Comparative study of interpretable and black-box machine learning for modeling mechanical and tribological properties of 3D-printed PLA/date pits composites

Abstract

3D-printed polymers have been applied in various fields. Machine learning (ML) has revolutionized material science by providing powerful tools for modeling complex properties. This study investigates the comparative performance of interpretable and black-box ML techniques in modeling the mechanical and tribological properties of 3D-printed PLA composites blended with varying amounts of date pit particles. While neural networks (NNs) can model complex input-output relationships with high accuracy, they function as black-box models, limiting the understanding of their predictions. To address this limitation, we propose Sequential-thresholded Least-squares Sparse Regression (SLSSR), an interpretable modeling approach. SLSSR constructs models using no more than five basis functions and achieves a mean absolute error of less than 2%, ensuring both accuracy and model transparency. Moreover, SLSSR outperforms NN, delivering better accuracy and a reduced standard deviation in predictions, particularly with smaller training datasets. These results demonstrate SLSSR’s effectiveness and potential as a reliable tool for material science applications, especially in data-limited scenarios.