Enhancing bending load prediction of CaCO₃-filled polypropylene composites via data augmentation and ensemble machine learning

Abstract

Predicting the mechanical properties of calcium carbonate (CaCO₃)-filled polypropylene (PP) composites is crucial for their industrial applications. However, conventional experiments are costly and time-consuming, resulting in limited data and challenges for accurate modeling under small-sample conditions. This study proposes a machine-learning–based framework to accurately predict the bending performance of CaCO₃-filled PP composites using limited experimental data. An integrated analytical methodology, combining orthogonal experimental design, cubic spline interpolation, and machine learning techniques, is developed to achieve this objective. An orthogonal experimental design is first conducted to obtain an initial dataset of 32 samples. Subsequently, cubic spline interpolation is applied to expand the dataset to 63, 125, and 249 samples. Three machine learning models—Support Vector Machine (SVM), Random Forest (RF), and eXtreme Gradient Boosting (XGBoost)—are constructed and optimized using grid search and five-fold cross-validation. The results demonstrate that data augmentation significantly improves the predictive performance of the models. Among them, the optimized XGBoost model trained on the 249-sample dataset achieves the best performance, with a coefficient of determination (R 2 ) of 0.8061 and the lowest mean squared error (MSE). Furthermore, correlation analysis reveals that the particle size distribution is the most influential factor affecting the bending load of the composites. Overall, this study presents an integrated framework based on cubic spline interpolation and ensemble machine learning methods, providing a viable approach for estimating the mechanical properties of polymer composites under small-sample cases.