Integrating experiments, finite element analysis, and explainable machine learning for natural fiber based hybrid composites

Abstract

The reliable prediction of mechanical properties in natural fiber–based composites is crucial for their advancement as sustainable structural materials. The development of natural fiber–based hybrid composites is often hindered by costly, resource-intensive, and time-consuming experimental iterations. While Machine Learning (ML) offers a powerful alternative for accelerating property prediction, standard models often function as 'black boxes,' which limits their trustworthiness and fails to provide the scientific insights needed for true material design. This study addresses this critical gap by developing and validating an integrated framework that combines experiments, Finite Element Analysis, and Explainable Machine learning to predict and interpret mechanical properties of banana fiber reinforced composites. Four supervised machine learning (ML) models—Random Forest, XGBoost, Gradient Boosting, and k-Nearest Neighbor—were developed and trained on a combined literature–experimental dataset. Among them, the XGBoost model exhibited the highest predictive accuracy, achieving R² values of 0.71 for tensile strength and 0.76 for flexural strength in validation. To address the “black box” challenge of ML, SHapley Additive exPlanations (SHAP) analysis was employed, revealing fiber length and filler percentage as the most influential features governing strength predictions. The integration of ML with experimental and numerical validation demonstrates a powerful framework for reducing trial-and-error in composite design, while providing interpretable insights into the role of key material parameters. This work establishes explainable ML as an effective and sustainable tool for optimizing natural fiber–based hybrid composites.