Machine Learning–Based Prediction and Optimization of Mechanical Properties in Natural Fiber Composites Considering Fiber Heterogeneity

ABSTRACT

Natural fiber‐reinforced composites are promising sustainable materials; however, their mechanical performance is difficult to optimize because natural fibers exhibit significant heterogeneity depending on fiber type, morphology, chemical composition, and processing history. In this study, a machine learning‐based framework was developed to predict and optimize the tensile and flexural strengths of natural fiber composites using a dataset of 270 samples collected from internal experiments and published literature. The dataset included fiber type, fiber form, matrix type, and surface treatment conditions as input variables, and the data were divided into training and test sets using an 80/20 split. Principal component analysis and clustering analysis were employed to investigate the heterogeneous structure of the dataset. Although silhouette analysis identified eight clusters, models trained using the entire dataset showed more stable predictive performance than cluster‐specific models. Compared with Random Forest, XGBoost achieved superior predictive accuracy, with R ² values of 0.980 and 0.987 and RMSE values of 5.46 and 4.44 MPa for tensile and flexural strength prediction, respectively. SHAP analysis revealed that fiber type and surface treatment conditions were the dominant factors affecting mechanical performance. Multi‐objective optimization using NSGA‐II identified flax/epoxy composites with alkali treatment as the optimal condition, and experimental validation showed good agreement between predicted and measured flexural strength values. The proposed framework provides an effective strategy for predicting and optimizing heterogeneous natural fiber composite systems while reducing experimental trial‐and‐error.