Explainable Neural Networks for Modeling Mechanical Properties of 3D-Printed Parts Using FDM

Abstract

This article presents a comprehensive methodology to predict critical mechanical properties, such as tensile strength and elongation, of 3D-printed parts using the FDM technique, through the development of an explainable multi-output neural network model. The study considers up to eight printing parameters, including layer height, infill pattern and density, print speed, cooling speed, retraction, material, and bed temperature. Through DoE, a reduced but representative set of experiments was generated, allowing the model to identify existing relationships with only 50 samples. The model achieves strong performance, with R² values above 0.93 and low relative errors (MAE < 5.3 %, RMSE < 6.3 %, MAPE < 7.54 %) for both outputs, ensuring reliable predictions and good generalization. Through the application of explainable artificial intelligence (XAI) techniques based on SHAP values, the most influential printing parameters were identified. Infill density and material type are the most critical factors in predicting tensile strength. In contrast, layer height, material, infill density, and infill pattern were found to be essential for predicting elongation. This approach improves FDM printing of medical devices and facilitates future optimizations, reducing time and resources needed for printing parameter adjustments.