Predicting the tensile properties of wood plastic composites using material extrusion with meta-based Few-Shot learning

Zhuo Zeng , Yan Zhou, Shifeng Wen , Cheng Zhou

Abstract

This study investigated the optimization of process parameters affecting the tensile properties of wood plastic composites (WPC) produced by material extrusion (MEX). In the field of MEX, determining the optimal process parameters is complex, and a lack of sufficient data constrains the application of existing applied machine learning methods. In this study, a meta-based few-shot regression framework involving semi-regression, meta-learning, and fine-tuning, was developed to accurately predict the tensile properties of MEX-formed WPC while avoiding overfitting. A small dataset with varying wood contents, printing temperatures, and layer thicknesses were used to predict Young’s modulus, tensile strength, and elongation at break, which demonstrated a 42.6%, 41.9%, and 40.3% improvement in accuracy, respectively, over multilayer perceptron (MLP) when applying the proposed method. The proposed method also demonstrated improved generalization and was validated against the MLP technique. The study concluded that meta-learning can effectively manage the small datasets typical in additive manufacturing, providing a robust tool for optimizing three-dimensional printing parameters.

Keywords

Material extrusion;Wood–plastic composites;Meta-learning

;Few-shot regression

https://www.sciencedirect.com/science/article/pii/S1359835X24006699