In a paper published in the journal Decision Analytics Journal, researchers explored ensemble machine learning (ML) models for predicting mechanical properties of 3D-printed polylactic acid (PLA) specimens, assessing five process parameters: build orientation, infill angle, layer thickness, printing speed, and nozzle temperature using data from 27 specimens.
Study: Ensemble Machine Learning Enhances 3D Printing. Image Credit: Nordroden/Shutterstock
Machine learning (ML) models, including gradient boosting regression (GBR), extreme gradient boosting regression (XGBR), adaptive boosting regression (ABR), random forest regression (RFR), and extremely randomized tree regression (ERTR), were employed.
The ERTR excelled in predicting tensile strength, while RF regression performed well in predicting surface roughness. Ensemble methods surpassed traditional models, proving effective in optimizing product design and mechanical properties through precise parameter adjustments.
Background
Past work has shown that the quality of 3D-printed parts in fused deposition modeling (FDM) is significantly influenced by process parameters such as layer thickness, printing speed, extrusion temperature, deposition direction, and build orientation.
Traditional manufacturing methods require multiple stages and are time-consuming compared to additive manufacturing, which can produce complex designs in a single step with minimal tooling. Extrusion-based FDM, known for its accessibility and affordability, is crucial in creating prototypes and functional parts. However, optimizing mechanical properties and surface quality remains challenging due to the complex interactions between these parameters.
Optimizing 3D Printing Parameters
The study utilized the American Society for Testing and Materials (ASTM) D638-14 Type I specimens for 3D printing with polylactic acid (PLA) material on a Prusa i3 MK-3 printer. Various parameters such as build angle, infill angle, layer thickness, printing speed, and nozzle temperature were systematically varied across 27 experimental combinations using a Taguchi L27 orthogonal array. This approach ensured comprehensive parameter space coverage to evaluate their impacts on surface roughness and tensile strength. ML models, including GBR, XGBR, ABR, RFR, and an ERTR, were developed to predict these mechanical properties based on the experimental data.
Each machine learning model underwent rigorous optimization using grid search methods to fine-tune hyperparameters such as learning rates and tree depths. The models were evaluated based on performance metrics, including root mean squared error (RMSE), mean absolute error (MAE), and mean absolute percentage error (MAPE). Results indicated that XGBR and ABR models generally outperformed GBR, RFR, and ERTR in predicting surface roughness and tensile strength, with variations observed based on training data size.
The findings underscore the effectiveness of ensemble learning techniques in capturing the complex relationships between process parameters and mechanical properties in 3D printed PLA specimens, offering insights for optimizing print settings to enhance part quality and performance.
Ensemble ML in 3D Printing
The study employed RFR and ERTR models to predict surface roughness and tensile strength in 3D-printed PLA specimens. RFR utilized 1000 trees without setting a maximum depth, while ERTR also used 1000 trees with unrestricted depth. The models were rigorously evaluated across training and test data splits (80-20 and 70-30), showing that larger training datasets generally improved prediction accuracy.
For instance, RFR achieved an RMSE of 0.4080 and 1.3863 for surface roughness and tensile strength, respectively, with 80% training data, compared to 0.6781 and 2.6095 with 70% training data. Similarly, ERTR demonstrated enhanced performance metrics with larger training datasets, highlighting the significance of data volume in model training.
The comparison among ensemble machine learning models revealed distinct performance characteristics. The XGBR and RFR models consistently outperformed others in predicting surface roughness, while ERTR excelled in predicting tensile strength. For both properties, models trained with 30% test data generally showed superior accuracy, indicating their robustness in capturing intricate relationships between process parameters and mechanical outcomes.
This comparative analysis underscores the effectiveness of ensemble techniques in optimizing FDM printing parameters for enhanced part quality and mechanical performance in PLA components. Furthermore, the interpretability of ensemble models provided valuable insights into parameter influences on prediction outcomes.
Visualizations of decision-making processes within Random Forests and ERTR models elucidated the criticality of factors like build angle and layer thickness in determining mechanical properties and surface finish. This transparency enhances process optimization by enabling informed adjustments to printing parameters, thereby supporting cost reduction, productivity enhancement, and environmental sustainability in additive manufacturing practices.
Conclusion
To sum up, the study applied ensemble ML models to predict the tensile strength and surface roughness of PLA specimens fabricated through FDM. The evaluation focused on five critical process parameters—build angle, infill angle, layer thickness, printing speed, and nozzle temperature—using GBR, XGBR, ABR, RFR, and ERTR.
Results indicated that ensemble models, particularly XGBR for surface roughness and ERTR for tensile strength, consistently outperformed traditional methods like k-nearest neighbors (KNN) and support vector machine (SVM) in predictive accuracy. These models exhibited superior performance metrics such as mean absolute and mean squared errors, highlighting their efficacy in capturing intricate dataset correlations and delivering precise predictions for FDM-printed PLA samples.
Looking ahead, while the study focused on five primary process parameters, future research could expand to include additional variables such as material type, cooling rate, and nozzle diameter. Exploring alternative ML methodologies such as recurrent neural networks (RNNs) and convolutional neural networks (CNNs) or integrating Bayesian Optimization could enhance predictive accuracy.
Incorporating domain-specific knowledge and physics-informed ML approaches could also improve model reliability and interpretability, offering deeper insights into material behavior under various printing conditions. These advancements are crucial for advancing mechanical property prediction in additive manufacturing, paving the way for more informed process optimizations and enhanced product quality in industrial applications.
Machine Learning Models for Turbulent Combustion Speed Prediction