Machine learning-assisted pattern recognition algorithms for estimating ultimate tensile strength in fused deposition modelled polylactic acid specimens

ABSTRACT

In this study, we investigate the application of supervised machine learning algorithms for estimating the Ultimate Tensile Strength (UTS) of Polylactic Acid (PLA) specimens fabricated using the Fused Deposition Modeling (FDM) process. 31 PLA specimens were prepared, with Infill Percentage, Layer Height, Print Speed, and Extrusion Temperature serving as input parameters. The primary objective was to assess the accuracy and effectiveness of four distinct supervised classification algorithms, namely Logistic Classification, Gradient Boosting Classification, Decision Tree, and K-Nearest Neighbor, in predicting the UTS of the specimens. The results revealed that while the Decision Tree and K-Nearest Neighbor algorithms achieved an F1 score of 0.71, the KNN algorithm exhibited a higher Area Under the Curve (AUC) score of 0.79, outperforming the other algorithms. The findings offer valuable insights into the potential use of machine learning techniques in improving the performance and accuracy of predictive models in additive manufacturing.

KEYWORDS:

• Additive manufacturing
• machine learning
• Fused Deposition Modelling
• classification algorithms

Introduction

In recent years, Artificial Intelligence (AI) has emerged as a transformative force across various industries, revolutionizing processes and driving innovation. The manufacturing and health sector is no exception, as it has experienced significant benefits from integrating AI-driven technologies [1–4]. Among the most significant benefits of AI in manufacturing is its capacity to enhance efficiency and productivity. AI-enabled systems can process enormous volumes of data in real time, allowing manufacturers to detect patterns and trends that can be harnessed for process improvement. Machine learning algorithms, a branch of AI, can evolve and refine over time, making manufacturing systems increasingly adept at forecasting equipment malfunctions and reducing downtime [5–9]. In a study conducted by Du et al. [10], the researchers examined the conditions leading to void formation in friction stir welded joints, as these voids negatively impact the mechanical properties of the joints. To investigate this phenomenon, the authors employed a decision tree and a Bayesian neural network. They analysed three types of input datasets, including unprocessed welding parameters and computed variables derived from both analytical and numerical models of friction stir welding. In a study by Roman Hartl et al. [11], the authors investigated the application of Artificial Neural Networks in analysing process data from friction stir welding to predict the quality of the resulting weld surface.

Artificial Intelligence is gaining interest in additive manufacturing industries like other industries. Du et al. [12] demonstrated that employing a synergistic approach that combines physics-informed machine learning, mechanistic modelling, and experimental data can mitigate the prevalence of common defects in additive manufacturing. By scrutinizing experimental data on defect formation for widely used alloys sourced from disparate, peer-reviewed literature, the researchers identified several crucial variables that elucidate the underlying physics behind defect formation. Maleki et al. [13] employed a machine learning-based methodology to explore the relationship between residual stress, hardness, and surface roughness (influenced by the applied post-treatments) and the depth of crack initiation sites as well as the fatigue life of post-treated additive manufactured samples. There have been other various research works that implemented machine learning in the domain of structural integrity [14–22]. Zhang et al. [23] uses three distinct machine learning techniques (Random Forest, XGBoost, and Adaboost) to predict the elastic modulus of triply periodic minimum surface (TPMS) structures for titanium, a biomedical material. Gong et al. [24] developed high prediction accuracy linear regression models. The first and second laws of thermodynamics provide a clear explanation for the decline of microstructure entropy, which is brought about by the system entering a self-similar phase during grain formation and has been further investigated through machine learning. Li et al. [25] used the gradient boosting decision tree (GBDT) machine learning method for the first time, boundaries in bainitic microstructure were automatically recognized and quantitative statistical analysis was carried out utilizing electron back-scatter diffraction (EBSD) data. Bobbili et al. [26] machine learning-based models such as Support Vector Machine (SVM), Logistic Regression, Decision Tree, Random Forest, Artificial Neural Network (ANN), and Gradient Boosting algorithm have been used to classify and estimate the phase prediction in HEAs. Zhang & Xu [27], statistical association between structural characteristics of the iron yoke and the core magnetic flux density for superconducting solenoids is explained using a Gaussian process regression model. Babu et al. [28] used the fused deposition modelling (FDM) technology to evaluate the mechanical properties of three-dimensional printed carbon fibre/polylactic acid (CF/PLA) composites. Different slicing parameters, such as layer heights or thicknesses, infill densities, and layer patterns, were used to construct the composites. Tensile, flexural, and interlaminar shear strength (ILSS) tests were performed on the 3D-printed composite samples in order to evaluate the impact of the process parameters on their mechanical properties. To assess the impact of the samples’ surface roughness, more research was done. Wu et al. [29] Graphene oxide (RGO)/polylactic acid printing filaments were employed in fused deposition modelling to create a unique gradient graphene composite consisting of three layers. Naveed [30] used thermoplastic material polylactic acid is utilized to build 3D parts at five various raster angles. Tensile properties of these parts are examined to determine the optimal raster location for creating the strongest 3D printed part.

The relationship between structural integrity and ultimate tensile strength is significant in the case of fused deposition modelled polylactic acid specimens. Structural integrity pertains to the capacity of a structure or material to endure loads and retain its form and functionality without experiencing failure. It encompasses various aspects such as strength, stiffness, durability, and resistance to deformation or breakage. Ultimate tensile strength measures the maximum stress a material can withstand before it fails under tension. It represents the peak load-bearing capability of a material and indicates its ability to resist being pulled apart or stretched. UTS is typically determined through tensile testing, where a specimen is subjected to progressively increasing tensile forces until it fractures. Regarding FDM PLA specimens, the structural integrity of the printed parts is influenced by multiple factors, including the design, print settings, material properties, and post-processing techniques. The ultimate tensile strength of the PLA specimens serves as a vital indicator of their capacity to bear loads and their resistance to tension.

This study marks the first endeavour to implement supervised machine learning classification algorithms for predicting the Ultimate Tensile Strength of Polylactic Acid specimens produced via the Fused Deposition Modelling process. We examined the applicability of four distinct supervised classification algorithms, i.e. Logistic Classification, Gradient Boosting Classification, Decision Tree, and K-Nearest Neighbour, in estimating the UTS of 31 PLA specimens, using Infill Percentage, Layer Height, Print Speed, and Extrusion Temperature as input parameters.

Problem statement

Accurately estimating the Ultimate Tensile Strength of Polylactic Acid specimens created through the Fused Deposition Modelling process is crucial for ensuring optimal performance and reliability in various applications. Traditional methods for determining UTS are labour-intensive and typically necessitate destructive testing. Consequently, there is a growing demand for a more efficient, non-destructive approach to predicting UTS by leveraging advancements in machine learning. This study aims to evaluate the accuracy and efficacy of four distinct supervised machine learning classification algorithms, i.e. Logistic Classification, Gradient Boosting Classification, Decision Tree, and K-Nearest Neighbour, in estimating the UTS of PLA specimens. Input parameters include Infill Percentage, Layer Height, Print Speed, and Extrusion Temperature. The primary challenge is to identify which algorithm, if any, exhibits superior performance in differentiating between the two classes of ultimate tensile strength within the dataset, ultimately determining the most suitable choice for classification in this research context. Furthermore, this study seeks to investigate the potential of machine learning-based classification algorithms in enhancing the performance and precision of predictive models within the additive manufacturing domain. As the first attempt to estimate the UTS of PLA specimens using these techniques, this research offers valuable insights and advances knowledge in this field.

Experimental procedure

The Fused Deposition Modelling process shown in Figure 1 works by creating three-dimensional objects layer by layer, using thermoplastic materials like polylactic acid. In this method, a computer-aided design model is prepared and converted into a compatible file format, sliced into thin horizontal layers by specialized software. These layers generate a set of instructions, or G-code, for the 3D printer to follow during the printing process. The printer’s extruder heats the PLA filament, a biodegradable material derived from renewable sources, and deposits it through a nozzle onto the build platform. As the extruder moves in the X and Y directions and the build platform moves in the Z direction, the object is formed layer by layer. The PLA material fuses with the previous layer and solidifies as it cools, creating the final 3D object. Support structures may be needed during printing for complex geometries or overhangs, and post-processing steps such as sanding or painting can be employed to achieve the desired finish. Fused Deposition Modelling samples were fabricated utilizing a Creality Ender 3 machine with a bed size of 220 × 220× 250 mm shown in Figure 2. The dimensions of the tensile specimens measured 63.5 × 9.53 × 3.2 mm, adhering to the ASTM D638 standard requirements as shown in Figure 3. The part design was created and converted into an STL file using CATIA software. The STL file was then processed into a machine-readable G-code file with the assistance of the Cura engine within Repetier software to build slicing of the file, as shown in Figure 4. In this research study, the dataset shown in Table 1 was initially converted into a CSV file format to facilitate its import into Google Colaboratory for developing machine learning-based classification algorithms using Python programming. Four distinct classification algorithms were employed for analysis, including Decision Tree, K-Nearest Neighbour, Logistic Regression, and Gradient Boosting Classifier. The material’s ultimate tensile strength was the basis for classification. If the UTS was below 80% of the base material’s UTS, it was labelled as ‘0’, while a value above 80% of the base material’s UTS was labelled as ‘1’. This labelling approach differentiated materials with relatively lower and higher tensile strengths. Experimentally obtained ultimate tensile strength of base material i.e. PLA with an average of 59.5 MPa. Thus, the target value set for classification is 47.6 MPa (which is 80% of 59.5 MPa).

Figure 1. Schematic representation of the Fused Deposition Modeling process.

Figure 2. Ender 3 3D printer.

Figure 3. Schematic sketch of tensile specimen [31].

Note: All dimensions are in mm

Figure 4. Tensile specimen a) before slicing, b) after slicing.

Table 1. Experimental dataset [31–34].

Infill percentage (%)	Layer height (mm)	Print speed (mm/sec)	Extrusion temp (${ }^{\circ }C$)	Ultimate Tensile Strength (MPa)	Classification Label
78	0.32	35	220	46.17	0
10.5	0.24	50	210	42.78	0
33	0.16	35	220	45.87	0
33	0.32	35	200	41.18	0
33	0.16	65	200	43.59	0
100	0.24	50	210	54.2	1
78	0.16	35	200	51.88	1
33	0.32	65	200	43.19	0
78	0.32	65	200	50.34	1
33	0.16	65	220	45.72	0
78	0.16	35	220	53.35	1
55.5	0.24	50	210	49.67	1
33	0.32	35	220	45.08	0
55.5	0.24	50	190	47.56	0
55.5	0.24	50	210	48.39	1
78	0.32	65	220	46.49	0
55.5	0.24	50	210	47.21	0
55.5	0.24	50	210	48.3	1
55.5	0.24	50	230	50.15	1
33	0.32	65	220	43.35	0
55.5	0.24	50	210	45.33	0
55.5	0.24	80	210	45.56	0
78	0.16	65	200	49.84	1
55.5	0.24	20	210	48.51	1
55.5	0.08	50	210	42.63	0
55.5	0.4	50	210	42.87	0
55.5	0.24	50	210	47.14	0
78	0.32	35	200	45.17	0
55.5	0.24	50	210	47.07	0
78	0.16	65	220	50.99	1
33	0.16	35	200	43.75	0

In present study, with 31 dataset 5-fold cross-validation was used and the dataset was divided into 5 subsets. This allowed to train and test the model five times, which resulted in a balance between variance reduction and computational efficiency.

To evaluate and compare the performance of these classification models, two key metrics were considered: the area Under the Receiver Operating Characteristic (ROC) Curve (AUC-ROC) and the F1 score. The AUC-ROC score measures the classifier’s ability to discriminate between the two classes, with a higher score indicating better performance. On the other hand, the F1 score represents the harmonic mean of precision and recall, providing a balanced evaluation of the model’s accuracy in terms of both false positives and false negatives. By comparing the AUC-ROC and F1 scores of the four classification algorithms, this research aims to identify the most suitable algorithm for predicting the ultimate tensile strength of materials based on the given dataset, ultimately contributing to a better understanding of material properties in additive manufacturing.

In case of logistic classification model the hyper parameters set are regularization strength of 0.1, type of regularization set as Lasso, solver algorithm as liblinear. For getting better precession in case of gradient-boosting classification model Learning Rate of 0.1, number of trees as default values, maximum depth of 5, subsample of 0.7, max features equal to square root of the total number of features, min samples split as default value, min samples leaf as default value. For getting better precession in case of decision tree classification model maximum depth of 5, max features equal to square root of the total number of features, min samples split as default value, min samples leaf as default value. In case of KNN the number of neighbours considered was 5 and Euclidean distance metrics was considered for the study.

Results and discussion

Metric features used in the present work

A confusion matrix serves as an essential evaluation tool for classification algorithms. It is a table that compares the true labels of a given set of test data with the predicted labels generated by the algorithm, as shown in Figure 5. The matrix consists of two rows and two columns, with the rows indicating the true labels and the columns representing the predicted labels. The four cells of the matrix reveal the number of instances that fall into each possible combination of true and predicted labels. The diagonal cells of the confusion matrix represent the number of instances where the predicted label matches the true label. In contrast, the off-diagonal cells signify the instances where the predicted label differs from the true label. This provides insight into the algorithm’s performance, including the true positive rate (TPR), false-positive rate (FPR), precision, recall, and F1 score.

Figure 5. Nomenclature of confusion matrix.

The TPR refers to the ratio of true positives out of all positive instances in the dataset shown in Equation 1. At the same time, the FPR represents the ratio of false positives out of all negative instances in the dataset shown in Equation 2. Precision is the ratio of true positives out of all predicted positives as shown in Equation 3, whereas recall is the ratio of true positives out of all actual positives as shown in Equation 4. The F1 score denotes the harmonic mean of precision and recall, as shown in Equation 5, and can be used to evaluate the overall performance of the classification algorithm.

(1) $TPR=\frac{TP}{TP+FN}$(1)

(2) $FPR=\frac{FP}{FP+TN}$(2)

(3) $Precision=\frac{TP}{TP+FP}$(3)

(4) $Recall=\frac{TP}{TP+FN}$(4)

(5) $F1-Score=2\times \frac{Precision\times Recall}{Precision+Recall}$(5)

The Receiver Operating Characteristic curve serves as a graphical evaluation method for binary classification models, illustrating the relationship between the true positive rate and the false-positive rate (FPR, or 1-specificity) across a range of decision thresholds. The Area Under the Curve is a scalar metric that quantifies the overall performance of the classifier by measuring the area beneath the ROC curve. Constructing the ROC curve involves plotting TPR against FPR for varying decision thresholds. To achieve this, classifier output probabilities are arranged in descending order, and the decision threshold is shifted from the highest to the lowest probability. For each threshold, TPR and FPR are calculated and plotted as a point on the ROC curve. The AUC metric is computed as the area beneath the ROC curve, with a range of 0 to 1, where a higher value signifies superior classifier performance. An AUC of 0.5 corresponds to a random classifier, while an AUC of 1 implies a flawless classifier. The AUC can be determined through trapezoidal or rectangular approximation techniques. The trapezoidal method entails summing the areas of trapezoids formed by consecutive points on the ROC curve depicted in Equation 6.

(6) $AUC={\sum }_{i=1}^{N-1}\frac{FPR\left(i+1\right)-FPR\left(i\right)\times \left(TPR\left(i+1\right)+TPR\left(i\right)\right)}{2}$(6)

Now, let us discuss the obtained values of the metric features by individual algorithms in the next subsection.

Logistic classification

Logistic classification is a way to predict whether something belongs to one category or another based on a set of features. In the present study, it has been used to predict whether the UTS of the additive manufactured specimen is greater or less than 80% of the UTS of the based material, as shown in Equation 7.

(7) $UTSofadditivemanufacturedspecimens\left(y\right)=\text{break}\left\{\begin{array}{c}0ify<80\mathrm{\%}oftheUTSofbasematerial\\ 1ify>80\mathrm{\%}oftheUTSofbasematerial\end{array}\right.$(7)

The input features are x₁ for Infill density, x₂ for Layer Height, x₃ for Print Speed, and x₄ for Extrusion Temperature. The UTS of the fabricated specimen can be represented (x₁, x₂, x₃, x₄).

The logistic classification model uses Equation 8 to make its predictions.

(8) $P\left(y=1|x_1,x_2,x_3,x_4\right)=\frac{1}{1+e^{-\left(w_0+w_1{x}_1+w_2{x}_2+w_3{x}_3+w_4{x}_4\right)}}$(8)

Where $P\left(y=1|x_1,x_2,x_3,x_4\right)$ represents the probability that the UTS of additive manufactured specimen belongs to the category labelled as 1, the w₀, w₁, w₂, w₃, and w₄ are the parameters of the model that is expected to learn from the given training data. The best values for w₀, w₁, w₂, w₃, and w₄ need to be found out that will make the model as accurate as possible. The cost function depicted in Equation 9 makes these parameters learn.

(9) $J\left(w_0,w_1,w_2,w_3,w_4\right)=\frac{1}{m}{\sum }_1^m-y\left(i\right).log\left(h\left(x\left(i\right)\right)\right)-\left(1-y\left(i\right)\right).log\left(1-\left(x\left(i\right)\right)\right)$(9)

Where $J\left(w_0,w_1,w_2,w_3,w_4\right)$ is the cost function that needs to be minimized, m is the number of training data provided, y(i) corresponds to the binary output for the i-th specimen, and h(x(i)) is the predicted probability of the i-th specimen having a UTS greater than or equal to 80% of the base material, based on the current values of w0, w1, w2, w3, and w4. Gradient descent is used iteratively to adjust the values of w₀, w₁, w₂, w₃, and w₄ to minimize the cost function $J\left(w_0,w_1,w_2,w_3,w_4\right).$ After that, the trained logistic regression model is used to predict the UTS of new specimens based on their infill percentage, layer height, print speed, and extrusion temperature. Figure 6 shows the obtained confusion matrix, and Figure 7 shows the obtained Receiver Operating Characteristic curve.

Figure 6. Confusion matrix obtained for Logistic classification algorithm.

Figure 7. ROC curve for Logistic classification.

Gradient boosting classification

The gradient boosting algorithm was used to iteratively build an ensemble of decision trees that minimized the classification error of the training data. The gradient-boosting classifier works similarly. It tries to guess the correct answer to a problem based on a set of input values. It uses many smaller decision rules to make its final prediction. Each decision rule is like a small, simple model that tries to predict the answer based on a few input values. The classifier starts by creating a first model that makes some predictions based on the input values. Then, it looks at the mistakes that the first model made and creates a second model that tries to correct those mistakes. It keeps doing this, creating many models and correcting its mistakes until it has a final model that is very good at predicting the correct answer. Each time the classifier creates a new model, it gives more weight to the input values that were difficult to predict correctly in the previous models. This helps the classifier focus on the input values that are most important for making a good prediction. Equation 10, based on Gradient Boosting classification, is used to make predictions on new specimens.

(10) $y\left(x\right)={\sum }_{i=1}^n{\gamma }_i{h}_i\left(x\right)$(10)

Where y(x) represents the predicted output (0 or 1) for a given set of input parameters x, the Sum(i = 1 to n) indicates that the contributions of each decision tree in the ensemble are summed, with ${\gamma }_i$ representing the weight assigned to each tree, $h_i\left(x\right)$ represents the output of the i-th decision tree, which depends on the values of the input parameters x. Figure 8 shows the obtained confusion matrix, and Figure 9 shows the obtained Receiver Operating Characteristic curve.

Figure 8. Confusion matrix obtained for Gradient Boosting classification algorithm.

Figure 9. ROC curve for Gradient Boosting classification.

Decision tree classification

Decision Tree Classification Algorithm asks a series of yes or no questions about the input variables to predict the output variable. The questions are organized into a tree-like structure, with the initial question at the top (the ‘root’ of the tree) and subsequent questions branching off. Each question splits the data into two groups based on the answer (e.g. infill percentage > 50% or not), and the process continues until a final prediction is made at the bottom of the tree (the ‘leaves’). The goal of the decision tree classifier is to ask questions that give the most information about the output variable with the fewest questions. This is done by selecting the best question to ask at each tree branch based on some criterion (e.g. information gain). Once the decision tree is constructed, it can predict new data by following the path from the root to the appropriate leaf node. Each leaf node corresponds to a particular value of the output variable, and the prediction is simply the value associated with the leaf node.

Let X = {x₁, x₂, …, x_n} be the set of input variables, and y be the output variable. A DTC can be represented by a tree T with a set of nodes V = {v₁, v₂, …, v_k} and edges E = {e₁, e₂, …, e_m}, where each node vi corresponds to a question about the input variables and each leaf node corresponds to a prediction of the output variable. The tree’s construction can be described using a set of splitting rules determining how to partition the data at each node. Let Q be the set of splitting rules, and let q(v) be the splitting rule at node v. Then, the tree can be constructed by recursively partitioning the data based on the splitting rules until all nodes are leaf nodes. The prediction of the DTC can be represented using a set of decision rules that determine which leaf node to assign a new input vector x to. Let R be the set of decision rules, and let r(v) be the decision rule at node v. Then, the prediction of the DTC for input vector x can be calculated by Equation 11.

(11) $\mathrm{y}=\mathrm{p}\left(\mathrm{x};\mathrm{T}\right)=\mathrm{r}\left({\mathrm{v}}_{\mathrm{j}}\right),\mathrm{i}\mathrm{f}\mathrm{x}\mathrm{s}\mathrm{a}\mathrm{t}\mathrm{i}\mathrm{s}\mathrm{f}\mathrm{i}\mathrm{e}\mathrm{s}\mathrm{t}\mathrm{h}\mathrm{e}\mathrm{c}\mathrm{o}\mathrm{n}\mathrm{d}\mathrm{i}\mathrm{t}\mathrm{i}\mathrm{o}\mathrm{n}\mathrm{s}\mathrm{o}\mathrm{f}\mathrm{t}\mathrm{h}\mathrm{e}\text{break}\mathrm{d}\mathrm{e}\mathrm{c}\mathrm{i}\mathrm{s}\mathrm{i}\mathrm{o}\mathrm{n}\mathrm{r}\mathrm{u}\mathrm{l}\mathrm{e}\mathrm{r}\left({\mathrm{v}}_{\mathrm{j}}\right)\mathrm{a}\mathrm{t}\mathrm{n}\mathrm{o}\mathrm{d}\mathrm{e}{\mathrm{v}}_{\mathrm{j}}$(11)

where v_j is the leaf node that x is assigned to based on the decision rules.

In the present work, the DTC shown in Figure 10 is constructed with the following hyperparameters: criterion = ‘entropy’, max_depth = 6, min_samples_leaf = 1, min_samples_split = 2, splitter = ‘best’. The criterion parameter specifies the quality of the split, with ‘entropy’ indicating that the information gain criterion is used. The max_depth parameter specifies the maximum depth of the tree, limiting the number of questions that can be asked. The min_samples_leaf parameter specifies the minimum number of samples required to be at a leaf node, while the min_samples_split parameter specifies the minimum number of samples required to split an internal node. The splitter parameter specifies the strategy used to choose the split at each node, with ‘best’ indicating that the best split is chosen based on the criterion. The DTC is trained on the X_train and y_train data using the fit method, allowing it to learn the patterns in the data and construct an appropriate decision tree. Figure 11 shows the obtained confusion matrix, and Figure 12 shows the obtained Receiver Operating Characteristic (ROC) curve.

Figure 10. Decision tree architecture obtained in the present work.

Figure 11. Confusion matrix obtained for decision tree classification algorithm.

Figure 12. ROC curve for decision tree classification.

K-Nearest neighbours classification

The KNN algorithm involves calculating the distance between the new specimen and each of the n training samples and selecting the K samples with the smallest distances to the new specimen. The value of K is a user-defined parameter and determines how many training samples are used to make the prediction. The distance between the new specimen and a training sample can be calculated using a metric such as Euclidean distance. Once the nearest K training samples have been identified, the predicted label for the new specimen is assigned based on the majority label among the K samples. That is, if the majority of the K nearest samples have a label of 0, then the new specimen is assigned a label of 0, and if the majority have a label of 1, then the new specimen is assigned a label of 1. Figure 13 shows the obtained confusion matrix, and Figure 14 shows the obtained Receiver Operating Characteristic curve.

Figure 13. Confusion matrix obtained for K-Nearest neighbour classification algorithm.

Figure 14. ROC curve for K-Nearest neighbour classification.

Table 2 shows the obtained values of the F1-Score of the implemented algorithms.

Table 2. Obtained F1-score of each implemented algorithm.

Algorithms	Obtained F1-Score
Logistic Classification	0.7143
Gradient Boosting Classification	0.5714
Decision Tree Classification	0.4286
K-Nearest Neighbours Classification	0.7143

The obtained results indicate that Logistic Classification and K-Nearest Neighbours Classification performed similarly, both achieving F1 scores of 0.7143. These algorithms demonstrated a strong ability to differentiate between the two classes of ultimate tensile strength in the dataset. On the other hand, the Gradient Boosting Classification algorithm, an ensemble method that combines weak learners to create a more accurate model, yielded a lower F1 score of 0.5714. This suggests that, in this particular case, the Gradient Boosting Classifier was less effective in classifying the material properties than the Logistic and KNN classifiers. Lastly, the Decision Tree Classification algorithm demonstrated the lowest performance among the tested algorithms, with an F1 score of 0.4286. This result indicates that the Decision Tree classifier’s ability to classify the material properties based on the dataset accurately could have been improved. Figure 15 shows the comparison of the AUC score of the implemented algorithms.

Figure 15. AUC score comparison of the implemented algorithms.

Considering that both K-Nearest Neighbours and Logistic Classification algorithms exhibit identical F1 scores. At the same time, KNN possesses a superior AUC score; it can be deduced that KNN is the more optimal algorithm for classification in this specific scenario. The F1 score, which represents the harmonic mean of precision and recall, provides a balanced evaluation of the model’s accuracy. In contrast, the AUC score quantifies the classifier’s capacity to differentiate between the two classes, with higher scores signifying enhanced performance. The observed higher AUC score for KNN demonstrates its increased effectiveness in distinguishing between the two classes of ultimate tensile strength within the dataset, rendering it the more favourable choice for classification in the context of this research.

Conclusion

In conclusion, this study has successfully investigated the application of supervised machine learning algorithms for estimating the Ultimate Tensile Strength of Polylactic Acid specimens fabricated using the Fused Deposition Modelling process. By preparing 31 PLA specimens and utilizing input parameters such as Infill Percentage, Layer Height, Print Speed, and Extrusion Temperature, we have assessed the accuracy and effectiveness of four distinct supervised classification algorithms: Logistic Classification, Gradient Boosting Classification, Decision Tree, and K-Nearest Neighbour. The results demonstrate that the K-Nearest Neighbour algorithm outperforms the other algorithms, achieving an F1 score of 0.71 and an Area Under the Curve score of 0.79. This highlights the superior ability of the KNN algorithm to differentiate between the two classes of ultimate tensile strength within the dataset, making it the most favourable choice for classification in this research context. As the first study to estimate the UTS of PLA specimens using machine learning-based classification algorithms, our findings provide valuable insights into the potential of these techniques for enhancing the performance and accuracy of predictive models in the additive manufacturing domain. This research paves the way for future work focused on refining these algorithms, optimizing the parameters, and expanding the application of machine learning in additive manufacturing to improve the quality and reliability of 3D-printed components.

Disclosure statement

No potential conflict of interest was reported by the author(s).

Data availability statement

The data that support the findings of this study are available from the corresponding author upon reasonable request.

Correction Statement

This article has been republished with minor changes. These changes do not impact the academic content of the article.

Additional information

Funding

The authors extend their appreciation to King Saud University for funding this work through Re-searchers Supporting Project number (RSP2023R164), King Saud University, Riyadh, Saudi Arabia.