Prediction and Explainable Analysis of Molecular Weight Distribution of Polystyrene Based on Machine Learning and SHAP

Abstract

Molecular weight distribution (MWD) is crucial for the product performance of polymers. In order to explore how process conditions affect molecules with different chain lengths, this study conducts a large number of polystyrene process simulations based on polymerization kinetics and validates them through the pilot plant data to generate a reliable dataset. Machine learning methods are employed to predict average molecular weights and conversion rates. Compared to extreme gradient boosting (XGBoost) and support vector regression (SVR), the fully connected neural network (FCNN) shows the best performance. Furthermore, an improved FCNN model with feature extractor and residual structure is developed to predict MWD accurately. The polymer molecules are divided into 10 bins based on chain length, and the influence of process conditions is revealed through SHapley Additive exPlanations (SHAP). Notably, reducing the feed mass fraction of ethylbenzene and increasing the charging coefficient of the second pre-polymerization reactor will lead to an increase of low molecular weight polymers. Raising the temperature of the second pre-polymerization reactor will promote a decrease in the proportion of small molecule polymers and ultra-large molecule polymers, thereby narrowing MWD. In addition, process conditions for polystyrene with specific target MWD can be effectively predicted by machine learning.