Experimental and AI-assisted optimization of mechanical performance of hybrid flax

Abstract

The research aims to bridge the gap by providing an integrated approach to material design, process optimization, and intelligent failure prediction. A novel approach to material design and process optimization was developed to examine the impact of stacking sequences and processing parameters on the mechanical behavior of flax-basalt composite laminates. Five laminate stacks were manufactured by hand lay-up and compression curing techniques. These stacks were then characterized by tensile, flexural, impact, interlaminar fracture toughness (ILFT), and energy absorption density (EAD) properties. Among the laminate stacks, B-F-F-F-B (S2) showed superior properties with tensile, flexural, impact, ILFT, and EAD properties of 150 MPa, 185 MPa, 14.2 J, 1580 J/m², and 0.68 J/cm³, respectively. To further optimize the material properties, the Response Surface Methodology (RSM) was combined with a novel approach based on a hybrid DQN-LSTM-Mother Optimization Algorithm (MOA) approach. In addition, scanning electron microscope (SEM) analysis was incorporated to interpret the material properties and provide a direct link to material structure and properties. This approach represents a major advancement over existing RSM-ANN and machine learning techniques for designing high-performance composite materials.

Keywords

hybrid flax–basalt composites stacking sequence mechanical properties interlaminar fracture toughness energy absorption density response surface methodology deep learning

Introduction

There has been an increase in demand for sustainable, high-performance composite materials. Flax fibers are commonly used due to their low density, degradability, and specific properties.^1,2 Flax fibers face limitations in moisture absorption, variability in property, and low strength. Hybridization with high-performance fibers such as basalt has been proposed to overcome the limitations of flax fibers. Hybridization has proven to enhance the strength and durability of composite materials.^3,4 The performance of hybrid flax and basalt laminates is largely influenced by processing parameters such as stacking sequence and resin content, which affect stress distribution, interfacial adhesion, and failure mechanisms.^5–7 Optimizing parameters is necessary to enhance the performance of the laminates.

Recent work has focused on hybrid composite systems and data-driven prediction methods, with greater emphasis on sustainability and optimization. For example, machine learning methods have been used to predict the mechanical behavior of hybrid natural fiber composites.⁸ In contrast, agro-waste composite optimization has been carried out using a combination of methods.⁹ Additionally, advanced hybrid composite systems and decision-making tools have been developed,¹⁰ while the impact of stacking sequence and thickness variation on the thermo-mechanical properties has been investigated.¹¹ While the above-mentioned research demonstrates the viability and effectiveness of using hybrid composite materials and prediction methods, the research has been largely restricted to material characterization and/or optimization methods. However, the integration of material design, process optimization, and failure prediction using AI methods has yet to be fully investigated, specifically in the context of hybrid flax-basalt composites. Nevertheless, it is worth noting that the majority of the literature focuses on either experimental characterization or optimization approaches, without a unified framework for material design, optimization, and failure prediction. RSM has been widely used for modeling and optimizing composite processing parameters with reduced experimental effort.^12,13 Meanwhile, artificial intelligence methods such as ANNs, CNNs, and regression models were used to predict complex nonlinear behaviors.^14,15 However, these methods often failed to incorporate feature selection, sequential learning, and multi-objective optimization. Besides, failure analysis by Scanning Electron Microscopy (SEM) is often qualitative, lacks predictive capabilities, and is limited in its integration with AI models.

Thus, the novelty of the present work lies in the formulation of a comprehensive experimental and computational framework that combines the RSM approach with the hybrid DQN-LSTM-MOA model for simultaneous optimization and prediction of mechanical properties of hybrid composites of flax and basalt. Moreover, the SEM technique is utilized for data-driven failure mode prediction by incorporating the microstructural analysis into the AI framework. This approach enables the direct linkage between processing parameters, microstructure, and mechanical performance, thereby facilitating the optimization of high-performance composites.

Literature study

Because of their biocompatibility, low mass density, and competitive specific strength, multiple researchers, including Irshad et al.,¹⁶ have studied the applications of flax fiber. They have also shown potential for use as lightweight structural materials. Factors such as inconsistent fiber properties, limited water uptake, and durability have limited the benefits of this fiber type over time (Yan and Xu).¹⁷ Studies have indicated that it is critical to increase the bond between the fiber and the polymer to improve the mechanical performance of flax-reinforced polymers.¹⁸

Basalt fiber is an environmentally friendly alternative for glass fiber that is produced from heating volcanic rock until it melts. It is able to produce stronger materials than traditional glass fiber with better thermal resistance and virtually constant chemical stability Chellan et al.¹⁹ Due to their high durability and ability to maintain their strength during thermal cycling, basalt composites are able to provide for high-end engineering applications Raja and Devarajan.²⁰

Combining flax and basalt in hybrid composites has demonstrated improved mechanical properties along with some sustainability benefits. Previous literature indicated that hybridizing the two materials increased laminate stiffness and interlaminar shear strength and reduced brittleness compared with solely using natural fibers. The strength of basalt combined with the flexibility or toughness of flax created a system with higher damage tolerance. However, there are inconsistencies between the studies regarding laminate configurations, which limit the ability to compare findings across studies (Dogar et al.²¹).

Numerous studies suggest that the stacking sequence strongly impacts load distribution, failure initiation, and energy absorption in hybrid composites. By judiciously locating the basalt and flax layers within these composite structures, significant gains in tensile, flexural, and impact characteristics can be achieved (Shi et al.²²). Additionally, the content of resin was revealed to be an important variable affecting wetting-out of the fibers, amount of voids, interfacial adhesion, and overall integrity of the laminate. Applying either too little or too much resin results in performance loss; thus, a need exists for the development of optimal resin content during lamination of hybrid composites Taheri et al.²³

RSM has been used to optimize composite manufacturing parameters because it can provide models of interactions among the manufacturing parameters with fewer experimental results. For example, RSM has been used to optimize fiber ratio, curing conditions, printing conditions, and reinforcement levels. RSM has been shown to predict responses that lead to optimal conditions for improving tensile, flexural, and interlaminar shear strength. This has been demonstrated by Thanikodi et al.²⁴

The effectiveness of deep learning techniques for accurately predicting and categorizing fiber composite meltdown mechanisms has been enhanced by advances over the past few years. The different neuro-network modalities used in past research include; CNNs (Convolutional Neural Networks) being used for fracture surface identification, MLPs (Multilayer Perceptrons) and DBNs (Deep Belief Networks) for correlating mechanical test data to how the composite broke down, Support Vector Regression and Random Forest Regression for identifying matrix cracking and fiber ply mechanisms, and Inception-V3 networks used to identify mixed mode failure from SEM images. All of these modeling techniques successfully extract microstructural features and therefore enable automated detection of fiber composite defects (Rothenhäusler et al.²⁵; Chang et al.²⁶; Wang et al.²⁷). However, hybrid composites of flax/basalt will require integrating deep learning techniques with experimental optimization techniques such as RSM; very little research has been conducted on this to date.

Research gap

Composites such as flax, basalt, and their hybrids have been investigated by various researchers, but most studies have been focused on specific aspects, such as material characterization, optimization, or AI applications. Although the RSM-machine learning approach has been explored in composites, its application to hybrid composites such as flax-basalt is less well established, especially for optimizing stacking sequence and resin content. Moreover, integrated approaches for processing parameters, microstructure, and multi-response mechanical properties are less explored, while advanced approaches for failure prediction have not been explored. To bridge these gaps, this study proposes a novel experimental-computational framework for hybrid feature selection, prediction, and optimization, as well as failure interpretation, integrating RSM with a hybrid DQN-LSTM-MOA model and SEM-based microstructural analysis. This approach provides a direct process-structure-property relationship, which is a breakthrough compared to existing RSM-ML models for hybrid flax-basalt composites.

Objectives

• To fabricate hybrid flax–basalt composite laminates with different stacking sequences and resin contents.

• To characterize the mechanical response of the developed laminates under tensile, flexural, and impact loading conditions.

• To investigate failure mechanisms using microscopic (SEM) observations for different laminate configurations.

• To apply RSM to identify the optimal combination of stacking sequence and resin percentage for improved mechanical properties.

• To integrate RSM-based optimization and deep learning–based failure prediction into a unified framework for designing high-performance hybrid flax–basalt composites.

Proposed methodology

A complete overview of the methodology for obtaining FLX-BAS composite laminate materials, along with their mechanical properties and microstructural characteristics, is presented in Figure 1. The approach starts with an appropriate selection of fiber and resin materials that are mechanically compatible and used in eco-friendly operations. Manufacturing of the composite laminates consists of five different stack arrangements of both types of fiber types, curing simultaneously under controlled ambient conditions to provide defect-free specimens. Mechanical properties of manufactured laminates are determined by conducting a series of laboratory tests, including tensile and flexural tests; energy-absorption tests; impact tests; and interlinear fracture-toughness tests. SEM image analysis has allowed for determination of the major failure mechanisms associated with composite laminates, such as inter-fiber matrix bonding failure, inter-layer matrix rupture, and fiber matrix debonding due to tensile loads, etc., and, through the process of RSM, has predict how each of these mechanisms will affect composite performance responses. The DQN-LSTM-MOA hybrid AI framework provides an additional tool for accurately predicting composite performance characteristics and optimizing laminate performance characteristics for a particular fabric stack sequence and/or composite fabrication parameters, thereby increasing the reliability of selecting those factors.

Figure 1.

Flow chart of proposed methodology.

Material selection

The objective of this research was to evaluate the compatibility of flax and basalt fibers in terms of the special properties they possess. Flax fibers are natural, renewable resources with low density and good tensile properties. They are also ductile materials with low environmental impact. On the other hand, basalt fibers, derived from volcanic rocks, exhibit exceptionally high tensile strength. They also possess high thermal stability and chemical durability. The combination of flax and basalt fibers ensures the composite material has the required balance of strength, stiffness, and impact resistance. The epoxy resin was adopted as the matrix material because it provides the required bonding to the fibers and, when handled properly, improves the mechanical behavior of the composite. The materials were obtained as fabric and the required ratios of resin and hardener were used to produce the composite materials. The bonding nature of the selected materials was later examined using SEM image analysis.

Fabrication of composites

The flax fibers used in the present work had an areal density of approximately 300 gsm in plain-weave fabric. The basalt fibers utilized in the present work had an areal density of 400 gsm. The flax fabric was first dried in an oven at 60°C for 24 h before composite processing. The epoxy resin system used in the current research comprised LY556 resin and HY951 hardener, mixed at a 10:1 weight ratio. The epoxy resin was mixed with the flax and basalt fibers. Hybrid composites were prepared using the hand lay-up technique and cured under a compression of 0.5 MPa at 100°C for 4 h. Degassing and rolling of the flax and basalt fabric were performed to ensure proper fibers wetting. Figure 2 illustrates the fabrication process of Hybrid Flax-Basalt composite laminates. The stacking sequences, including pure forms (FFFFF, BBBBB) and hybrid forms (BFFFB, FBBBF, FBFBF), were designed to investigate interface interactions, stress distributions, and reinforcement. Triplicate laminate samples were prepared to ensure reproducibility. The void content of the fabricated laminates was evaluated using the density method, comparing the measured laminate density with the theoretical density calculated from the fiber and resin proportions. The void fraction was found to be below 3%, indicating good impregnation of the fibers and uniform laminate quality.

Figure 2.

Fabrication of hybrid flax-basalt composite laminates.

Stacking sequence and specimen preparation

The stacking sequence (Figure 3) significantly influences load transfer, stiffness, and failure behavior in laminated composites. Five representative configurations were selected: all-flax (S1) as the baseline, basalt-outer (S2: BFFFB) for improved surface strength, alternating (S3: FBFBF) for balanced load sharing, basalt-rich (S4: FBBBF) for higher rigidity, and all-basalt (S5) as the synthetic benchmark. These configurations enable a systematic evaluation of the effect of ply arrangement on mechanical performance and failure mechanisms in hybrid flax–basalt composites.

Figure 3.

Stacking sequence.

The stacking sequences were selected before the experiment to investigate different structural configurations. These include the natural fiber baseline (S1), the synthetic fiber baseline (S5), and various hybrid configurations (S2, S3, and S4). This approach will provide a comprehensive evaluation of the laminate’s properties, considering the influence of stacking sequence, including basalt placement at the outer surfaces, the alternating effect, and the basalt core. Although the number of stacking sequences used in this study is limited, the chosen configurations represent the major laminate configurations in terms of outer layer reinforcement, core reinforcement, and alternating effects.

Selection of parameters

Selection of relevant input and output parameters is critical for understanding how fabrication conditions affect the mechanical properties of hybrid flax-basalt composite laminates. Four major fabrications, which were critical input parameters, were selected for this study based on their significant influence on composite matrix interaction, quality, and structural integrity. The parameters include fiber weight fraction, epoxy resin content, curing temperature, and curing time. The levels of the factors were determined for the BBD based on the results of preliminary experiments, material limitations, and literature on hybrid composites of natural and synthetic fibers. The fiber weight fraction range was determined to be 30–50%, as it is critical for proper impregnation without compromising matrix interaction. The range for epoxy resin content was determined to be 50–70%, as it is critical for proper impregnation and minimizing voids. The range for excessive epoxy resin content would result in decreased mechanical properties of composites. The range for curing temperature was determined to be between 80 and 120°C, as it is critical for proper curing without degrading the matrix. The range for curing time was determined to be between 2 and 6 h, as it is critical for proper curing without degrading the matrix. The range was determined based on recommendations for the epoxy resin system (LY556 and HY951).

Mechanical testing

The mechanical properties of hybrid flax-basalt laminates were tested under standardized conditions. Specimens were prepared according to ASTM standards and tested at room temperature under controlled laboratory conditions. The experiments were conducted in triplicate for all stacking sequences (S1 to S5). All mechanical properties were tested in triplicate for all stacking sequences. This is according to standard experimental procedures for composite materials. The average values were determined for all properties measured. Moreover, all specimens were tested under controlled conditions to account for natural fiber properties. This ensured consistent results for all specimens.

Tensile strength

The tensile test was carried out according to ASTM D3039 on a Universal Testing Machine (UTM) equipped with wedge grips. Loading was performed at a crosshead speed of 2 mm/min until fracture occurred. Figure 4 shows tensile testing and a fractured tensile specimen, which indicates the modes of failure in the fibers and the matrix material.

σ_{t} = σ_{f, f l a x} V_{f,, f l a x} + σ_{f, b a s a l t} V_{f, b a s a l t} + σ_{m} V_{m}

(1)

where,

V_{f}, V_{m}

denotes the volume fraction of fiber and matrix and

σ_{f}, σ_{m}

are the TS and fiber and matrix.

Figure 4.

Tensile strength.

The above relation (equation (1)) is a simplified rule of mixtures. This rule assumes an ideal load sharing mechanism and an ideal bond condition. However, the actual behavior of composite materials, especially those with a hybrid laminate configuration, may not follow this rule due to various factors. Therefore, this equation is an approximate analytical tool to understand the trend of TS. This simplified rule is applicable to woven composite materials. This is because this rule provides an approximate indication of the tensile behavior of composite materials. Although the load-sharing mechanism is complex due to crimped fibers and the heterogeneous fiber distribution, this rule is applicable for understanding the trend in tensile behavior. This is because this rule provides an approximate indication of the tensile behavior and the contributions of both flax and basalt fibers to the tensile strength.

Flexural strength

Flexural strength and modulus were measured using the three-point bending test consistent with ASTM D790. The span depth ratio was 16:1, and the crosshead speed was 2 mm/min. The images of the specimens after the bending test are presented in Figure 5. These images show the laminate bending.

σ_{f} = \frac{3 P L}{2 b h^{2}}

(2)

where,

P

denotes applied load at fracture

L

is the support span

b

indicates specimen width and

h

represents laminate thickness

Figure 5.

Flexural strength.

Impact energy

Impact resistance was evaluated using the Izod Impact Test (ASTM D256). The energy absorbed in the fracture process was measured. Figure 6 displays the laminates after impact testing, demonstrating their impact resistance and energy absorption. These values can be calculated using equation (3).

E_{i m p a c t} = \frac{W_{a b s o r b e d}}{b h}

(3)

Figure 6.

Impact energy test.

Interlaminar fracture toughness

The delamination toughness was quantified by the Double Cantilever Beam (DCB) test according to ASTM D5528. The value of GIc was calculated using the Modified Beam Theory based on load-displacement measurements. Figure 7 displays the DCB test specimens with visible crack propagation and delamination. The GIc was obtained using equation (4).

G_{I C} = \frac{3 P δ}{2 b (a + Δ)}

(4)

where,

P

is the applied load,

δ

is the load point displacement,

b

is the specimen width,

a

is the crack length, and

Δ

is the correction factor obtained from compliance calibration. This approach ensures accurate evaluation of delamination resistance using the load–displacement response recorded during DCB testing.

Figure 7.

Interlaminar fracture toughness test.

Energy absorption density

EAD, which represents the amount of energy a material can absorb per unit volume before failure, is summarized by the area under the Stress/Strain Curve, with the total toughness of a given material equal to this area. EAD can be calculated using equation (5), and the schematic representation in Figure 8 represents EAD from the Stress/Strain Curve.

E A D = \frac{A_{c u r v e}}{V}

(5)

where

A_{c u r v e}

is the area under stress-strain curve up to failure and

V

is the volume of the specimen.

Figure 8.

Schematic illustration of EAD derived from the stress–strain curve.

Microstructural analysis (SEM)

To identify the different failure mechanisms, a SEM provided further insight into broken samples resulting from mechanical testing. Prior to SEM analysis, all fractured samples were prepared by mounting and coating them. Following sample preparation, SEM inspections took place with different degrees of magnification and included observations on how matrix cracking, fiber breakage, delamination, and void formations all contributed to mechanical failure of the laminate samples. Additionally, these observations were related back to the mechanical performance associated with each laminate’s stacking sequence and resin content in order to comprehend how each of these factors influenced laminate failure behavior. The void fraction of the fabricated laminates was also assessed by image-based analysis of the SEM images. In this regard, the void area was identified according to the contrast difference between the matrix material and the voids, and the area fraction was estimated accordingly.

Response surface methodology analysis

RSM was used to establish how much input parameters (fiber weight fraction (A), polymer resin content (B), curing temperature (C), and curing duration (D)) affect the mechanical properties of hybrid flax and basalt laminates. A Box-Behnken design with 30 experimental trials was used to create valid statistical regression models for five response variables that include tensile strength, flexural strength, impact energy, inter-laminar fracture toughness (IFT), and EAD.

The construction of a second-order polynomial equation allows RSM to link the predictor Variables with measured Responses while minimizing the number of experiments. The coded values for each factor were created based on the researcher-specified ranges, ensuring that the experimental design points are evenly distributed within the experimental domain. The responses of each output were fitted to a standard quadratic model as shown in equation (6) as described by Boobalan et al.²⁸

Y = β_{0} + \sum_{i = 1}^{4} β_{i} X_{i} + \sum_{i = 1}^{4} β_{i i} X_{i}^{2} + \sum_{i = 1}^{4} β_{i j} X_{i} X_{j}

(6)

where,

Y

indicate the predicted response

β_{0}

is the intercept

β_{i}

represents the linear coefficients

β_{i i}

is the quadratic coefficients

β_{i j}

illustrate the interaction coefficients

X_{i}

denotes the coded input variable.

To verify the reliability of the regression models, multicollinearity among the input variables was evaluated before model development. The correlation matrix of the input factors showed low correlation coefficients ( $| r | < 0.8$ ), indicating that the variables are statistically independent. In addition, the Variance Inflation Factor (VIF) values for all factors were found to be less than 5, confirming the absence of significant multicollinearity. These results confirm that the selected process parameters can be varied independently and that the developed RSM models are statistically robust.

Application of AI

The integrated framework uses experimental RSM results as input to a DQN for feature selection and a subsequent LSTM for predicting mechanical properties. Finally, the predicted results are optimized using the MOA to obtain the optimum processing parameters. The DQN architecture consists of fully connected layers with a learning rate of 0.001 and a replay memory. Similarly, the LSTM architecture consists of 64 hidden units, and training is performed over 100 epochs using the Adam optimizer. The network’s input parameters are fiber weight fraction, epoxy weight fraction, curing temperature, and curing time. Similarly, the output parameters consist of TS, FS, IE, ILFT, and EAD. The dataset consists of 30 BBD experiments. The dataset was split in the ratio 80-20 for training and testing. The performance was evaluated using MSE, MAE, and RMSE. Despite the dataset’s limited size, the structured DOE design and optimized network architecture minimize the risk of overfitting. The close correlation between the predicted and experimentally obtained results demonstrates the robustness and reliability of the DQN-LSTM-MOA framework for predicting the mechanical properties.

The dataset consists of 30 BBD experiments. However, the structured design ensures high information density in the dataset. Recent research has successfully used such DOE-based datasets for machine learning-based prediction and optimization of composite properties. These structured datasets have been successfully used to obtain reliable predictions with machine learning algorithms, even with a limited number of samples.

Deep Q network for feature selection

To determine the most influential input parameters from the RSM dataset, an automated process is used to develop a DQN. The learning process of the DQN aims to minimize the difference between the predicted Q-value and the Q-value for each state and action pair. This process is defined mathematically by the Deep Q-Network loss function shown in equation (7), as presented in Shonkwiler et al.²⁹

L_{D Q N} = {(y_{t} - Q (s, α, θ))}^{2}

(7)

where,

L_{D Q N}

represent the Loss function of the DQN used to train the network,

y_{t}

is the Target Q-value calculated from the Bellman equation, representing the expected reward for taking action

α

in state

s

Q (s, α, θ)

denotes Predicted Q-value for state

s

and action

α

by the DQN with parameters

θ

s

is the Current state,

α

predicted Action taken and

θ

shows the Expectation operator, averaging over all state-action pairs in the training dataset. This equation enables the model to select only the most relevant features, which are then supplied to the LSTM model for improved prediction accuracy.

The Deep Q-Network (DQN) was selected for feature selection over traditional methods such as PCA or Genetic Algorithms (GA) because reinforcement learning–based selectors dynamically learn the importance of features in non-linear and interdependent spaces. Unlike PCA, which only captures linear correlations, or GA, which relies on stochastic search without adaptive learning, DQN optimizes feature selection through reward-based learning. Similar approaches have been shown to significantly improve feature selection performance in engineering applications (Yang et al.³⁰).

LSTM model for mechanical property prediction

Five mechanical responses are predicted using the selected DQN features as input to the LSTM. The LSTM’s training goal is to reduce the error between predicted and experimental values. The LSTM uses Mean Squared Error (MSE) to measure the gap between predictions and reality as shown in equation (8) Deng et al.³¹

M S E = \frac{1}{N} \sum_{i = 1}^{N} {‖ Y^{i} - \hat{Y} ‖}^{2}

(8)

where,

N

is total number of training samples,

Y^{i}

denotes Actual value of the mechanical property,

\hat{Y}

represent LSTM-predicted value. This equation represents the primary loss function used to train and evaluate the predictive capability of the LSTM network.

Although the LSTM is designed for time-series data, it can be used to model datasets with highly interdependent variables. In the present work, the variables in the experimental data are the different parameters involved in the process, which have high interdependency between the mechanical responses. This means that the variables can be considered as a series of dependent features rather than independent features. This is where the LSTM model is used for modeling the complex nonlinear relationship between the variables by utilizing the memory-based learning capacity. This is because the model can better represent the relationship between the features compared to other machine learning models. This is why the LSTM model is used in the present work for modeling the complex relationship between the variables involved in the hybrid composite systems.

Moreover, the use of deep learning models, specifically the hybrid machine learning model, has been justified by recent studies on modeling complex composite systems, especially those involving nonlinear interactions between the parameters and the mechanical responses (Saha et al.³²).

Optimization using mother optimization algorithm

The MOA receives LSTM-generated feedback on the optimal weight fiber fraction, amount of resin, temperature, and cure time (multi-response), using the overall desirability function as the decision criterion to guide an optimal/correct path forward for these experiments.

D (x) = {(\prod_{j = 1}^{5} d_{j} ({\hat{y}}_{j}))}^{1 / 5}

(9)

This equation (9), Kiratli³³ converts all predicted mechanical outputs into a single performance index, helping MOA determine the globally optimal processing parameters.

To confirm the reliability of the hybrid DQN–LSTM–MOA framework, the optimized parameters were experimentally tested. The agreement between predicted and experimental values is quantified using percentage error.

% E r r o r = 100 % \frac{\hat{y} - y_{\exp}}{y_{\exp}}

(10)

where,

y_{\exp}

denotes the experimental value of the mechanical property. A lower error indicates higher accuracy of the proposed AI-assisted optimization model.

Result and discussion

Experimental result

All results are for various stacking sequence configurations, with thickness held constant to evaluate material distribution properties. As shown in Table 1, best-performing configuration is S2 (BFFFB). This is due to the properties of the lamina, where maximum stress is experienced at the outermost layers. The basalt layers at the outermost points offer maximum strength and stiffness, whereas the flax layers offer maximum ductility and energy absorption. On the other hand, configuration S1 has the lowest performance due to poor bonding, moisture, and void content. Configuration S5 has poor ductility despite its high stiffness. Hybrid composite configurations S3 and S4 show fair performance with poor stress distribution. Similar improvements in mechanical properties with hybridization and efficient interactions between fibers and matrices can be found in recent studies on green composite materials, Saha et al.³⁴

Table 1.

Mechanical properties of hybrid flax–basalt laminates.

Stacking sequence	Tensile strength (MPa)	Flexural strength (MPa)	Impact energy (J)	ILFT (J/m²)	Energy absorption density (J/cm³)
S1: FFFFF	95	130	9.5	950	0.42
S2: BFFFB	150	185	14.2	1580	0.68
S3: FBFBF	125	160	12.1	1320	0.55
S4: FBBBF	138	175	13.0	1450	0.61
S5: BBBBB	142	178	13.4	1500	0.64

It is noteworthy that the experimental results reported in Table 1 are based on variations in stacking sequence at constant processing conditions.

Stress–strain performance of hybrid laminates

To analyze the deformation mechanism and failure mode of such composite structures, the tensile and flexural stress-strain behaviors of hybrid flax-basalt laminae S1–S5 are illustrated in Figure 9(a) and (b). As far as tensile stress-strain graphs are concerned, it can be concluded that the stacking sequence plays an important role in determining tensile properties. It can be seen that the S2 laminate has the highest tensile strength and the most uniform strain. This is because the basalt fibers on the outer surface and the flax fibers in the core provide the best conditions for stress distribution and delayed failure. However, the S1 laminate has lower tensile strength but a higher strain because the flax fibers are ductile. Additionally, the S5 laminate exhibits high stiffness but fails abruptly at a low strain. However, the S3 and S4 laminates have intermediate values, thus proving the effect of the stacking sequence on the tensile properties. From the flexural stress-strain curves, it is evident that the S2 laminate has the best bending properties and the highest tensile stress at a moderate strain. However, the S1 laminate has poor bending properties because the bonding forces between the flax and basalt fibers are low, and the S5 laminate has a brittle fracture because the flax fibers have a poor strain value.

Figure 9.

(a) Tensile and (b) flexural stress–strain curves of hybrid flax–basalt laminates (S1–S5).

Microstructural evolution

In Figure 10, it is clear that there is a significant difference in the microstructural characteristics of the three materials. The basalt material (Figure 10(a)) shows fiber pull-out due to the lack of bonding between the two materials. The flax material (Figure 10(b)), on the other hand, shows the presence of voids resulting from moisture content, which affects the bonding between the two materials. The hybrid material (Figure 10(c)), however, shows improved bonding between the two materials with fewer voids.

Figure 10.

Micrographs of (a) basalt fiber (b) flax fiber and (c) flax-basalt fiber.

The void content was quantitatively assessed through SEM images by defining the pore regions based on contrast differences. After analyzing multiple images, it was confirmed that the porosity was less than 3% in hybrid laminates, indicating better interfacial bonding and resin impregnation between fibers and matrices than in pure flax or basalt laminates. Moreover, the failure mechanisms were well captured in SEM images, including fiber pull-out, matrix cracking, and interfacial debonding, confirming the AI model’s predictions.

The failure initiation mechanisms different across stacking sequences. In S2 (BFFFB), failure was initiated at the surface of the basalt layers, while in S1, failure was initiated at the matrix regions due to poor bonding between the matrices and fibers. In S5, brittle failure was observed at the surface, while in S3 and S4, mixed-mode was observed at the interface between the fibers and matrices. These failure mechanisms confirm that the stacking sequence affects failure initiation and failure propagation, consistent with the better mechanical performance of the laminate compared to other laminates, as discussed in the previous study by Saha and Kumari.⁸

RSM results

RSM experimental design and data’s

Results from the RSM design matrix experiment are presented in Table 2 and were collected after running 30 experiments involving several parameters simultaneously to determine the effect of their combination on the mechanical characteristics of the flax-basalt hybrid composite material. Mechanical characteristics determined in this experiment included Tensile Strength, Flexural Strength, Impact Resistance, ILFT and Energy Absorption Index (EA). Generally, it can be concluded that these mechanical characteristics increased with increasing basalt fiber percentage and that the thermal curing process improved. Of the samples tested across the 30 runs, those with additional fiber reinforcement at 50%, a resin content of 70%, and elevated curing temperatures (100–120°C) responded best. These results indicate that hybrid composites are sensitive to the materials ratio used and to the temperature at which they are manufactured.

Table 2.

RSM experimental results.

Run	A: Fiber weight fraction	B: Epoxy resin content	C: Curing temperature	D: Curing time	Tensile strength	Flexural strength	Impact energy	ILFT	Energy absorption density
Run	%	%	°C	h	MPa	MPa	J	J/m²	J/cm³
1	40	60	100	4	224	285	7.7	454	4.3
2	40	60	100	4	222	300	7.3	458	4.6
3	40	60	80	6	210	270	6	430	4.2
4	30	60	100	2	180	230	4	300	3.2
5	40	70	120	4	240	309	8.4	500	5
6	40	50	100	6	207	260	6	420	4
7	50	60	100	6	260	330	7.8	550	5.5
8	40	60	120	2	225	290	6.93	460	4.7
9	50	70	100	4	275	340	10	600	6
10	50	60	80	4	250	320	8	530	5.2
11	40	60	80	2	205	300	5.6	420	4.1
12	50	60	120	4	265	335	9.3	560	5.6
13	40	60	100	4	226	288	7	455	4.7
14	30	60	80	4	185	240	4.2	310	3.3
15	30	60	100	6	190	245	5.42	320	3.5
16	40	60	100	4	222	289	7	455	4.6
17	40	70	100	2	230	300	8	500	5
18	30	50	100	4	170	220	4	290	3
19	40	60	100	4	221	281	7.5	500	4.9
20	30	70	100	4	203	260	6	400	4
21	40	50	100	2	205	260	6	420	4
22	40	50	80	4	195	255	5	410	3.8
23	30	60	120	4	185	240	5	310	3.3
24	40	60	120	6	235	305	7.4	480	4.8
25	40	70	100	6	245	310	8.5	510	5.1
26	40	50	120	4	210	275	6.32	440	4.3
27	40	70	80	4	230	302	7.19	480	4.8
28	40	60	100	4	222	282	7	452	4.6
29	50	60	100	2	255	325	9	570	5.7
30	50	50	100	4	250	320	8	550	5.5

Table 2 presents the results of 30 RSM experiments in which fiber weight fraction, epoxy content, curing temperature, and curing time were varied. The high mechanical properties in this table are the results of optimized processing parameters rather than the stacking sequence. Therefore, Table 2 illustrates the possibility of performance enhancement via RSM optimization.

ANOVA analysis

ANOVA results, reported in Table 3, the quadratic model created to predict all of these values correspond with all the response variables of interest at that level of significance which means that statistically, they are all significant (p < 0.0001), however based on the analysis, it shows that they do have a strong influence or have a large amount of F-Value (which indicates how the model fits against random variation). Their respective F-Values are as follows: Tensile strength (133.52), Flexural strength (48.03), Impact energy (82.56), ILFT (68.13), EAD (50.72). This is evidence that the input parameters identified do indeed affect the mechanical response of Laminates on a large scale, and lack of fit remains negligible, thus the responses correlate closely with those of the original data, and the models therefore can be relied upon or used as baselines for prediction and optimization.

Table 3.

ANOVA for quadratic model.

Responses	df	F-value	p-value	Model	Lack of fit
Tensile strength	14	133.52	< 0.0001	Significant	Not significant
Flexural strength	14	48.03	< 0.0001	Significant	Not significant
Impact energy	14	82.56	< 0.0001	Significant	Not significant
ILFT	14	68.13	< 0.0001	Significant	Not significant
EAD	14	50.72	< 0.0001	Significant	Not significant

The Fit statistics (Table 4) demonstrate an excellent correlation between the observations of the experimental and predicted data with R² being >0.97 for all outcome variables. The tensile strength produced the highest R² of 0.9920, reflecting model’s accuracy very close to perfect. The adjusted and predicted R² values were very close together, indicating that the model is very strong and unlikely to suffer from overfitting. All outcome variables had low Standard Deviations, indicating that the experimental error is very small and that the results of the testing reflect a high level of precision. Values for Adequate Precision greater than 27 suggest that the model has an appropriate signal-to-noise ratio and can be used to maximize the design space. Collectively, the statistical results supported that RSM is effective for understanding the potential performance and for optimizing the Hybrid Composite Materials through RSM.

Table 4.

Fit statistics.

Responses	Std. Dev. of residuals	R²	Adjusted R²	Predicted R²	Adeq precision
Tensile strength	3.29	0.9920	0.9846	0.9577	45.0530
Flexural strength	6.63	0.9782	0.9578	0.9083	27.2255
Impact energy	0.2432	0.9872	0.9752	0.9546	35.5252
ILFT	14.37	0.9845	0.9701	0.9480	31.0061
EAD	0.1579	0.9793	0.9600	0.9258	27.6089

To confirm that the high R² values (>0.97) are not due to overfitting, cross-validation and analysis of the predicted residuals have been performed. It has been found that the adjusted and predicted R² values are very similar to the R² values for all response variables, confirming the model’s high predictive capability with minimal signs of overfitting. Furthermore, the residual plots show a random pattern around zero with no notable trends, thus confirming the high capability of the RSM model. In addition, plots of actual versus predicted values for all response variables show strong agreement, with points closely clustered along the 45° line.

Residuals versus predicted plots

The residuals versus predicted plots Figure 11(a)–(e) for each mechanical property show the differences between predicted and experimental results. The random distribution of the residuals around the zero axis indicates that the predicted values unbiased and that the errors are homoscedastic. This is supported by the narrow range of residuals, indicating that the RSM-AI model’s predctions agree well with the measured values.

Figure 11.

Residuals versus predicted plots (a) tensile strength, (b) flexural strength, (c) impact energy, (d) ILFT, and (e) EAD. The random scatter around zero confirms unbiased predictions and homoscedasticity.

Predicted versus actual plots

The predicted-versus-actual plots in Figure 12(a)–(e) show a strong correlation between model predictions and experimental results for all mechanical properties. Most data points align closely with the 45° reference line, demonstrating high predictive accuracy of the model for tensile strength, flexural strength, impact energy, ILFT, and EAD. Minor deviations reflect experimental variability and material heterogeneity but remain within acceptable error limits, confirming the reliability and robustness of the developed optimization framework.

Figure 12.

predicted versus actual plots (a) tensile strength, (b) flexural strength, (c) impact energy, (d) ILFT, and (e) EAD.

Normal probability plots of residuals

In the following figures, normal probability plots of residuals for various response variables are depicted Figure 13(a)–(e). In each plot, the values of externally studentized residuals are placed on the X-axis while the expected value of normal probability is taken on the Y-axis. The red straight line shows the perfect normal distribution. Data points that follow the red line indicate that the residuals are normally distributed, an essential requirement of regression models, such as RSM. It can be seen from the plots that the data points closely follow the red line. A few points show minor deviations at the tails, which is common in experimental data, but overall, the residuals appear reasonably normally distributed. This confirms the adequacy of the fitted models and supports the reliability of the predicted results from the RSM analysis.

Figure 13.

Normal probability plots of residuals corresponding to (a) tensile strength, (b) flexural strength, (c) impact energy, (d) ILFT, and (e) EAD. Points close to the diagonal line indicate approximate normality of residuals and validate the regression models.

Sensitivity analysis

Sensitivity analysis was performed using RSM-based perturbation plots (Figures 14 and 15) to evaluate the influence of input parameters on mechanical responses, a method widely used in composite optimization studies, Gillani et al.³⁵ The results show that fiber weight fraction (Factor A) has the most significant effect, as indicated by steep slopes, due to its direct role in load transfer and reinforcement efficiency, leading to improvements in flexural strength, tensile strength, impact energy, ILFT, and EAD. Epoxy resin content (Factor B) has a moderate positive influence, enhancing fiber–matrix adhesion and reducing void formation. In contrast, curing temperature (Factor C) and curing time (Factor D) show minimal impact, with nearly flat trends indicating that once adequate curing is achieved, further variation does not significantly affect performance. Overall, across processing conditions, the results confirm that mechanical properties are primarily governed by reinforcement content and interfacial bonding.

Figure 14.

Perturbation plot of, (a) tensile strength and (b) flexural strength.

Figure 15.

Perturbation plot of, (a) impact energy and (b) ILFT and (c) EAD.

Desirability analysis

Figure 16 presents the results of the multi-response optimization using RSM. The desirability index is 0.995, indicating that the chosen optimal solution meets all required mechanical responses. The optimized parameters were found to be the fiber weight fraction at 49.9%, the content at 69.72%, the curing temperature at 107.93°C, and the curing time at 4.219 h. The responses were the tensile strength (277.54 MPa), flexural strength (343.45 MPa), impact energy (10.15 J), ILFT at (600.34 J/m²), and EAD (6.02 J/cm³).

Figure 16.

RSM optimization plot.

As shown in the experimental validation table (Table 5), all prediction errors are within 5%, indicating strong agreement between predicted and experimental results. The enhancement in mechanical properties can be attributed to improved transfer and interfacial properties. More fibers content will lead to better efficiency, and the resin content will help with better wetting and minimize voids. Therefore, maintaining an appropriate fiber proportion is essential.

Table 5.

Validation of RSM-optimized parameters.

Property	Predicted	Experimental	Error (%)
Tensile strength (MPa)	277.54	270	2.72
Flexural strength (MPa)	343.45	335	2.46
Impact energy (J)	10.15	9.9	2.46
ILFT (J/m²)	600.34	585	2.56
EAD (J/cm³)	6.02	5.85	2.82

Machine learning model comparison and validation

Figure 17(a) illustrates the variation in tensile strength across 30 experimental runs, comparing the experimental values with predictions from multiple machine-learning and statistical models, including, DBN-RFR, RSM, CNN-SVR, and the proposed hybrid model (DQN-LSTM-MOA). The experimental tensile strength fluctuates noticeably from run to run due to variations in process parameters, while all predictive models follow the same overall trend. Among these, the proposed model consistently aligns more closely with the experimental data, capturing both peak and low tensile-strength values more accurately than other methods. In contrast, the CNN-SVR model shows more deviation and larger oscillations, indicating lower prediction stability.

Figure 17.

Comparison for prediction performance (a) tensile and (b) flexural strength.

Figure 17(b) presents the flexural strength results for the same 30 runs, again comparing experimental responses with the outputs of the different prediction models. Similar to the tensile case, the flexural strength exhibits noticeable fluctuations across the runs. The proposed model demonstrates the closest agreement with the experimental measurements, successfully replicating the major variations and peak responses. Other models, such as DBN-RFR and RSM, follow the general pattern but show larger prediction gaps at several runs. The CNN-SVR model shows the highest deviation and instability, with larger spikes and dips than the actual flexural strength values.

Figure 18(a) presents the variation of impact energy across 30 experimental runs and compares the experimental results with predictions from different models. The experimental impact energy fluctuates significantly between runs due to changes in fabrication parameters, and all models attempt to follow this trend. Among them, the proposed model provides the best match to the experimental data, accurately capturing both the high- and low-impact energy values. Other models show moderate deviations, with RSM exhibiting the greatest variation and occasional overshoot, indicating lower prediction reliability.

Figure 18.

Comparison for prediction performance (a) IE (b) ILFT and (c) EAD.

Figure 18(b) shows the experimental values exhibit strong variations, reflecting sensitivity to process factors such as fiber–matrix bonding and laminate integrity. The proposed model consistently aligns more closely with the experimental pattern, effectively tracking the peaks and troughs. Models such as DBN-RFR and CNN-SVR follow the general trend but display noticeable discrepancies at several points. The RSM model again shows larger oscillations and reduced stability, deviating more from the experimental measurements. Figure 18(c) illustrates the variation in energy-absorption density, comparing experimental results with predictions from the different modeling approaches. The trend shows multiple fluctuations across the runs, and the proposed model provides the most accurate and stable prediction, matching the experimental curve with minimal deviation.

The performance of the proposed DQN–LSTM–MOA model was compared with several baseline and advanced ML models, including ANN, SVR, Random Forest, CNN–SVR, DBN–RFR, and RSM. All models were trained and evaluated using the same dataset and identical conditions to secure a fair and unbiased comparison. The results of this comparative analysis are presented in Table 6.

Table 6.

Performance comparison of different predictive models.

Model	MSE	MAE	RMSE	Prediction accuracy (%)
ANN	18.52	3.74	4.30	91.2
SVR	15.36	3.21	3.92	92.8
Random forest (RF)	12.84	2.95	3.58	94.1
CNN–SVR	10.27	2.61	3.20	95.6
DBN–RFR	8.95	2.34	2.99	96.3
Proposed (DQN–LSTM–MOA)	3.95	1.32	1.98	99.2

To further validate the reliability of the model comparison, a statistical analysis was conducted using one-way ANOVA and t-tests on the prediction errors (MSE values) at a 95% confidence level. From the results, it was evident that the proposed DQN-LSTM-MOA model showed a statistically significant improvement over the other models (p < 0.05), thereby confirming that the proposed model’s better performance was not due to random variation.

From the comparative analysis, it was evident that the proposed DQN-LSTM-MOA model outperforms conventional models such as ANN, SVR, and RF in terms of prediction accuracy and error minimization. Although ANN and SVR have high errors due to their inability to effectively handle complex nonlinear relationships, the RF model shows moderate improvement but lacks the capability to handle interdependencies. Other hybrid models, such as CNN-SVR and DBN-RFR, also have better performance due to the deep features extracted in the context of the proposed framework. However, the proposed DQN-LSTM-MOA model achieves the best prediction accuracy of 99.2%, thereby demonstrating its robustness in predicting the mechanical behavior of hybrid flax-basalt composites.

Comparison with previous hybrid composite studies

A comparison of relevant composite studies on hybrid material systems is presented in Table 7. Although previous works by Sharma et al.³⁶ and Fadıl et al.³⁷ have discussed basalt-based systems, other studies have discussed different types of hybrid composites, like flax E-glass Koppula et al.,³⁸ flax carbon Wang et al.,³⁹ and synthetic-only hybrids Bayar and Babaoğlu.⁴⁰ Lack of studies has discussed the optimization of flax-basalt composites. In the present study, AI techniques have been successfully implemented for the first time to optimize flax-basalt composites, demonstrating improvement in tensile, flexural, impact, ILFT, and energy absorption properties compared to individual components and other hybrid composites.

Table 7.

Comparative analysis of previous hybrid composite studies and the present flax–basalt study.

Study	Objective	Materials used	Technique/fabrication	Outcomes	Limitations
Sharma et al.³⁶	Impact of basalt filler on mechanical properties	Basalt fiber reinforced epoxy composites	Fabrication unspecified (likely hand lay-up/compression), mechanical testing (tensile, flexural)	Basalt reinforcement improves mechanical properties relative to neat resin; property trends with filler loading reported	Does not involve hybridization with natural fibers, no hybrid flax context
Fadıl et al.³⁷	Influence of stacking sequence on mechanical and impact behavior	Basalt, aramid, balsa fiber layers in epoxy	Sandwich composite fabrication (likely hand lay-up/press), mechanical and impact testing	Stacking sequence affects bending and impact behavior; sustainable composite design insights	Not directly flax hybrid; focuses on basalt/aramid/balsa system, limited relevance to flax–basalt
Koppula et al.³⁸	Effect of flax/E-glass layer arrangement on composite performance	E-glass fiber & flax fiber reinforced polymer composites	Layered hybrid composite fabrication (method not detailed) and testing	Stacking arrangement influences tensile/flexural behavior; combining natural and synthetic fibers yields balanced performance	Not flax–basalt hybrid; uses E-glass instead of basalt
Wang et al.³⁹	Mechanical performance and aging behavior of flax–carbon hybrids	Flax + carbon fiber epoxy hybrid composite	Conventional lay-up, low-velocity impact tests, aging treatments	Hybridization improves impact resistance and aging response compared to pure flax; tensile/flexural performance trends reported	Not flax–basalt hybrid; different synthetic reinforcement fiber
Bayar and Babaoğlu⁴⁰	Punch shear behavior and machine learning prediction	Basalt + glass fiber reinforced composites	Composite fabrication, shear testing, machine learning analysis	Demonstrates effect of hybrid basalt/glass system on shear performance and ML prediction capability	Not natural fiber hybrid; relevance is limited to synthetic-only hybrid composites
Present study	Stacking sequence and process optimization using RSM + AI	Flax and basalt fiber epoxy composites	Hand lay-up + compression curing; mechanical testing; RSM and hybrid AI framework	Optimal BFFFB configuration achieved high tensile, flexural, impact, ILFT, and EAD; robust predictive framework	Limited stacking sequences; only short-term mechanical tests; no long-term durability

Conclusion

• The study demonstrates that hybridization of flax and basalt fibers significantly enhances the mechanical performance of composite laminates.

• The S2 (B–F–F–F–B) configuration exhibited the best overall performance, achieving 150 MPa tensile strength, 185 MPa flexural strength, 14.2 J impact energy, 1580 J/m² ILFT, and 0.68 J/cm³ EAD.

• Improved performance is attributed to effective stress distribution, reduced void content, and enhanced fiber–matrix interfacial bonding, as confirmed by SEM analysis.

• The RSM-based optimization and DQN–LSTM–MOA model demonstrated high prediction accuracy and effective multi-response optimization capability.

• Overall, the proposed integrated experimental–computational framework provides a reliable approach for designing high-performance hybrid flax–basalt composites.

Limitations and future scope

The study is limited by the use of hand lay-up fabrication, which may introduce variability in fiber wetting and void content. Additionally, only a limited number of stacking sequences were investigated, and long-term durability aspects such as fatigue, moisture, and thermal aging were not considered. Future work should focus on advanced manufacturing techniques, such as vacuum infusion and expanded stacking configurations, as well as durability analysis under environmental conditions. Incorporating larger datasets and advanced optimization algorithms can further improve the predictive capability and practical applicability of the proposed model.

Footnotes

ORCID iD

Sabbah Ataya

Ethical considerations

Ethical approval is not required for this study, as it does not involve human or animal participants. Therefore, ethical approval is not applicable.

Consent to participate

Formal consent is not essential for this particular type of research.

Author contributions

Every author has made an equal contribution to the work. All authors reviewed the manuscript.

Funding

The authors received no financial support for the research, authorship, and/or publication of this article.

Declaration of conflicting interests

The authors declared no potential conflicts of interest with respect to the research, authorship, and/or publication of this article.

Compliance with ethical standards

Potential Conflict of Interest Disclosure: None of the authors have any potential conflicts of interest.

Replication of results

The results reported in this study can be reproduced using the methods, data, and parameters described in the manuscript.

Code availability statement

The computational code associated with this research is currently kept confidential due to institutional restrictions and author agreements.

Data Availability Statement

The majority of the data supporting the findings of this study are presented within the manuscript in the form of tables and figures. Additional raw data may be made available from the corresponding author upon reasonable request.*

References

Puttegowda

. Eco-friendly composites: exploring the potential of natural fiber reinforcement. Discov Appl Sci 2025; 7: 1–24. https://doi.org/10.1007/s42452-025-06981-8

Lakshmaiya

Udhayakumar

Sathiyamurthy

, et al. Multi-functional natural fiber composites using flaxseed and cotton: tailoring acoustic, mechanical, and thermal properties for eco-friendly applications. Discov Appl Sci 2025; 7: 906. https://doi.org/10.1007/s42452-025-07345-y

Mehaba

Toubal

Belouadah

. Advantages of in-situ impregnation 3D printing in continuous flax/PLA biocomposites for optimized mechanical properties in structural applications. J Mater Res Technol 2025; 39: 355–367.

Gao

Jiang

Lei

, et al. Comparisons between basalt for continuous fiber and ordinary basalt. Ceram Int 2025; 51: 132–146. https://doi.org/10.1016/j.ceramint.2024.10.383

Doğan

Yapici

Gemi

, et al. Investigation of the stacking sequence and cutting parameters effect on hole morphology in hybrid FML composites. Composites, Part B Eng 2025; 300: 112464. https://doi.org/10.1016/j.compositesb.2025.112464

Hasegawa

Yamazaki

Miura

, et al. Comprehensive experimental and molecular-dynamics studies of a secondary bonded epoxy–epoxy joint: towards dominant adhesion mechanism ensuring integrity of structural composite joints. Int J Adhesion Adhes 2025; 145: 104229. https://doi.org/10.1016/j.ijadhadh.2025.104229

Husain

Khan

Mourad

AHI

. Stacking sequence effects on flexural stiffness, failure progression, and energy absorption in asymmetric CFRP laminates. Results Eng 2025; 27: 105863. https://doi.org/10.1016/j.rineng.2025.105863

Saha

Kumar

Mahakur

, et al. Development and machine learning based prediction of carbon/hemp/ramie sandwich composite as sustainable and ecofriendly solution for automobile application. Mater Today Commun 2024; 41: 110817. https://doi.org/10.1016/j.mtcomm.2024.110817

Kumar

Saha

Zindani

. Agro-waste-based polymeric composite laminates for aerospace cabin interior and identification of their optimal configuration. Biomass Convers Biorefinery 2024; 14(24): 31907–31923. https://doi.org/10.1007/s13399-023-04914-2

10.

Kumar

Mahakur

Mishra

, et al. Performances and decision framework of CNT-infused bio-based hybrid composites for lightweight smart structures. Sci Rep 2026; 16: 8531.

11.

Kumar

Bhowmik

Zindani

. Effect of stacking sequence and thickness variation on the thermo-mechanical properties of flax-kenaf laminated biocomposites and prediction of the optimal configuration using a decision-making framework. Int Polym Process 2023; 38(3): 404–423. https://doi.org/10.1515/ipp-2023-4341

12.

Sharma

Bharti

. Optimization of resin printing parameters for improved surface roughness using metaheuristic algorithms: a multifaceted approach. J Mater Eng Perform 2025; 34: 8525–8539. https://doi.org/10.1007/s11665-024-10296-x

13.

Sekar

Narayanan

. Enhancing mechanical and thermal properties of isophthalic polyester resin composites reinforced with graphene oxide and nanosilica using RSM and ANN. Polym Test 2025; 149: 108876. https://doi.org/10.1016/j.polymertesting.2025.108876

14.

Hamzat

Murad

Adediran

, et al. Fiber-reinforced composites for aerospace, energy, and marine applications: an insight into failure mechanisms under chemical, thermal, oxidative, and mechanical load conditions. Adv Compos Hybrid Mater 2025; 8: 152. https://doi.org/10.1007/s42114-024-01192-y

15.

Azad

Prabhakar

Kim

, et al. Deep learning-based autonomous morphological fracture analysis of fiber-reinforced composites. Eng Fail Anal 2025; 170: 109292. https://doi.org/10.1016/j.engfailanal.2025.109292

16.

Irshad

Zafar

Pathak

, et al. Microwave-assisted compression moulding of Flax/Bio-PBS composites using an innovative spring-loaded fixture for sustainable manufacturing. Manuf Lett 2025; 47: 6.

17.

Yan

. Lightweight composite materials in automotive engineering: state-of-the-art and future trends. Alex Eng J 2025; 118: 1–10. https://doi.org/10.1016/j.aej.2024.12.002

18.

Yazik

MHM

. Modelling of 3D-printed natural fibre-reinforced polymer biocomposite damage progression. In: Damage analysis of natural fiber-reinforced polymer biocomposites. Woodhead Publishing, 2026, pp. 421–454.

19.

Chellan

Mondal

MIH

Joseph

. Basalt fibres and their applications in the automotive industry. In: Technical organic and inorganic fibres from natural resources. Elsevier Science Ltd, 2025, pp. 599–623.

20.

Raja

Devarajan

. Study on the mechanical and thermal properties of basalt fiber-reinforced lead oxide nanoparticle polymer composite for advanced energy storage applications. J Energy Storage 2025; 108: 115079. https://doi.org/10.1016/j.est.2024.115079

21.

Dogar

MMA

Mubashar

Masud

, et al. Influence of fibre stacking sequence on impact resistance and residual strength in flax/basalt hybrid laminates. Appl Compos Mater 2025; 32: 681–702. https://doi.org/10.1007/s10443-024-10294-1

22.

Shi

Zhang

, et al. Effects of the internal pressure and stacking sequence on the transverse impact behavior of hybrid uniaxial-biaxial braided composite tubes. Eng Fract Mech 2025; 313: 110667. https://doi.org/10.1016/j.engfracmech.2024.110667

23.

Taheri

Ghiaskar

Chowdhury

. Flexural and shear behavior of basalt-Recyclamine composites: a comparative study with conventional epoxy systems. Appl Compos Mater 2025; 32: 1–30. https://doi.org/10.1007/s10443-025-10370-0

24.

Thanikodi

Rathinasamy

Solairaju

. Developing a model to predict and optimize the flexural and impact properties of jute/kenaf fiber nano-composite using response surface methodology. Int J Adv Manuf Technol 2025; 136: 195–209. https://doi.org/10.1007/s00170-024-13975-0

25.

Rothenhäusler

Albuquerque

Sticher

, et al. Application of convolutional neural networks and ensemble methods in the fiber volume content analysis of natural fiber composites. Mach Learn Appl 2025; 19: 100609. https://doi.org/10.1016/j.mlwa.2024.100609

26.

Chang

Wang

, et al. High-precision identification and classification of alloy fatigue microcracks through deep learning and in-situ SEM. Comput Mater Sci 2025; 252: 113795. https://doi.org/10.1016/j.commatsci.2025.113795

27.

Wang

Huang

Guo

. A two-stage deep learning framework for predicting crack patterns and mechanical properties of unidirectional composites with void defects. Compos Sci Technol 2025; 271: 111357. https://doi.org/10.1016/j.compscitech.2025.111357

28.

Boobalan

Sathish

Saravanan

. Synergistic effects of SiO2 and MWCNTs in glass/basalt-reinforced composites via BBD-RSM. Chem Pap 2025; 80: 1–20. https://doi.org/10.1007/s11696-025-04398-6

29.

Shonkwiler

Fenrich

, et al. Deep reinforcement learning for stacking sequence optimization of composite laminates. Manuf Lett 2023; 35: 1203–1213. https://doi.org/10.1016/j.mfglet.2023.08.133

30.

Yang

Zeng

, et al. Feature selection optimization algorithm based on evolutionary Q-learning. Inf Sci 2025; 719: 122441. https://doi.org/10.1016/j.ins.2025.122441

31.

Deng

Zhi

, et al. Spatio-temporal prediction of curing-induced deformation for composite structures using a hybrid CNN-LSTM and finite element approach. Compos Sci Technol 2025; 268: 111225. https://doi.org/10.1016/j.compscitech.2025.111225

32.

Saha

Kulkarni

Kumari

. Development of bambusa tulda fiber-micro particle reinforced hybrid green composite: a sustainable solution for tomorrow’s challenges in construction and building engineering. Constr Build Mater 2024; 441: 137486. conceptually supports hybrid composite development (discussion context). https://doi.org/10.1016/j.conbuildmat.2024.137486

33.

Kiratli

. Compressive characteristics of E-glass/epoxy composites modified with multi-walled carbon nanotubes: multi-objective optimization using response surface methodology combined with different algorithms. Eur Phys J Plus 2025; 140: 666. https://doi.org/10.1140/epjp/s13360-025-06598-1

34.

Saha

Kumari

. Effect of alkaline treatment on physical, structural, mechanical and thermal properties of Bambusa tulda (northeast Indian species) based sustainable green composites. Polym Compos 2023; 44(4): 2449–2473. https://doi.org/10.1002/pc.27256

35.

Gillani

Khan

Shah

. Sensitivity analysis of reinforced aluminum based metal matrix composites. Materials 2022; 15(12): 4225. https://doi.org/10.3390/ma15124225

36.

Sharma

Meena

Kumar

. Effect of filler percentage on physical and mechanical characteristics of basalt fiber reinforced epoxy based composites. Mater Today Proc 2020; 26: 2506–2510. https://doi.org/10.1016/j.matpr.2020.02.533

37.

Fadıl

Erkliğ

Bulut

. The influence of fiber stacking sequence on the mechanical and impact performance of sustainable epoxy/basalt/aramid‐balsa sandwich composites. Polym Adv Technol 2025; 36(12): e70465. https://doi.org/10.1002/pat.70465

38.

Koppula

Karachi

Kumar

, et al. Investigation into the mechanical characteristics of natural fiber-reinforced polymer composites: effects of flax and e-glass reinforcement and stacking configuration. Mater Today Proc 2024; 115: 82–88. https://doi.org/10.1016/j.matpr.2023.07.020

39.

Wang

Xian

. Mechanical, low-velocity impact, and hydrothermal aging properties of flax/carbon hybrid composite plates. Polym Test 2020; 90: 106759. https://doi.org/10.1016/j.polymertesting.2020.106759

40.

Bayar

Babaoğlu

. Mechanical performance and machine learning-based prediction of punch shear strength in basalt/glass fiber reinforced composites. Compos Appl Sci Manuf 2026; 204: 109620. https://doi.org/10.1016/j.compositesa.2026.109620

Experimental and AI-assisted optimization of mechanical performance of hybrid flax–basalt composite laminates

Abstract

Keywords

Introduction

Literature study

Research gap

Objectives

Proposed methodology

Material selection

Fabrication of composites

Stacking sequence and specimen preparation

Selection of parameters

Mechanical testing

Tensile strength

Flexural strength

Impact energy

Interlaminar fracture toughness

Energy absorption density

Microstructural analysis (SEM)

Response surface methodology analysis

Application of AI

Deep Q network for feature selection

LSTM model for mechanical property prediction

Optimization using mother optimization algorithm

Result and discussion

Experimental result

Stress–strain performance of hybrid laminates

Microstructural evolution

RSM results

RSM experimental design and data’s

ANOVA analysis

Residuals versus predicted plots

Predicted versus actual plots

Normal probability plots of residuals

Sensitivity analysis

Desirability analysis

Machine learning model comparison and validation

Comparison with previous hybrid composite studies

Conclusion

Limitations and future scope

Footnotes

ORCID iD

Ethical considerations

Consent to participate

Author contributions

Funding

Declaration of conflicting interests

Compliance with ethical standards

Replication of results

Code availability statement

Data Availability Statement

References