Machine Learning-Driven Prediction of Polypropylene Alloy Properties: A Bayesian Network–XGBoost Framework Robust to Missing Data

Abstract

Polypropylene alloys exhibit a complex multiscale structure, posing challenges for traditional analytical methods in uncovering the relationships between structural and mechanical properties. To address this, a novel machine learning framework that integrates graphical modeling with regression techniques is proposed to enhance interpretability and predictive accuracy. Specifically, a Bayesian network is constructed to elucidate the dependencies among structural variables, guiding feature selection in the regression stage, and a hierarchical XGBoost regression model based on this network (BNXGB) is developed to predict impact strength (IS) and flexural modulus (FM). Compared to a conventional XGBoost model (XGB), BNXGB demonstrates superior performance, achieving higher R2 score for both IS and FM. More importantly, BNXGB handles missing input features by leveraging the network to impute them from parent nodes, thereby maintaining robust performance under incomplete data. Even under the extreme condition when all features except the four root nodes are missing, BNXGB still preserves reasonable R2 scores for both IS and FM, whereas XGB's performance drops sharply. These results demonstrate the robustness and practical applicability of the BNXGB framework for polymer property prediction under missing data, potentially reducing the need for exhaustive material characterization.