Data-Driven 3D-Printed Material Data Prediction From Benchmark Specimens

Abstract Predicting the mechanical behavior of 3D-printed products requires accurate material data. However, this is challenging because the printing process significantly affects the mechanical behavior of 3D-printed parts. Consequently, filament material data from the datasheet may not accurately represent the behavior of a user’s product if the printing process of the benchmark test specimen differs from that of the user. To address these issues, we propose a data-driven method for predicting the mechanical behavior of 3D-printed parts from benchmark specimen test results. First, we generate a dataset of 10,000 points, each including filament material data, benchmark test data, 3D-printing process parameters, and the corresponding effective material data. This dataset includes features often overlooked in prior studies, such as interlayer bonding perfection and benchmark test data. Next, we train diverse single and stacked machine learning models with different input features. The statistical analysis shows that multilayered perceptron (MLP) outperforms other models in both feature extraction and downstream tasks. Finally, predictions are validated against experimental data. Statistical analysis confirms that incorporating interlayer bonding perfection, benchmark specimen data, and stacking models enhances prediction accuracy. The best-performing model, which utilizes an MLP for feature extraction and a polynomial model for downstream prediction, achieves average errors of 0.06 GPa (4%) for Young’s modulus and 1.3 MPa (5%) for tensile yield strength prediction, significantly outperforming previous methods.