Tuning of the tensile property of HDPE/CaCO3/PKS composite using machine learning

Abstract

Particulate palm kernel shell (PKS) and calcium carbonate (CaCO 3 ) reinforced high-density polyethylene (HDPE) composites offer advantages of environmental friendliness, high stiffness, and sustainability. However, the empirical design of these materials remains costly and time-consuming due to the complex, non-linear interplay between filler composition, particle morphology, and mechanical response. Here, we present a machine learning (ML) framework to predict the full tensile stress-strain response of PKS/CaCO 3 /HDPE composites, providing a more comprehensive representation of mechanical behaviour than single-value tensile strength prediction. Using 8,284 stress-strain data points from 8 experimental compositions derived from our previous experimental work, we engineered key features including volume fraction, interparticle spacing, surface area interaction, and strain-extension coupling to capture underlying deformation mechanisms. Nine ML algorithms were benchmarked, with Extra Trees Regressor emerging as the best-performing model. A full-feature Model A (including tensile modulus) achieved R 2 =0.9998, MAE=0.0385N/mm 2 , and RMSE=0.0621N/mm 2 on the held-out test set. A composition-only Model B (tensile modulus excluded) achieved R 2 =0.9982, MAE=0.0634N/mm 2 , and MAPE=0.46%, demonstrating that accurate prediction is achievable without any prior mechanical testing. Learning curve analysis confirmed negligible overfitting in both models, with train-CV generalisation gaps of 0.0002 (Model A) and 0.0005 (Model B). SHAP analysis revealed non-linear dependencies on tensile modulus, strain, PKS content, and interparticle spacing. This ML framework reduces the need for extensive mechanical testing, accelerates sustainable composite design, and provides interpretable structure-property insights — contributing to more efficient, data-driven development of bio-based green materials aligned with SDG9.