Experimental Characterization and a Machine Learning Framework for FDM-Fabricated Biocomposite Lattice Structures

Highlights

What are the main findings?

• The manufacturing parameters significantly influence the mechanical performance of wood–PLA composite lattice structures. Specifically, the lower layer height (0.06–0.10 mm), the moderate nozzle temperature (195–205 °C), the lower printing speed (40 mm/s), and the infill density of 70% result in a superior compressive strength and compressive modulus due to improved interlayer bonding and enhanced material deposition.
• The eXtreme Gradient Boosting (XGBoost) model successfully predicts the compressive behavior of wood–PLA composite simple cubic lattice structures with high accuracy, achieving a minimal error range of 2–8%. Using a dataset of approximately 12,000 data points, the model effectively establishes the relationship between printing parameters and compressive performance.

What is the implication of the main finding?

• The influence of the manufacturing parameters is not solely dependent on the lattice structure but is strongly governed by the interaction between the material properties and processing conditions. A systematic evaluation through stress–strain analysis and microscopic observations helps to better understand and establish the relationship between the process parameters and compressive performance.
• The developed machine learning framework demonstrates strong predictive capability for lattice structures, confirming its suitability for performance prediction. This approach supports efficient decision-making, optimizing process parameters, and handling of large datasets, enabling a better integration of manufacturing variables with mechanical outcomes in advanced lattice structures.

Abstract

The present study investigates simple cubic lattice structures fabricated through an FDM-based three-dimensional (3D) printing method using wood–polylactic acid (wood–PLA) bio-composite filament and develops a data-driven framework to predict their mechanical response. The design of experiments (DOE) was developed using a response surface methodology (RSM) based on a central composite design (CCD) that was implemented in Design-Expert software (Version 13). During fabrication, four different manufacturing parameters—the layer height, the printing speed, the nozzle temperature, and the infill density—were considered. The compressive strength and compressive modulus were evaluated experimentally, and the corresponding stress–strain responses were examined. The results reveal that the layer height is the most influential parameter, where lower layer heights (0.06–0.1 mm) significantly improve both the compressive strength and the modulus due to enhanced interlayer bonding and reduced void formation. The printing speed and the nozzle temperature also play critical roles, where lower printing speeds (≈40 mm/s) and moderate nozzle temperatures (≈195–205 °C) promote more uniform material deposition and improved interlayer bonding, while higher speeds (≥60 mm/s) and excessive temperatures (≈225 °C) lead to reduced bonding quality and a deterioration in mechanical performance. In contrast, the infill density exhibited a non-monotonic influence, where intermediate levels (around 70%) provided an improved performance under combinations of the low layer height (≈0.1 mm), the low printing speed (≈40 mm/s), and the moderate nozzle temperature (≈195–215 °C), suggesting an interaction-driven effect rather than a purely density-dependent trend. To complement the experimental findings, a machine learning model based on eXtreme Gradient Boosting (XGBoost) was developed using 12,000 data points that were derived from stress–strain curves. The model successfully predicted continuous mechanical responses with errors in the range of 2–8% for unseen specimens, suggesting its capability to capture the relationship between printing parameters and mechanical behavior within the studied design space. Overall, the study highlights that the mechanical properties of wood–PLA lattice structures can be effectively tailored by choosing an appropriate printing parameter control and demonstrates the feasibility of using machine learning to estimate mechanical performance without additional physical testing within the defined parameter domain.

1. Introduction

The composite materials have gained significant attention in modern high-tech industries and in advanced applications due to their superior mechanical properties, low weight, and multifunctional capabilities compared with single-constituent or conventional materials [1]. The composite materials can be fabricated using different types of reinforcements, such as metals, polymers, ceramics, and natural fibers. Among these, polymeric composites are widely adopted in FDM printing because they provide a favorable mechanical strength, durability, and prolonged service life. However, the major drawbacks of polymeric composites include the release of certain volatile compounds, extremely long degradation times, and their contribution to carbon emissions, which significantly affect ecosystems, pollute the environment, and accelerate global warming [2]. In recent years, natural fiber-reinforced composites (NFRCs) have gained significant attention due to their ability to provide satisfactory mechanical properties and performance, including adequate strength, wide availability, recyclability, biodegradability, and low manufacturing cost. Most importantly, NFRCs align well with circular economy principles when compared with conventional synthetic or polymeric composites. Moreover, NFRCs are capable of mitigating the presence of volatile organic compounds commonly found in conventional polymer-based composites. In addition to these advantages, the use of natural fibers contributes to reduced carbon emissions and helps address climate change concerns, which are key industrial demands of the 21st century [3]. Consequently, the market size and production rates of NFRCs are increasing significantly as replacements for conventional or polymeric materials, with applications in biomedical, building and construction, automotive, aerospace, and related industries [2,3]. Natural fibers are primarily classified based on their sources; for example, plant fibers include wood, jute, bamboo, and PLAF, while animal fibers include sheep wool and others. Among these, plant fibers and their particles are extensively adopted into the composite industry due to their low preprocessing requirements, reduced fabrication complexity, and minimal post-processing needs [2,4].

Lattice structures are an advanced class of structures that are widely adopted in current industries, where voids are intentionally introduced to reduce the weight and material usage while achieving a high strength-to-weight ratio compared with conventional solid or block-type structures. Particularly in the aviation, aerospace, and automobile industries, the primary objective is to reduce structural weight to improve fuel efficiency. Consequently, lattice structures have become a preferred choice in these industries [5]. The application of lattice structures further depends on the component location, the functional requirements, and the design intent. In the biomedical industry, lattice structures are primarily used to promote biological responses, such as cell adhesion and proliferation, while maintaining a low weight and adequate mechanical strength [6]. Based on the type of lattice structures, the lattice can be categorized into simple cubic (SC), face-centered cubic (FCC), body-centered cubic (BCC), diamond-centered cubic, octet-centered cubic, truss-centered cubic, and so on [7]. However, among the various lattice structures, the simple cubic (SC) lattice is relatively simple in geometry, offers high design flexibility, and can provide superior mechanical properties. For instance, this study reported that SC lattices exhibited higher stiffness, yield strength, plateau stress, and energy absorption compared with other topology-optimized structures and body-centered cubic (BCC) lattices [8], which reflects that the types of lattices are responsible for providing different mechanical properties and performances. For any type of lattice structure, geometric parameters, such as the strut diameter, the strut volume fraction, and the repeating unit cell, are critical factors in addition to the unit cell topology itself. These parameters can significantly influence—and even markedly alter—the mechanical responses that are under investigation. For example, this research conducted a comparison between different strut diameters (0.5–1.5 mm), strut lengths (12–18 mm), and repeating units (3 × 3 × 3 and 6 × 6 × 6). The study revealed that a higher number of repeating units (6 × 6 × 6) with a strut diameter of 1.5 mm provided superior compressive properties compared with structures having fewer repeating units (3 × 3 × 3) [9]. However, the complexity of these structures arises because the repeating unit size, strut diameter, and strut length are not always identical, making it challenging to establish a standardized experimental procedure. This variation occurs because strut geometry, orientation, and connectivity differ depending on the lattice type, leading to changes in the structural configuration and mechanical responses. An earlier study investigated the effects of unit cell size and strut volume fraction, revealing that variations in these parameters significantly increased or decreased the compressive strength, elastic modulus, and other mechanical properties as their values changed [10]. However, the study established certain linear relationships, showing that an increase in lattice volume fraction led to a higher compressive modulus and strength. This behavior is primarily attributed to an increased stress concentration and load-bearing capacity around the pores. Conversely, an increase in unit cell size resulted in a reduction in the overall mechanical properties of the structure [10].

To fabricate the lattice structures, a wide range of conventional and advanced manufacturing technologies has been developed. Among these, additive manufacturing (AM) is particularly advantageous due to its design freedom, sustainability, lower carbon emissions, reduced energy consumption, and capability to produce intricate geometries. Based on dimensional flexibility, AM technologies are commonly classified into three-dimensional (3D), four-dimensional (4D), and five-dimensional (5D) printing [2]. In 4D printing, stimulus-responsive behavior is introduced as an additional functional feature. However, compared with 4D and 5D printing, 3D printing is more widely adopted in the current composite industries because it enables the fabrication of complex geometries with fewer complications, shorter production times, and lower costs. 3D printing technologies are broadly categorized into seven types, including material extrusion, vat photopolymerization, sheet lamination, powder bed fusion, material jetting, binder jetting, and directed energy deposition [11]. Among these, material extrusion, commonly known as fused deposition modeling (FDM) or fused filament fabrication (FFF), has gained significant attention due to its low manufacturing time, ease of material handling, wide range of feedstock materials, relatively low energy consumption, and strong alignment with circular economy principles. Consequently, FDM-based 3D printing is extensively employed for fabricating intricate geometries such as lattice structures in current applications [12]. In FDM printing machines, manufacturing parameters have a significant influence on the mechanical properties of the printed parts. In addition, the structural performance of components may also depend on the selected manufacturing parameters. For example, this research investigated the effects of key manufacturing parameters—namely layer height, printing speed, nozzle temperature, and infill density—using the Taguchi method on hexagonal lattice structures fabricated from a PLA/walnut shell composite. The results revealed that each parameter had a significant influence on compressive strength. Among these, layer height and nozzle temperature showed the highest contributions, accounting for 37.07% and 38.05%, respectively [13]. The relationship between the manufacturing parameters and the response variables is not always linear or identical. A similar approach was applied to a PLA-based fluorite lattice structure. The investigation revealed that, among the parameters that were considered, nozzle temperature made the most significant contribution (72.44%), followed by printing speed (22.81%), while layer height showed a negligible effect. The optimal results were obtained at a layer height of 0.16 mm and a nozzle temperature of 205 °C, yielding a compressive strength of 12.22 MPa [14]. These findings indicate that lattice performance depends not only on FDM fabrication parameters but also significantly on the material used.

Particularly in composite-based lattice structures, significant research efforts have been carried out. For example, a study designed a planar lattice structure using polyamide reinforced with carbon fiber and reported favorable thermal dimensional stability when fabricated using the FDM method [15]. In another investigation, polylactide (PLA)/polyhydroxyalkanoate (PHA)-based lattice composites with different variants of triply periodic minimal surface (TPMS) structures—such as gyroid, Kelvin, and Schwarz-D—were fabricated using FDM. A comparative analysis revealed that the gyroid TPMS exhibited the highest elastic modulus of approximately 252.32 MPa [16]. A similar approach was adopted in another study, which considered three different lattice geometries (gyroid, diamond, and primitive) with cell sizes of 8 mm and 12 mm. Based on the data analysis, the diamond lattice with an 8 mm cell size showed a superior mechanical performance, achieving an elastic modulus of 0.549 GPa and a compressive strength of 12.28 MPa. In contrast, the 12 mm cell size resulted in a lower elastic modulus of 0.364 GPa and a compressive strength of 11.41 MPa. These results indicate that both the lattice geometry and the dimensional parameters significantly influence the mechanical performance in addition to the material selection [17]. Furthermore, a comparative experimental and finite element method (FEM) study was conducted on the FDM-fabricated TPMS gyroid lattice structures with varying unit cell configurations. The investigation found that the 2 × 2 × 2 unit cell configuration exhibited a higher elastic modulus compared to the other configurations, while similar elastic moduli were observed for the 4 × 4 × 4 unit cell arrangement [18]. In another research, carbon fiber-reinforced polyethylene terephthalate glycol (PETG) composite octagonal lattice structures were fabricated using FDM, maintaining a constant cell size of 3 mm while varying the manufacturing parameters such as layer height, printing speed, nozzle temperature, line width, and infill density. The statistical analysis revealed that most of the considered parameters had negligible effects on the mechanical responses; however, the layer height showed a notable contribution of 49.347% to the compressive strength [19]. A similar approach was also adopted in a study involving the FDM-printed carbon fiber-reinforced PLA matrix composites with an advanced hybrid lattice design. The results demonstrated substantial improvements compared to the conventional lattice structures, including increases of 12.7 times in the elastic modulus, 5.4 times in the yield strength, and 4.4 times in the energy absorption capacity [20].

Artificial intelligence (AI) and machine learning (ML) are advanced predictive techniques that have been widely adopted in modern industries to estimate mechanical performance and material properties with high accuracy, reducing the complexity associated with the traditional statistical and theoretical computation methods [21]. Numerous machine learning models have recently been applied in the composite materials industry [3]. Among these, classical ML models such as linear regression, decision trees, random forests, and support vector machines are commonly used. In contrast, ensemble methods, including XGBoost, CatBoost, and LightGBM, are particularly effective for handling complex datasets and capturing nonlinear relationships. Furthermore, neural networks and deep learning approaches are increasingly employed in dense network frameworks and multimodal frameworks for advanced predictive investigations [3]. The prediction accuracy is commonly evaluated using performance metrics such as R², MAE, MSE, RMSE, and MAPE.

In the context of composite material performance and property prediction, ML models have been extensively adopted. Artificial neural network (ANN) models have been used to predict the influence of FDM process parameters, such as infill density and printing orientation, on polyamide–carbon fiber and polyamide–glass fiber composites. These studies reported a high prediction accuracy, with the R² values approaching 0.99 for tensile strength and elastic modulus prediction, outperforming the conventional statistical methods [22]. Similarly, XGBoost models applied to the FDM-printed natural fiber-reinforced composites (NFRCs) achieved R² values in the range of 0.99–1.00 for predicting tensile, flexural, and impact properties, surpassing the linear regression and SVM models [23]. However, a multi-objective optimization framework was developed in this research for the FDM-printed carbon fiber-reinforced nylon composites using different AI models. The ANN model outperformed linear regression, achieving an R² value of approximately 0.99 when predicting tensile and impact strength, while the PSI–VIKOR method was used to determine the optimal manufacturing parameters [24].

For triply periodic minimal surface (TPMS) lattices, a machine learning framework has been developed for the FDM-printed PLA-based lattice structures. The comparative studies among different ML models revealed that random forest and decision tree models outperformed the others, achieving R² values of 0.99–1.00 and RMSE values between 0.08 and 0.12, outperforming the CNN and Bayesian regression models [25]. In another study, an auxetic foam-type lattice structure was designed, and its mechanical performance was predicted based on the lattice type, cell size, and wall thickness using various ML models. Among these, ANN achieved the best performance, outperforming linear regression, random forest, and decision tree models, with an R² value of 0.93 and demonstrating low prediction errors. The results indicated that cell size and wall thickness were the most influential parameters governing the mechanical behavior [26].

Despite the growing interest in ML-based prediction of lattice composite performance, studies focusing specifically on lattice-reinforced composites remain limited. To predict the flexural properties of the FDM-printed Kevlar-reinforced ABS lattice composites, several ML models—including linear regression, ANN, random forest, and SVM—were applied to TPMS lattice geometries such as gyroid and diamond. The results demonstrated that the random forest model achieved the best prediction performance, with R² = 0.93, RMSE = 0.09, and MAE = 0.07 [21]. Similarly, an AI-integrated framework was employed to investigate the performance of the TPMS lattice structures that were fabricated via FDM from the PLA/HAP/GO composites, considering parameters such as strut thickness, porosity, and cell geometry. The study reported approximately 20% improvement in energy absorption and a 15% enhancement in thermo-mechanical properties compared to traditional designs. In this framework, AI primarily functioned to correlate the structural parameters with mechanical properties and to guide lattice optimization, ultimately demonstrating suitability for bone scaffold applications [27].

From the above literature review, it is evident that although the FDM-based 3D printing of lattice structures, the influence of printing parameters on their mechanical properties, and the machine learning based prediction of composite mechanics have been widely studied, their integration with a natural fiber-reinforced composite (NFRC) lattice remains limited. The existing studies primarily focus on the conventional polymers or advanced lattice geometries, whereas less attention is given to the wood–PLA-based lattice structures and their compressive behavior. In addition, most machine learning investigations emphasize a discrete property prediction rather than capturing the continuous stress–strain response as a function of the process parameters. Therefore, there is a need for a systematic framework that combines experimental characterization with data-driven modeling for NFRC lattice structures.

Accordingly, the primary objective of this study is to investigate the influence of FDM printing parameters on the compressive behavior of wood–PLA lattice structures, to identify optimal parameter combinations using a statistical design framework, and to develop a machine learning-based model that is capable of predicting their mechanical response directly from the printing parameters within the defined design space. In this manuscript, Section 2 presents the material selection, fabrication process, and design of experiments, and Section 3 discusses the stress–strain behavior obtained for each design of experiments (DOE). In Section 3, the experimental results corresponding to each DOE are analyzed in detail, including machine learning-based prediction and optimization. Section 3 also focuses on microscopic analysis and provides mechanistic interpretations of the observed behavior with respect to the selected parameters. Finally, Section 4 outlines future research perspectives and Section 5 summarizes the key conclusions.

2. Materials, Design of Experiments, Experimental Setup

2.1. Materials

In this research, a wood–PLA composite containing 10% wood particles was used. The wood–PLA filament was commercially purchased from Qidi, a manufacturer of filaments for FDM 3D printing [28]. Although various types of commercial wood–PLA filaments are available, a 10% wood content was selected for this study because these proportions are more suitable for NFRCs. A comparative study evaluated the effect of wood content ranging from 0% to 50% in PLA composites, focusing on tensile strength [29]. The results revealed that the highest tensile strength (57 MPa) was obtained at 10% wood content. As the wood proportion increased, both the tensile strength and the elastic modulus decreased linearly. This reduction may be attributed to increased brittleness at higher wood loadings, leading to an earlier collapse or failure of the material. Furthermore, the lowest tensile strength (30 MPa) was observed at 50% wood content. However, in terms of dimensional stability, the 10% wood–PLA composite exhibits slight deviations in the moisture absorption and thickness swelling ratio compared with pure PLA [30]. Nevertheless, due to the relatively low wood content, the dimensional changes in the printed parts remain limited and do not significantly affect their structural stability. The density of the considered wood–PLA filament for this study ranged from 1.22 to 1.24 g/cm³ [28].

2.2. Design of the Experiment

The design of the experiment was developed using the statistical analysis software Design-Expert (Version 13). Four critical input parameters were considered: layer height, nozzle temperature, printing speed, and infill density. The previous studies have indicated that, for NFRCs, lower layer heights, lower nozzle temperatures, and moderate to high infill densities are generally more suitable for the FDM-fabricated parts [11]. However, these variables do not always produce consistent outcomes, as the results may vary depending on the geometry and structural configuration of the printed components. Therefore, the selected parameter ranges were determined based on the existing literature and the manufacturer’s recommended settings [11,28]. A central composite design (CCD) was employed to structure the experimental plan. The design consisted of 30 experimental runs, including 24 unique design points and 6 replicates at the center points. The detailed design matrix is presented in

Table 1. The experimental design matrix for the wood–PLA lattice composite prepared using a central composite design (CCD).

Input Variables	Low Level	High Level
Layer height (mm)	0.1 mm	0.3 mm
Printing speed (mm/s)	40 mm/s	60 mm/s
Nozzle temperature (°C)	195 °C	215 °C
Infill density (%)	70%	100%

.

Apart from these variables, several additional printing parameters were kept constant throughout the investigation, as previous studies have indicated that they do not have a significant influence on the measured outcomes. These fixed parameters are listed in

Table 2. The constant process parameters used in the FDM 3D printing machine.

Constant Parameters	Values
Infill pattern	Cubic
Travel speed	100 mm/s
Build plate temperature	60 °C
Fan speed	100%
Z offset	0.0 mm
Top layers and bottom layers	4
Travel speed	100%

.

2.3. Sample Fabrication and Experimental Setup

For this study, the sample dimensions were kept consistent with those reported in the reference study according to ASTM D695 [12], with a length of 36 mm, a width of 36 mm, and a height of 30.21 mm. The specimens were fabricated using an FDM (Fused Deposition Modeling) 3D printer (QIDI brand, model QIDI Max 4) with a build volume of 390 × 390 × 340 mm³ and a nozzle size of 0.4 mm. To fabricate the lattice structure, a CAD model of a simple cubic lattice was first designed in ABAQUS CAE (Version 2017). The designed lattice was then exported as an STL file, which was subsequently imported into the QIDI slicing software (Version 6.5). Based on the designed experiments, the manufacturing parameters were set to generate the required G-code for printing. The generated G-code was then transferred to the FDM printer for fabrication.

During fabrication, nozzle clogging was encountered due to the use of natural fiber-reinforced wood–PLA filament. To mitigate this issue, the nozzle was thoroughly cleaned before and after each fabrication process. Figure 1a illustrates the FDM fabrication process, while Figure 1b presents the fabricated wood–PLA composite lattice structure.

The primary objective of this research was to investigate the compressive behavior of the wood–PLA composite lattice. The compression testing was performed using an MTS 322 test frame with a maximum load capacity of 100 kN. The loading rate can significantly influence the mechanical response of the lattice structures. To ensure consistency and avoid unintended variations in the results, a crosshead speed of 2 mm/min was selected, following previously reported studies [14,31]. Furthermore, to improve the accuracy and reliability of the experimental results, the compression displacement was limited to 22 mm. Beyond this displacement, the lattice structure exhibited densification behavior, effectively transforming into a sandwich-like compact structure and demonstrating an artificially elevated compressive strength. Therefore, restricting the compression bar travel ensured that the results reflected the intrinsic compressive performance of the lattice structure prior to full densification. The compression test setup is illustrated in Figure 1c.

Finally, to compare the mechanical performance within the designed DOE framework, the fracture regions of the wood–PLA lattice composites were examined using a microscope, as shown in Figure 1d. The microscope used in this study was a Zeiss Stemi 2000-C model. The fractured surfaces were analyzed at different magnifications to evaluate the failure modes and crack propagation characteristics.

After conducting the compression test, the load–displacement curve was obtained for each experimental design. For the lattice structures, two stress definitions are commonly used: (i) nominal stress, calculated based on the total cross-sectional area, and (ii) effective stress, calculated using the load-bearing solid area after excluding void regions. In this study, the effective cross-sectional area was adopted to better represent the actual load transfer through the lattice structure. This approach is consistent with analyses where the mechanical response is interpreted based on the solid fraction of the structure [12,32,33].

The stress values were calculated by dividing the applied load by the effective cross-sectional area. The total cross-sectional area of the specimen was 1296 mm². After excluding the void region (577 mm²), the effective load-bearing area was determined to be 719 mm². This effective area was used for all stress calculations of the wood–PLA composite lattice.

To evaluate and understand the compressive behavior of the wood–PLA lattice composite, two key responses were considered: the compressive strength and the compressive modulus. These responses were determined for each DOE specimen based on its corresponding stress–strain curve. The previous studies have shown that higher values of compressive strength and compressive modulus indicate a stronger and stiffer material or structure compared to those with lower values [12]. The compressive strength and compressive modulus that are reported in this study represent the average of five experiments, which were conducted to minimize experimental errors and to improve the accuracy and reliability of the overall investigation.

2.4. Machine Learning Modeling

In this research, machine learning (ML) models were employed as an efficient predictive tool to reduce the reliance on destructive experimental testing and computationally expensive, time-consuming numerical simulations. The traditional experimental and finite element analyses, although accurate, require significant material preparation, testing time, and computational resources. By learning the relationship between the printing parameters and the corresponding compressive strength and the compressive modulus response from a limited set of experimental data, the developed ML models enable a rapid and reliable prediction for new configurations.

The dataset that was constructed for machine learning prediction in this study was derived from 24 unique specimen configurations, each tracked via a unique specimen ID. The experimental stress–strain data for each specimen was uniformly sampled to yield 500 discrete points per curve. Consequently, a comprehensive dataset comprising 12,000 rows and 6 columns was generated. For the development of the predictive model, stress was defined as the target feature, while layer height, printing speed, nozzle temperature, infill density, and strain served as the input features.

Table 3. The dataset for machine learning modeling.

Specimen_ID	Layer Height, mm	Printing Speed, mm/s	Nozzle Temperature (°C)	Infill Density (%)	Strain	Stress
1	0.2	50	185	85	0	0.018539
1	0.2	50	185	85	0.000159	0.019113
1	0.2	50	185	85	0.000318	0.019687
.	.	.	.	.	.	.
.	.	.	.	.	.	.
.	.	.	.	.	.	.
.	.	.	.	.	.	.
24	0.3	40	215	70	0.081526	2.617483
24	0.3	40	215	70	0.081689	2.614002

represents the dataset used for this study.

The models utilized in this work are based on ensemble learning techniques. The fundamental principle of ensemble learning is to employ more than one learner and integrate their predictions into a single model that is stronger and more accurate than any individual learner. By combining multiple models, ensemble learning improves predictive performance, robustness, and generalization capability. This framework includes several methodologies, among which bagging and boosting are the most prominent. Boosting is a major ensemble learning technique and is regarded as one of the most powerful learning methods, forming the foundation of advanced algorithms such as XGBoost, LightGBM, and CatBoost [34]. In this research, only the XGBoost model is used.

2.4.1. XGBoost

Extreme Gradient Boosting (XGBoost) is a scalable implementation of gradient tree boosting that constructs new models sequentially to correct the pseudo-residuals of the previously built ensemble [35]. The high performance of XGBoost arises from both algorithmic and system-level improvements. At the algorithmic level, XGBoost enhances traditional gradient boosting by incorporating a regularized learning objective that includes both L₁ and L₂ penalties, which penalize model complexity and reduce overfitting. Figure 2 illustrates the algorithm of XGBoost.

Another important feature of XGBoost is its sparsity-aware split-finding algorithm, which allows the efficient handling of sparse data and missing values by learning optimal default paths during tree construction. From a system-level perspective, XGBoost is designed for scalability and computational efficiency through techniques such as cache-aware access patterns and out-of-core computation for datasets that exceed system memory capacity. These improvements enable XGBoost to achieve high predictive accuracy while maintaining fast training and evaluation performance [36].

2.4.2. Training Strategy and Evaluation Metrics

To ensure a fair evaluation and prevent data leakage, the dataset was divided into an 80% training set and a 20% testing set. Instead of a standard random split, the data was strictly grouped by Specimen_ID. This grouping is crucial when testing physical data, as it ensures that all the data points belonging to a single 3D-printed specimen remain entirely within either the training or the testing set. This approach forces the model to predict the behavior of truly unseen physical specimens, rather than just memorizing fragments of a stress–strain curve that it has already been exposed to.

The machine learning models were trained on the 80% subset using 5-fold cross-validation. A fixed random seed was applied throughout this process to guarantee reproducible results and to minimize the random sampling bias. Finally, the model’s performance and predictive accuracy were quantified using three standard regression metrics: the coefficient of determination (R²), the root mean square error (RMSE), and the mean absolute error (MAE). The model performance is evaluated using three standard regression metrics, as described below.

Root Mean Squared Error (RMSE):

The RMSE measures the square root of the average squared differences between predicted and observed values, thereby assigning greater weight to larger prediction errors and emphasizing model sensitivity to significant deviations. Equation (1) represents the numerical formulation of the RMSE.

$$\begin{array}{c}RMSE=\sqrt{{\displaystyle \frac{1}{n}}{\displaystyle \sum _{i=1}^n}{\left(y_i-\widehat{y_i}\right)}^2}\end{array}$$

(1)

Mean Absolute Error (MAE):

The MAE computes the average absolute difference between the predicted and observed values, providing an intuitive and easily interpretable measure of the overall model prediction accuracy. Equation (2) is the numerical equation of the MAE.

$$\begin{array}{c}MAE={\displaystyle \frac{1}{n}}{\displaystyle \sum _{i=1}^n}\left|y_i-\widehat{y_i}\right|\end{array}$$

(2)

Coefficient of Determination (R²):

R² quantifies the proportion of variance in the observed data explained by the model, where a value of 1 represents a perfect prediction. The numerical formula of R² is given in Equation (3).

$$\begin{array}{c}{R}^2=1-{\displaystyle \frac{\sum _{i=1}^n{\left(y_i-\widehat{y_i}\right)}^2}{\sum _{i=1}^n{\left(y_i-\overline{y}\right)}^2}}\end{array}$$

(3)

This multi-metric evaluation framework ensures that each model is optimized not only for minimizing absolute and squared prediction errors but also for accurately capturing the variance in the experimental stress response across different DOE.

2.4.3. Hyperparameter Tuning and Optimization

The performance of modern machine learning (ML) systems is fundamentally governed by two distinct classes of variables: model parameters, which are learned directly from data during training, and hyperparameters, which are external configuration variables that define the learning process, model complexity, and optimization behavior [37]. Unlike model parameters, hyperparameters are not updated automatically during training and must be specified prior to model fitting. Inappropriate hyperparameter selection may therefore lead to underfitting, overfitting, or unstable learning behavior.

The systematic process of identifying the optimal hyperparameter configurations is referred to as hyperparameter tuning or hyperparameter optimization [38]. In ensemble-based learning algorithms, hyperparameters such as tree depth, learning rate, number of estimators, sampling ratios, and regularization coefficients strongly influence the bias–variance trade-off and the overall generalization performance. Consequently, the careful optimization of these hyperparameters is essential, particularly when working with experimentally derived datasets of a limited size.

Several strategies exist for hyperparameter optimization, including manual tuning, grid search, randomized search, and Bayesian optimization. Among these approaches, a randomized search has been shown to be more efficient than a grid search for high-dimensional hyperparameter spaces, as it explores a broader range of configurations while requiring significantly fewer evaluations [39]. Rather than exhaustively testing all parameter combinations, a randomized search samples the hyperparameter values from predefined distributions, enabling an effective exploration of the search space with a reduced computational cost.

In this study, randomized search-based hyperparameter tuning was applied uniformly in the XGBoost model to ensure consistency and fair comparison. For each model, a total of 500 random hyperparameter configurations (n_iter = 500) were evaluated using 5-fold cross-validation (cv = 5). The model performance during tuning was assessed using the negative root mean squared error (neg_root_mean_squared_error) as the optimization criterion, ensuring a direct minimization of prediction error. To improve computational efficiency, parallel processing was enabled using n_jobs = −1, allowing the utilization of all the available processing cores. A fixed random state (random_state = 42) was employed to ensure the reproducibility of the optimization results, while verbose = 1 was used to monitor the tuning progress.

The optimized hyperparameter sets obtained from this procedure were subsequently used to train the final models. The selected hyperparameters and their corresponding search ranges for each model are summarized in the following

Table 4. The hyperparameter tuning ranges and the optimal values for XGBoost.

Machine Learning Technique	Hyperparameters	Range Tested
XG Boost	Learning rate	0.01–0.19
	Max depth	3–10
	Subsample	0.4–0.6
	Colsample bytree	0.4–0.6
	N estimators	50–300

. This standardized and systematic tuning strategy ensures that all the models operate under near-optimal conditions and enables a reliable performance comparison across the different ensemble learning techniques.

In summary, to ensure accurate predictions on unseen data, the input dataset is grouped based on specimen ID and split into an 80/20 training-to-testing ratio. An XGBoost base model is trained first, followed by a hyperparameter-tuned model. After evaluating the comparisons, the highest predictive accuracy is found in the base model. Consequently, this base model is deployed to predict the continuous stress–strain curves for the unseen data. Finally, the compressive modulus and compressive stress are extracted from these predicted curves and evaluated against the actual experimental data. Figure 3 represents the workflow for machine learning modeling in the present study.

3. Results and Discussions

3.1. Stress–Strain Curve Analysis

In Figure 4, the stress–strain behavior of wood–PLA composite lattice structures is depicted based on the design of experiments from DOE-1 to DOE-30. The variations observed among the thirty curves clearly demonstrate that the stress–strain response of the wood–PLA composite largely depends on the processing parameters. In the stress–strain curve, the region up to the first peak stress represents the elastic region, the peak stress corresponds to the compressive strength, while the post-peak region represents failure after the peak stress.

Experiments such as DOE-1, DOE-4, DOE-10, DOE-14, DOE-16, and DOE-18, with moderate layer height (0.2 mm) and minimum infill density (85%), provide a stable elastic region and a balanced peak stress in the stress–strain curve shown in Figure 4. On the other hand, DOE-2, DOE-3, DOE-6, DOE-7, DOE-9, and DOE-15, which are printed with a lower layer height (0.1 mm) and higher nozzle temperatures (up to 215 °C), show an improved initial slope in the elastic region, mainly due to enhanced interlayer bonding and better filament fusion.

In particular, the influence of the infill density is observed in DOE-5, DOE-6, DOE-9, DOE-17, DOE-26, and DOE-28, where 100% infill density provides higher peak stress and a delayed collapse compared to lower infill densities such as DOE-2, DOE-7, DOE-22, and DOE-29. For example, DOE-29, with a 55% infill density, exhibits an earlier collapse despite having a moderate layer height (0.2 mm), a moderate printing speed, and a moderate nozzle temperature. This specimen could not support the overall lattice structure due to a lack of internal support.

However, the nozzle temperature also significantly affects the stress–strain behavior of the wood–PLA lattice composite. For instance, DOE-1 was printed at 185 °C compared to DOE-20 at 225 °C. A higher temperature promotes a stronger interlayer adhesion in the composite, reduces voids in the structure, and results in smoother transitions in the stress–strain response. However, excessively high temperatures may also cause unintentional difficulties such as imperfect bonding, poor adhesion, and a distortion of the intended structure.

In terms of printing speed effectiveness, it can be observed that moderate speeds (40 mm/s), as seen in DOE-5, DOE-15, DOE-17, and DOE-24, provide more stable stress–strain curves than DOE-11 and DOE-12, which were printed at higher speeds of 60–70 mm/s. This is possibly because the increased printing speed reduces the bonding time, which may introduce slight inconsistencies in the internal structure and earlier drops in the stress–strain curve.

The larger layer heights (0.3 mm) generally reduce stiffness due to thicker filament deposition and comparatively weaker bonding between layers, as seen in DOE-17 and DOE-26. Although these two experiments consist of high infill density (100%), the thicker layers imbalance the overall structure. On the other hand, a lower layer height may improve the mechanical integrity; however, the residual stresses may increase depending on the temperature, speed, and infill density combinations.

3.2. Compressive Strength and Compressive Modulus Analysis

Table 5. The experimental design and corresponding compressive strength and compressive modulus results.

Sl No.	Layer Height (mm)	Printing Speed mm/s	Nozzle Temperature (°C)	Infill Density (%)	Compressive Strength (MPa)	Compressive Modulus (MPa)
1	0.2	50	185	85	3.56 MPa	110 MPa
2	0.1	60	215	70	3.23 MPa	124 MPa
3	0.1	60	215	100	3.06 MPa	113 MPa
4	0.2	50	205	85	3.27 MPa	111 MPa
5	0.1	40	195	100	3.48 MPa	117 MPa
6	0.1	60	195	100	3.42 MPa	112.66 MPa
7	0.1	60	195	70	3.62 MPa	139 MPa
8	0.06	50	205	85	3.85 MPa	140 MPa
9	0.1	40	215	100	3.35 MPa	125 MPa
10	0.2	50	205	85	3.27 MPa	111 MPa
11	0.2	70	205	85	3.48 MPa	132 MPa
12	0.2	30	205	85	3.46 MPa	139 MPa
13	0.2	50	205	50	3.13 MPa	108 MPa
14	0.2	50	205	85	3.27 MPa	3.27 MPa
15	0.1	40	195	70	4.00 * MPa	157 * MPa
16	0.2	50	205	85	3.27 MPa	111 MPa
17	0.3	40	195	100	2.80 MPa	91.44 MPa
18	0.2	50	205	85	3.27 MPa	111 MPa
19	0.2	50	205	85	3.27 MPa	111 MPa
20	0.2	50	225	85	3.02 MPa	85 MPa
21	0.3	40	215	100	2.76 MPa	93 MPa
22	0.3	60	215	70	2.77 MPa	83 MPa
23	0.3	40	195	70	2.9 MPa	102 MPa
24	0.3	40	215	70	2.77 MPa	80.66 MPa
25	0.3	60	195	70	2.72 MPa	79.66 MPa
26	0.3	60	215	100	2.9 MPa	80 MPa
27	0.1	40	215	70	4.11 * MPa	139 * MPa
28	0.3	60	195	100	2.66 * MPa	76 * MPa
29	0.2	50	205	55	3.2 MPa	120 MPa
30	0.32	50	205	85	2.89 MPa	84 MPa

presents the 30 design of experiments (DOE) trials that were conducted to evaluate the effect of key FDM process parameters on the compressive strength and compressive modulus of wood–PLA composite lattice structures. The reported results represent the average values obtained from the three samples that were tested. Among the experimental runs, the highest compressive strength of 4.11 MPa was achieved at a layer height of 0.1 mm, nozzle temperature of 215 °C, printing speed of 40 mm/s, and infill density of 70%. Similarly, the maximum compressive modulus of 157 MPa was obtained at a layer height of 0.1 mm, nozzle temperature of 195 °C, printing speed of 40 mm/s, and infill density of 70%. In contrast, the lowest compressive strength of 2.66 MPa and the compressive modulus of 76 MPa were observed at a layer height of 0.3 mm, printing speed of 60 mm/s, nozzle temperature of 195 °C, and infill density of 100%. The highest and lowest values of compressive strength and modulus are highlighted in Table 5 using an asterisk (*) for clarity.

It is clearly evident from Table 5 that the layer height produces the strongest effect on the compressive performance of the wood–PLA lattice composite. The experimental results show that lower layer heights in the range of 0.06–0.1 mm significantly improved the compressive strength, reaching the maximum value of 4.11 MPa. A similar trend was observed for the compressive modulus, as the highest modulus values among the 30 DOE trials were also obtained at low layer heights. This improvement can be attributed to an enhanced interlayer fusion, a reduced void formation, and a stronger bonding between deposited filaments. Conversely, increasing the layer height to 0.3–0.32 mm substantially reduced the compressive strength to approximately 2.6–2.9 MPa and the compressive modulus to around 76–80 MPa. This reduction reflects a poorer interlayer adhesion, an insufficient fusion, a weaker fiber–matrix interaction, a reduced load-bearing capacity, and an earlier structural failure during compression.

In terms of printing speed, lower speeds generally provide better compressive strength and stiffness compared to higher speeds. For example, DOE-15 and DOE-27 achieved the highest compressive strength (4.11 MPa) and modulus (157 MPa) when the printing speed was 40 mm/s. However, when the printing speed increased to 60–70 mm/s, the compressive strength and modulus decreased to approximately 3.23–3.62 MPa and 111–139 MPa, respectively. This reduction is mainly due to faster cooling rates at higher speeds, which limit polymer diffusion and weaken interlayer bonding. Nevertheless, this trend is not strictly uniform across all cases. An exception is observed in DOE-11, where, despite the highest speed of 70 mm/s, the lattice exhibited a moderate compressive strength (3.48 MPa) and modulus (132 MPa). This behavior may be because of the combined influence of a moderate nozzle temperature (205 °C) and an infill density of 85%, which likely promoted improved material fusion and interfacial bonding.

With respect to nozzle temperature, most DOE’s conducted within a moderate temperature range of 195–205 °C produced relatively stable and higher modulus values, reaching up to 140–157 MPa. However, when the nozzle temperature increased to 225 °C, the compressive modulus dropped significantly to approximately 85 MPa. This decline suggests that excessive thermal input may cause degradation of the wood–PLA composite, reduce dimensional stability, and over-soften the PLA matrix, ultimately weakening the lattice structure. A similar influence of the nozzle temperature on the mechanical performance of NFRCs has also been reported in previous studies [40].

Finally, the infill density also influenced the compressive behavior by controlling the internal structural support of the lattice. However, it does not follow a strictly monotonic trend. The specimens fabricated with a 100% infill demonstrated moderate compressive strength and modulus in the range of 3.35–3.48 MPa and 117–125 MPa, respectively. However, intermediate infill levels, such as 85%, often provided improved stiffness and strength compared to a 100% infill, likely due to an optimized balance between the internal support and the structural flexibility. Interestingly, a lower infill density of 70% produced the highest compressive strength and modulus in several cases, particularly in DOE-15 and DOE-27. A similar phenomenon regarding the infill density was observed in this research, where a 50% infill density exhibited a higher fatigue performance than the 90% and 100% infill configurations in the lattice structures [41]. This indicates that an optimal infill level may exist where the lattice geometry effectively distributes compressive loads while maintaining efficient bonding and a reduced stress concentration. Therefore, the influence of the infill density should be interpreted as condition-dependent and governed by interactions among the printing parameters rather than as a purely density-driven effect.

Overall, the DOE results confirm that the compressive performance in the FDM-printed wood–PLA lattice composites is governed by the combined effects of layer height, printing speed, nozzle temperature, and infill density, with layer height being the most dominant parameter affecting both compressive strength and stiffness.

3.3. Microscopic Analysis

A comparative study was conducted among the different design of experiments (DOE) considerations. To further correlate the manufacturing parameters with the compressive strength and compressive modulus, microscopic analysis was performed. Figure 5a–c represents DOE-1, whereas Figure 5d–f corresponds to DOE-20. The primary difference between these two DOE conditions is the printing temperature. As observed in Figure 5a–c, the lower printing temperature of 185 °C resulted in improved interfacial bonding at the strut junctions, promoting a more effective load transfer within the lattice structure compared to DOE-20. In contrast, the higher printing temperature of 225 °C, shown in Figure 5d–f, exhibited noticeable fiber pull-out, indicating a weaker adhesion between the wood fibers and the PLA matrix. This behavior is likely associated with thermal degradation or matrix softening at elevated temperatures. Therefore, excessively high printing temperatures are not suitable for natural fiber–reinforced wood–PLA composite lattice structures. To enhance the mechanical performance, low to moderate printing temperatures are more appropriate, as they help preserve interfacial bonding and maintain desirable compressive strength and modulus.

In Figure 6a–c, a comparison was also conducted between two printing speeds, 40 mm/s and 60 mm/s. As observed in Figure 6a–c, the specimen printed at the higher speed of 60 mm/s (DOE-3) exhibited a lower compressive strength and compressive modulus. This reduction can be attributed to irregular material deposition and the presence of interlayer gaps, which weaken the structural integrity of the lattice. In contrast, the sample fabricated at the lower printing speed of 40 mm/s (DOE-9) showed a more uniform filament deposition and improved interlayer bonding. Consequently, the lower printing speed resulted in enhanced compressive strength and overall mechanical performance.

To evaluate the effect of layer height, a comparison was conducted between DOE-8 and DOE-30. In DOE-8, the layer height was 0.06 mm, whereas in DOE-30 it was 0.32 mm. All the other printing parameters, including nozzle temperature, printing speed, and infill density, were kept constant to isolate the influence of layer height. As shown in Figure 7a–c, the specimen fabricated with a 0.06 mm layer height exhibits more uniform material distribution and minimal interlayer gaps, resulting in improved interfacial bonding and a more efficient load transfer throughout the lattice structure. This enhanced structural integrity leads to a higher compressive strength and compressive modulus. In contrast, the sample printed with a 0.32 mm layer height, as seen in Figure 7d–f, demonstrates noticeable interlayer gaps and an irregular filament deposition, which create structural discontinuities and stress concentration sites. These microstructural defects weaken the overall lattice architecture, resulting in a reduced compressive strength and compressive modulus compared to the specimen printed with the lower layer height.

To understand the role of the infill density, a comparison was conducted between DOE-2 and DOE-3. The primary difference between these two experimental conditions is the infill density, which was 70% for DOE-2 and 100% for DOE-3, while all the other printing parameters remained constant. As shown in Figure 8a–c, the specimen fabricated with a lower infill density of 70% exhibits a smoother surface morphology and a more uniform material distribution within the lattice structure. This improved deposition promotes effective load transfer, resulting in a higher compressive strength and compressive modulus.

In contrast, the sample printed with a 100% infill density (Figure 8d–f) demonstrates excessive material accumulation, overlapping filament layers, and irregular deposition at the junction regions. Such congestion introduces stress concentration sites and reduces the structural uniformity, leading to a lower compressive strength and compressive modulus. These findings suggest that a moderate infill density (70%) provides a more optimized microstructure and a comparatively better mechanical performance for wood–PLA lattice structures.

Finally, a comparison was conducted between DOE-27 and DOE-28, representing the highest and lowest compressive strength values, respectively. As shown in Figure 9a–c, DOE-27, fabricated with a lower layer height (0.1 mm), higher nozzle temperature (215 °C), lower printing speed (40 mm/s), and moderate infill density (70%), exhibits a uniform material distribution and a well-defined node formation within the lattice structure. The improved interlayer bonding and controlled filament deposition promote efficient load transfer, resulting in the highest compressive strength (4.11 MPa) and compressive modulus (139 MPa).

In contrast, DOE-28, produced with a higher layer height (0.3 mm), higher printing speed (60 mm/s), lower printing temperature (195 °C), and 100% infill density, shows signs of over-extrusion, irregular filament stacking, and less consolidated junction regions in Figure 9d–f. The combination of thicker layers, insufficient thermal bonding, and excessive material accumulation leads to structural discontinuities and stress concentration sites within the lattice. Consequently, DOE-28 demonstrates a significantly lower compressive strength (2.66 MPa) and compressive modulus (76 MPa).

3.4. Machine Learning Results

3.4.1. Hyperparameter Optimization and Model Generalization

To accurately capture the highly non-linear stress–strain behavior of the 3D-printed structures, an eXtreme Gradient Boosting (XGBoost) regression model was employed. Initially, a base XGBoost model was trained using a standard parameter configuration (n_estimators = 200, learning_rate = 0.1, max_depth = 6). Evaluating the base model revealed a degree of overfitting: while it achieved a near-perfect training R² of 0.9999, the testing R² was 0.9368 with a root mean square error (RMSE) of 0.2937 MPa. Although the test score indicated strong generalization to unseen specimens, the extreme gap between the training and testing metrics suggested that the base model was memorizing noise within the training data.

To mitigate this memorization and explore potential improvements in predictive accuracy, a randomized hyperparameter tuning approach was executed using 5-fold cross-validation. The optimized hyperparameters for the XGBoost algorithm are summarized in

Table 6. The optimized hyperparameters for XGBoost.

ML Model	Hyperparameters	Best Values
XG Boost	Learning rate	0.1228
	Max depth	4
	Subsample	0.8586
	Colsample bytree	0.8061
	N estimators	58

. By constraining the maximum tree depth and introducing randomization through subsampling, the model was mathematically prevented from over-relying on specific training artifacts.

As presented in

Table 7. The performance evaluation of base vs. tuned XGBoost models.

Model	Train R2	Test R2	Train RMSE	Test RMSE	Train MAE	Test MAE
Base model	0.9999	0.9368	0.012	0.2937	0.0082	0.1788
Hyperparameter tuning	0.9972	0.929	0.0586	0.3113	0.0435	0.1847

, the hyperparameter tuning process successfully reduced the model’s overfitting. The training R² was lowered to 0.9972, and the training RMSE was increased to 0.0586 MPa, indicating the model was no longer memorizing the data perfectly. On the unseen testing set, the tuned model achieved an R² of 0.9290, an RMSE of 0.3113 MPa, and a mean absolute error (MAE) of 0.1847 MPa. While the tuning successfully narrowed the gap between the training and testing metrics, proving the regularization worked, the overall predictive accuracy remained highly comparable to the base model. This confirms that XGBoost configurations are highly robust, successfully capturing approximately 93% of the variance in the unseen experimental stress–strain curves across the relevant testing regimes.

3.4.2. Validation of Key Mechanical Properties

To see how well the machine learning framework actually works in practice, we tested its predictions against the real-world experimental results of three completely unseen specimens (Specimens 1, 9, and 17).

As shown by the tight clustering in the parity plots in Figure 10. and the exact values in

Table 8. The comparison of the experimental vs. the ML-predicted mechanical properties.

Specimen No.	Compressive Strength (Exp)	Compressive Strength (ML)	Error (%)	Compressive Modulus (Exp)	Compressive Modulus (ML)	Error (%)
1	3.56	3.32	6.74	110	112	1.82
9	3.35	3.26	2.69	125	116	7.2
17	2.80	2.86	2.14	91.44	84	8.14

, the XGBoost model proved highly capable of predicting macro-mechanical properties using nothing but the initial 3D printing parameters. When predicting the compressive strength, the model’s estimates were remarkably close to the physical tests, keeping the error margin strictly between 2.14% and 6.74%. It was just as precise when predicting the compressive modulus, with error rates staying extremely low, between 1.82% and 8.14%.

These low error rates are a strong indicator that the model has not just memorized the training data but has genuinely learned the physical relationship between the printing parameters and the mechanical behavior of the material. Ultimately, this proves that our framework can serve as a highly reliable, non-destructive tool. It can confidently predict how well an additively manufactured lattice will perform and absorb energy, which could drastically cut down the time, material waste, and cost of physical testing.

4. Future Perspectives

This research investigated the compressive behavior of a lattice wood–PLA composite fabricated using an FDM 3D printing machine. Although the study incorporated artificial intelligence approaches to enhance the material performance and included microscopic analysis, several scientific gaps remain that can be addressed in future work.

The study primarily compared the compressive strength and the compressive modulus. While these comparisons are appropriate and effective for evaluating wood–PLA lattice structures, future research could explore the correlation between the manufacturing parameters and the mechanical performance using statistical methods such as ANOVA, regression analysis, and multi-objective optimization. Such investigations may provide deeper insights into how printing conditions influence structural integrity.

The mechanical properties were compared and analyzed through microscopic analysis to understand how manufacturing defects are linked to the deterioration of strength. However, further studies could be conducted to gain a deeper understanding of their effects on the outcomes through scanning electron microscopy (SEM) analysis.

This research focused on the compressive behavior of wood–PLA composite simple cubic lattice structures, particularly the compressive strength and the compressive modulus. However, the porosity is also an important parameter that can significantly influence the mechanical properties of lattice structures. Therefore, future studies will investigate the influence of manufacturing parameters on the porosity of wood–PLA composite lattice structures and their relationship with mechanical performance.

In addition to porosity, future research could focus on the tensile, flexural, impact, and fatigue resistance, as well as the surface roughness of wood–PLA composites, particularly for lattice structures. Different machine learning approaches could also be applied to further improve the mechanical properties and the overall performance of these structures.

This research only focused on the XGBoost model prediction of compressive properties. However, future research could be conducted by considering other machine learning models and comparing them with the existing literature to improve prediction efficacy and to better understand the compressive performance of different types of materials, particularly for the lattice structures.

Although suitable mechanical strength was achieved, the applicability of this lattice composite for specific engineering fields remains unclear. Future studies should identify the potential real-life applications, including the types of loads and environments in which this material can be effectively utilized.

In addition, the sustainability aspects of the wood–PLA lattice composite were not examined. Important factors such as recyclability, renewability, and environmental impact require further investigation before the material can be proposed for large-scale commercial applications.

This study was limited to experimental analysis; however, future work could integrate numerical approaches such as finite element modeling to better predict the failure mechanisms and improve the lattice design.

Finally, only one machine learning model was employed in this research. Further improvements could be achieved by implementing advanced deep learning techniques to optimize configuration, enhance service life prediction, and provide industries with a clearer understanding of how different processing parameters affect mechanical properties.

5. Conclusions

Natural fiber-reinforced composites (NFRCs) have attracted significant attention in engineering industries due to their sustainability and favorable mechanical properties. This study comprehensively investigated the effect of FDM printing parameters on the compressive behavior of wood–PLA composite lattice structures, with particular emphasis on the compressive strength and the compressive modulus. In addition, a machine learning model was employed to enhance the prediction of stress–strain behavior and improve the accuracy of mechanical performance estimation. The key findings of this research are summarized as follows:

A moderate layer height (0.2 mm) and infill density (85%), as observed in DOE-1, DOE-4, DOE-10, DOE-14, DOE-16, and DOE-18, produced more stable stress–strain curves compared to other design configurations. However, lower nozzle temperatures and lower printing speeds also resulted in an improved mechanical performance of the wood–PLA lattice composites.

The highest compressive strengths of 4.00 MPa and 4.11 MPa were achieved in DOE-15 and DOE-27, respectively. These specimens were printed with a small layer height (0.1 mm), nozzle temperatures ranging from 195 to 215 °C, low printing speed (40 mm/s), and relatively low infill density (70%). However, the lowest compressive strength and modulus were observed in DOE-28, which was printed with a larger layer height (0.3 mm), nozzle temperature of 195 °C, higher printing speed (60 mm/s), and full infill density (100%).

Validation using unseen test specimens confirmed the ML algorithm’s practical reliability. The prediction error for compressive strength ranged between 2.14 and 6.74%, while the compressive modulus errors ranged between 1.82 and 8.14%. These low discrepancies demonstrate that the tuned XGBoost model can reliably estimate both strength and stiffness directly from the printing parameters.

The microscopic analysis revealed important morphological features of the lattice composites, including fiber–matrix adhesion quality, interlayer gaps, material distribution uniformity, and fracture regions. The microstructural observations further validated the findings obtained from the DOE analysis.

In most cases, the specimens printed with a low layer height (0.1 mm), low printing speed (40 mm/s), moderate nozzle temperature (185–195 °C), and infill density ranging from 70% to 85% exhibited an improved fiber–matrix adhesion, reduced interlayer gaps, and more uniform material distribution within the lattice structure. Moreover, these optimized parameter combinations corresponded to the higher compressive strength and compressive modulus values, indicating a strong relationship between microstructural integrity and mechanical performance.

This study investigated the compressive behavior of sustainable wood–PLA composite lattice structures fabricated using FDM, incorporating one popular AI-based ensemble machine learning method. The investigation provided a thorough analysis of the effectiveness of key printing parameters on the compressive properties of the lattice composites. The findings offer valuable insights for manufacturers and researchers by supporting the adoption of this systematic approach into decision-making processes and improving the understanding of how processing parameters influence the mechanical performance of wood–PLA lattice structures. Furthermore, due to their sustainability, low cost, and acceptable compressive strength, these wood–PLA FDM-printed lattice composites can be considered for engineering applications involving low-load requirements, particularly within the compressive strength range of approximately 2.0–4.2 MPa.