 1.023 to 1.032. These values, achieved with an f of 50 mm/min and an HSS-TiN tool, are comparable to or better than the delamination levels usually reported for carbon and glass-fiber composites machined with conventional or coated tools under dry or minimum-quantity lubrication conditions (F_d = 1.05–1.35)^96–98. Therefore, in terms of machinability and damage resistance, this biocomposite meets the industrial standards required for many lightweight structural and semi-structural applications, including automotive interiors, sporting equipment, and certain aerospace components. When considering its excellent hole quality along with its high specific strength, the complete biodegradability of its Bio-EP matrix, and the sustainable, non-food-source origin of its fibers, Sr biocomposite presents a compelling, eco-friendly alternative to traditional synthetic composites in industries focused on sustainability.

Although the optimised drilling window proposed in this study demonstrates strong potential for industrial application, several practical aspects should be considered. First, progressive tool wear may gradually increase thrust force and consequently the delamination tendency during long-term production. Second, laminate thickness and stacking sequence variations may alter the local stiffness distribution and modify the delamination response. Finally, at higher production rates, heat accumulation and process instability may influence damage evolution. These factors were beyond the scope of the present experimental programme but should be addressed in future implementation studies^99–101.

Conclusion

This study offers the first systematic investigation of drilling-induced delamination in Syagrus romanzoffiana fiber–reinforced bio-epoxy (Bio-EP-SrF) composites, defining its fundamental machining behavior through combined experimentation and modeling. It represents the first dedicated analysis of drilling in this previously unexamined biocomposite, showing that its delamination response is mainly influenced by feed rate (f) and drill diameter (d), within an optimal spindle speed (N) range of about 1200 rpm. This intermediate speed reduces damage by balancing thrust force reduction via matrix softening while avoiding excessive thermal degradation at the fiber–matrix interface. A direct comparison of tool coatings demonstrated that high-speed steel coated with titanium nitride (HSS-TiN) drills consistently yield lower delamination factors (F_d) than HSS tools under same conditions, thanks to less friction, better thermal stability, and reduced fiber pull-out.

Mechanistically, f was identified as the most critical parameter, with a reduction from 150 to 50 mm/min lowering d by 45–50%, confirming a thrust-driven damage mechanism. Increasing d from 5 to 10 mm resulted in a 12–18% increase in delamination, consistent with the scaling of thrust force and the damage zone. Under optimized parameters (f = 50 mm/min, N ≈ 1200–1400 rpm), minimum d of 1.023 (HSS-TiN) and 1.032 (HSS) were achieved, demonstrating machinability competitive with that of conventional synthetic composites such as carbon fiber-reinforced polymer and glass fiber-reinforced polymer under dry drilling. These outcomes highlight damage mechanisms and optimal conditions that differ from those reported for analogous fibers (e.g., jute or flax), which points to the importance of fiber-specific machining studies to enable accurate industrial adoption.

The contribution of the modeling is twofold. Statistically significant quadratic Response Surface Methodology (RSM) models offer interpretable insights into parameter interactions, especially the synergistic f × N effect. Additionally, developed Artificial Neural Network (ANN) models achieved superior predictive accuracy (R² > 0.96), better capturing nonlinear damage behavior than RSM. Importantly, this work provides fully reproducible ANN architectures and closed-form equations, enabling direct industrial application—a feature often missing in the literature on Natural fiber-reinforced polymer machining.

While the identified optimal conditions—low feed rate, moderate spindle speed, and HSS-TiN tooling—offer a robust processing window suitable for industrial use, their practical application entails certain limitations. These include the progressive wear of tools in high-volume production, which may increase drilling forces, and the restriction of the results to a specific laminate thickness, as thicker stacks could worsen exit delamination.

The findings provide a strong processing window tolerant to small parameter changes, making them easily applicable to shop-floor conditions. Using HSS-TiN tooling with optimized parameters allows for high-quality hole production without complex cooling or backing supports. Future work should focus on combined force–temperature measurements, tool geometry optimization for SrF morphology, and evaluating the fatigue performance of drilled joints. Expanding the ANN datasets to include variables such as fiber volume fraction and laminate thickness would help make the models more general.

Finally, this work establishes Bio-EP-SrF as a viable, machinable, and sustainable alternative to synthetic FRPs, provides material-specific drilling guidelines, and shows that ANN-based modeling offers improved predictive ability for the complex machining behavior of natural fiber composites.

Author contributions

Oussama Ferfari: Investigation, Methodology, Conceptualization, Formal analysis, Writing – review & editingAhmed Belaadi: Investigation, Methodology, Supervision, Methodology, Conceptualization, Writing - review & editing.Prabu Krishnasamy: Investigation, Writing - review & editing.Djamel Ghernaout: Investigation, Writing - review & editing.Herbert Mukalazi: Investigation, Writing - review & editing.

Data availability

The datasets generated and/or analyzed during the current study are available from the corresponding author on reasonable request.

Declarations

Competing interests

The authors declare no competing interests.