Artificial Neural Network Approach for Assessing mechanical properties and Impact Performance of Natural-Fiber Composites Exposed to UV Radiation

1. Introduction

Over the past two decades, biocomposites of short natural fiber thermoplastics (SNFT) obtained via injection molding (SNFT) have garnered considerable attention [1,2,3]. They are widely used in various industries, including construction, automotive, and packaging [4,5]. SNFT biocomposites offer high-specific mechanical properties and are eco-friendly [6,7,8]. Nevertheless, one major challenge in developing natural-fiber composites is their response to environmental factors, such as moisture, ultraviolet (UV) radiation, and heat, which can degrade their mechanical performance and limit their applicability, especially for outdoor applications [9,10,11,12,13].

Researchers have extensively investigated the aging of SNFTs via exposure to real climates (natural aging) [14,15,16,17] or simulated conditions in laboratory chambers (artificial aging) [7,18,19,20,21]. When SNFTs are exposed to UV rays and/or high temperatures, photooxidation reactions occur in the lignin of natural fibers, and moisture accelerating these reactions [7,22]. The velocity of photo-oxidation reactions in biocomposites depends on the chemical composition of the natural fibers. Peng et al. [20] investigated the impact of the chemical composition of natural fibers on the performance degradation of polypropylene reinforced with short wood fibers. They revealed that the lignin in natural fibers can act as a UV absorber, mitigating the degradation of the mechanical properties. Similarly, Nasri et al. [7] reported that biocomposites reinforced with wood pine fibers, which have high lignin content, exhibited less significant degradation in mechanical properties than flax fibers. Indeed, under the influence of UV radiation, the biocomposites undergo photo-oxidation reactions at the lignin level of the natural fibers, resulting in chemical changes, increased roughness, and degradation of mechanical properties. The photo-oxidation reactions strongly depend on the chemical composition of the natural fibers, with biocomposites containing flax fibers showing more degradation due to their higher cellulose content in their structures. More detailed information regarding the degradation of SNFT biocomposites under UV aging can be found in previous reviews [22,23].

Despite numerous and diverse research papers on natural-fiber-reinforced composites, the range of applications involving this type of material in engineering design is still limited. Notably, biocomposite structures designed for outdoor applications are vulnerable to low-speed impacts, and the strength of these materials becomes even more critical when exposed to the combined effects of UV radiation and/or moisture. It is well known that aging tests require numerous samples and are time-consuming and expensive. Aging (natural or artificial) requires exposure to various environmental conditions for months or even years [18,19]. Therefore, developing efficient prediction models to assess the mechanical resistance of biocomposites and their evolution based on exposure time to ultraviolet radiation and/or moisture is of great importance.

Artificial neural networks (ANNs) have emerged as highly effective methods for linear and nonlinear predictions concerning the mechanical properties of composite materials, considering their various constituents (such as fibers, matrix, and particles). An ANN creates predictive regression models based on experimental data [24]. Numerous researchers have successfully utilized ANN algorithms to predict the mechanical properties of biocomposite materials. For instance, Stamopoulos et al. [25] developed two ANN models trained using a multiscale methodology to predict the mechanical properties of matrix-dominated composites (for example, transverse strength, transverse stiffness, bending strength, flexural modulus, and short-beam strength), demonstrating consistency with experimental results. Yang et al. [26] developed an ANN model to predict the residual strength of carbon fiber-reinforced carbon (CFRC) following low-velocity impacts. The model, which was trained using finite-element analysis results, accurately established a nonlinear relationship between the impact parameters and residual strength, thus reducing computational costs and time compared to traditional methods. Fan et al. [27] trained an ANN-based model using limited test data to predict the tensile strengths of composite laminates with open holes. Altabey and Noori [28] developed a neural-network model for predicting the fatigue life of CFRC considering factors such as fatigue stress ratio, fiber orientation, materials, and loading conditions. Mohsin et al. [29] developed prediction models using ANN algorithms to predict the compressive strength and dry thermal conductivity of hemp-based biocomposites. Experimental records were used to train the models and demonstrate their accuracy and feasibility. These models offered significant time savings compared with laboratory tests. Zhang et al. [24] provided an in-depth overview of the use of ANNs for the mechanical modeling of composite materials.

According to the literature, the use of ANN algorithms represents a robust approach for modeling complex nonlinear connections between inputs and outputs when obtaining a precise analytical expression is challenging. The strength of ANNs lies in their ability to analyze complex data, identify patterns and relationships, and accurately predict the mechanical behavior of biocomposites under various conditions. However, models should be based on easily measurable physical indicators. The data obtained from cut specimens of large components are primarily limited to laboratory use and are not applicable for in-service detection. Ideally, these data should be collected directly from a real structure without causing its destruction or alteration of its functionality. To achieve this, we have developed an efficient ANN-based model for predicting the low-velocity impact properties of aged biocomposites, specifically polypropylene reinforced with short flax or pine fibers. The novelty of our approach lies in predicting both mechanical and low-velocity impact properties through nondestructive testing, which involves measuring parameters such as mass, hardness, roughness, and resonant frequencies, providing the ANN model with its originality. By better understanding the behavior of these materials, it becomes possible to integrate them more effectively into various applications, thereby contributing to the development of environmentally friendly design solutions. In this context, tensile and impact samples were exposed to two accelerated aging programs: UV aging in dry and humid environments. Subsequently, we evaluated the changes in the properties of the biocomposites, and the proposed ANN model was finally validated with experimental results.

2. Methodology

This section summarizes the approach adopted to predict the long-term mechanical properties and low-velocity impact properties of the PP30-F and PP30-P biocomposites subjected to accelerated weathering. The collected data were used to develop an ANN model. Two aging conditions were used in this study: aging by UV rays with or without moisture. The input parameters of the model were the mass (M), hardness (H), mean roughness (Ra), and natural frequencies (bending and torsion modes, respectively f_b and f_t), and the output parameters were the maximum impact load and absorbed energy. After the ANN model was validated, such models were used to predict the tensile strength (R) and the impact performance (F and E). Figure 1 presents the methodology of this study.

Artificial neural network

An ANN is a network composed of perceptron cells linked by weighted interactions. Figure 2 presents the architecture of the ANN used in this study, which was divided into three parts: input, hidden, and output layers. The model was developed using the ANN toolbox implemented in MATLAB R2020 software. Before the ANNs were trained, the data were normalized to the range of –1 to +1 using the Z-score technique. This was done to ensure consistency with the transfer function used in the hidden and output layers. Then, the ANN models were trained, tested, and validated using a backpropagation algorithm. Subsequently, a set of data (input data and the corresponding output values) was applied to the developed ANN model to calibrate the weighting factors. A test set was used to select the best ANN through the calculation of the linear correlation coefficient (R²) and mean squared error (MSE).

A database was constructed on the basis of the experimental results. It included 70 sets of data used to validate the ANN model for predicting the impact performance (maximum load, absorbed energy and tensile strength that is, F, E and R, respectively) of the biocomposites. Initially, six input parameters were selected: the hardness (H), mean roughness (Ra), and natural frequencies (torsion and bending modal, i.e., f_t and f_b, respectively).

3. Experimentation

3.1. Materials and manufacturing

The materials used in this study were biocomposites of polypropylene reinforced with either 30 wt.% flax fiber (FF30P233-00) or 30 wt.% pine wood fiber (WP30P233-00) purchased from Rhetech Inc. (Whitmore Lake, MI, USA). Flax fibers differ from pine fibers with regard to their chemical and geometric compositions. Flax fibers are rich in cellulose and have a higher length-to-diameter (L/d) ratio than pine fibers [7], whereas pine fibers are richer in lignin than flax fibers. This difference in fiber composition may influenced the mechanical and aging properties of the studied short-fiber composites.

A 100-ton press (Zhafir Zeres series ZE900/210, Haitian Inc.) was employed to perform injection molding of the impact samples, in accordance with the ASTM D-2856 standard. The injection temperature was maintained at 200 °C. To prevent the occurrence of microvoids and porosity in the samples post-injection, the biocomposite granules were dried at 80 °C for 2 h prior to injection.

3.2. Aging conditions

Two environmental conditions were considered in this study:

-

Condition 1 (UV without moisture): The samples were subjected to UV aging using UVA-340 fluorescent lamps emitting irradiance at a wavelength of 340 nm. Aging was performed using a QUV/SE aging apparatus (Q-Lab Co., USA). Over a period of two months, the samples were exposed to 8 h of UV radiation each day at an irradiance of 1.55 W/m², and the temperature was maintained at 60 °C.

-

Condition 2 (UV with moisture): The samples were subjected to the same conditions as Condition 1 for two months. However, after each UV exposure at 60 °C, the samples were subjected to 4 h of water condensation at 50 °C.

The aging conditions were conducted in accordance with ASTM G53, standard practice for artificial UV-aging of non-metallic materials.

3.3. Experimental setup and procedure

3.3.1. Mass Measurement:

Specimen mass was measured using a precision electronic scale with accuracy up to 10^-3 g, providing the mass (M) in kilograms (kg).

3.3.2. Roughness Measurement:

Surface roughness was evaluated via a 3D laser confocal microscope (Keyence, Canada). The roughness parameter Ra was determined in millimeters (mm).

3.3.3. IET (Impulse excitation technic):

Impact Echo Testing (IET) was conducted following ASTM E-1876-09 on an IMCE machine. Signal processing, facilitated by Resonant Frequency and Damping Analyzer (RFDA) software, yielded resonance frequency (f) results in Hertz (Hz).

3.3.4. Hardness Test:

Hardness assessments adhered to ASTM E329 standards utilizing an HM-100 Economical manual-type hardness testing machine (810-124). The resulting hardness values (H) are reported in HRC (Rockwell C scale).

3.3.5. Tensile test

We utilized an Instron model LM-U150 electromechanical testing apparatus, which was equipped with a 10 KN load cell, to rigorously examine the tensile strength properties of materials. Our experimental protocol adhered to the ASTM D638 standard for such tests. The tests were performed at a controlled displacement rate of 1 mm/min.

3.3.6. Drop-Weight Impact Test:

The ASTM D-5628 standard guided drop-weight impact tests on an Instron machine (Model CEAST 9350) equipped with a 22-kN load cell. Employing an initial impact energy of 5 Joules and a 5.4 kg impactor, impact force (F) values were measured in kilogram meters per second squared (kg⋅m⋅s^-2), while impact energy (E) values were quantified in Joules (J).

The details of each test (device and measurement method) are presented in Appendix 1. The equipment and samples used in this study are presented in Appendix 2.

4. Results and discussion

4.1. Experimental results

The evaluation results are presented in Table 1. Both biocomposites exhibited changes in their physical and mechanical properties over time. These changes were more significant under the second treatment (UV irradiation with moisture), possibly because photooxidation reactions are the main contributors to SNFT property degradation [7] and that moisture accelerates these reactions [22]. Consequently, a higher degree of degradation was observed under the second condition (UV irradiation with moisture) for both materials.

4.1.1. Mass

The masses of the samples decreased with an increase in the exposure time for both conditions (UV with and without moisture). These mass losses were mainly due to the degradation of natural fibers by photo-oxidation reactions [22].

4.1.2. Hardness

Prior to aging, the PP30-F biocomposites exhibited higher hardness than the PP30-P biocomposites; the corresponding hardness values were H = 9.12 and 8.72 HRC, respectively. After aging, a reduction in the hardness was observed for both materials, as shown in Figure 3. However, after 1440 h of UV exposure under dry conditions, the measured hardness of the PP30-F biocomposite was 6.21 HRC, while that of the PP30-P biocomposite was 6.44 HRC. Similarly, under humid conditions, the corresponding values were 5.54 and 5.71 HRC, respectively. This reduction is mainly attributed to the scission of the polymer chains, which led to the formation of surface cracks and embrittlement of the material. The number of chain scissions increased with the exposure time, resulting in shorter polymer chains and degradation of all the mechanical properties. PP30-P exhibited a smaller hardness loss than PP30-F, which can be explained by the antioxidant effect of the lignin in pine fibers [29].

4.1.3. Roughness

The measured results for the surface roughness presented in Figure 4 indicate that both biocomposites initially had smooth and intact surfaces with a roughness of approximately 2.5 μm. However, after UV exposure in dry or moist conditions, both biocomposites exhibited rough surfaces. The increase in the surface roughness of the UV-exposed biocomposites is attributed to polymer-chain scission resulting from photo-oxidation. Polymer-chain scission is also responsible for the formation of microcracks on the surface of SNFT biocomposites during aging [22]. The PP30-P biocomposite exhibited fewer large cracks than the PP30-F biocomposite under both conditions. The lignin content of pinewood fibers is at least seven times higher than that of flax fibers. This suggests that the presence of lignin in pinewood had an antioxidant effect, delaying surface degradation, as previously reported [7]. After 1440 h of UV exposure in dry conditions, the average roughness (Ra) measured for the PP30-F biocomposite was 19.86 µm, and that for the PP30-P biocomposite was 10.02 µm. Similarly, the corresponding values for humid conditions were 32.41 and 16.73 µm, respectively.

4.1.4. Natural frequencies

Figure 5 shows the natural frequencies of the unaged and aged biocomposite samples at 1440 h. The results indicate reductions in the natural frequencies. The UV irradiation was reduced the resonant frequencies of the composites by degrading the polymer materials in both the matrix and the reinforcing fibers. In the case of UV aging combined with moisture, the resonant-frequency reduction was accelerated. Photooxidation causes the splitting of the polymer chains of the thermoplastic matrix, resulting in surface microcracks. When composites are exposed to moisture, water can be absorbed by natural fibers or by the interfaces between the fibers and matrix. This water absorption causes swelling of the fibers and matrix, which affects the fiber–matrix adhesion and the mechanical properties of the composite, reducing the resonant frequency.

4.1.5. Tensile test

Figure 6 shows the tensile strength of both aged and unaged biocomposite samples after 1440 hours. The results reveal a decrease in the tensile strength following aging. This reduction in tensile strength is attributed to the effects of UV radiation on the polymeric materials within the natural fibers, particularly lignin, and the polypropylene matrix. This may be explained by the fact that photo-oxidation induces superficial microcracks in the biocomposites, which act as stress concentration points within the samples, consequently leading to a degradation in mechanical properties such as tensile strength. When UV exposure is combined with moisture, the degradation process is amplified. Indeed, moisture accelerates the photo-oxidation reactions, further promoting degradation.

4.1.6. Low-velocity impact properties

The low-velocity impact properties, such as the impact resistance and absorbed energy of aged biocomposites were studied to evaluate the effects of aging on their behavior under impact loads. The results (Figure 7 and Figure 8) indicated significant changes in these properties with an increase in the exposure time, which are mainly attributed to the formation of microsurface cracks during the aging of the biocomposite samples. These microcracks acted as initiation points for damage under an impact load, amplifying the local stresses within the biocomposites. The amplification of local stresses resulting from surface microcracks reduced the tensile strength and absorbed energy of the aged biocomposites.

The PP30-P biocomposites exhibited less degradation than the PP30-F biocomposites with regard to their impact properties. This difference is attributed to the antioxidant effect of the lignin present in wood pine fibers. Lignin suppresses crack propagation during impact by delaying the initiation and evolution of surface microcracks.

Prior to aging, the PP30-F biocomposites exhibited higher strength than the PP30-P biocomposites, with a maximum load (F) of 951.72 N and energy absorption (E) of 3.86 J, compared with a maximum load of 862.28 N and energy absorption of 4.21 J for PP30-P. However, after aging, the PP30-F biocomposites exhibited more significant degradation in both properties. After 1440 h of exposure, the PP30-F biocomposites exhibited a maximum load of 798.84 N and energy absorption of 3.25 J in the first condition (UV without moisture) and a maximum load of 867.18 N and energy absorption of 3.78 J in the second condition (UV with moisture). Meanwhile, the PP30-P biocomposites exhibited a maximum load of 760.82 N and energy absorption of 3.11 J in the first condition and a maximum load of 749.94 N and energy absorption of 3.68 J in the second condition.

Table 1 presents the experimental results for the physical and mechanical properties.

4.2. ANN approach

4.2.1. ANN model validation

The performance of the ANN model was evaluated according to the convergence of the MSE. The best validation performance was observed after four epochs (MSE = 22). Figure 9 and Figure 10 present plots of the linear regression coefficients. As shown, the model fit the data well; the global correlation coefficients (R) in the case of PP30-F were 0.999 for training, 0.997 for testing, and 0.999 for validation, and in the case of PP30-P, they were 0.997, 0.999, and 0.999, respectively. In addition, the training, testing, and validation stages of the model for prediction were positive. This suggests that the model learned effectively, generalized well to new data, and didn’t overfit the training dataset. It indicates a promising performance and adds credibility to the model’s ability to make accurate predictions.

4.2.2. ANN prediction

Figure 11, Figure 12 and Figure 13 present comparisons between the prediction results of the ANN model and the experimentally obtained results. Overall, the proposed model provided accurate results. The ANN results agreed well with the experimental data concerning the impact load, energy absorbed and tensile strength, with maximum errors of 3, 5% and 6%, respectively. However, it is important to recognize that despite the overall accuracy of our predictions, slight differences between the results of the ANN algorithm and the experimental data may appear. These deviations could arise from several factors, such as natural variations in material properties, environmental conditions during experimental testing, or even potential inaccuracies in the parameters of the algorithm itself. However, despite these minor deviations, our model is capable of accurately predicting the mechanical impact performance of aged SNFT biocomposites.

5. Conclusion

In external applications, biocomposites are vulnerable to environmental aging and low-velocity impacts. Aging reduces the resistance of biocomposites to low-impact collisions. Thus, for avoiding damage, it is important to predict the impact resistance of aged biocomposites. In this study, a novel ANN prediction model was developed to predict the durability of aged biocomposites with regard to their mechanical and impact properties. This network was applied to two biocomposite materials (polypropylene reinforced with 30 wt.% flax or pine fibers) under two environmental conditions (UV aging in a dry or moist environment). Unaged and aged biocomposites were evaluated via nondestructive tests, such as mass, hardness, roughness, and IET tests, as well as destructive tests (tensile tests and low-velocity impact tests) to evaluate these properties with respect to the exposure time. The results indicated that all the mechanical properties were degraded after aging. The degradation was more significant under Condition 2 (UV with moisture) than under Condition 1 (UV without moisture). Moreover, PP30-F exhibited a higher degree of degradation than PP30-P under both conditions. Subsequently, the impact properties of the two composite materials (tensile strength and absorbed impact energy) were predicted using the developed ANN algorithm, for which nondestructive test results were used as inputs. The proposed ANN model proved to be a reliable prediction tool, as indicated by strong agreement between the experimental and predicted results. The correlation coefficient R was 0.999 for the two biocomposites.