Comparative Study of Machine Learning Models for Optimal Prediction of Printed-Line Features in Material Extrusion Additive Manufacturing

Material extrusion (MEX), commonly known as fused deposition modeling (FDM), has become a widely adopted additive manufacturing (AM) technology owing to its low equipment cost and broad polymer compatibility. However, the geometric fidelity of the printed line often suffers from defects that compromise overall part quality. Specifically, residual edge non-uniformity degrades surface finish, while uncontrolled line width variability causes undesired gaps or overlaps that undermine mechanical performance. Therefore, ensuring an accurate line width and low edge non-uniformity is essential for advancing material extrusion toward high-precision industrial applications. In this study, a machine learning framework is proposed for the rapid prediction and analysis of printed line characteristics. Nozzle temperature, print speed, and material flow rate were considered as input process parameters. Mean line width and edge non-uniformity were taken as the target responses. Four representative machine learning algorithms (XGBoost, BPNN, GPR, and SVR) were adopted for model development. To enhance predictive accuracy, these models were optimized using Particle Swarm Optimization for automatic hyperparameter tuning. Subsequently, comparative evaluations identified GPR as the optimal predictive model. Furthermore, a SHAP-based interpretability analysis was conducted, revealing that nozzle temperature dominates line width, while the flow rate governs edge non-uniformity. Consequently, this interpretable and computationally efficient surrogate modeling approach provides a robust foundation for future closed-loop quality control and inverse process design.