Comparative Performance of Bootstrap-Optimized Machine Learning Models for Predictive Modeling of Industrial LDPE Production

Abstract

This paper presents a bootstrap-enhanced machine-learning framework for predictive modeling of industrial low-density polyethylene (LDPE) production. The high-pressure LDPE tubular reactor is a nonlinear and thermally sensitive process in which operating variables strongly affect reaction temperature, monomer conversion, and melt flow index. A mechanistic reactor model was first used to generate a multiple-input multiple-output dataset through sensitivity analysis of key operating parameters. Three supervised learning models, Gaussian Process Regression (GPR), Support Vector Machine (SVM), and Decision Tree regression, were trained and optimized using bootstrap resampling and grid-search hyperparameter tuning. Model performance was evaluated using the coefficient of determination for three product and process indicators. GPR gave the most balanced performance across all outputs, with R² values of 0.9227 for temperature, 0.9261 for monomer conversion, and 0.9246 for melt flow index. SVM produced slightly higher R² values for temperature and conversion but performed poorly for melt flow index. Decision Tree regression was less accurate overall but showed acceptable performance for melt flow index. The results indicate that bootstrap-based hyperparameter optimization improves predictive robustness and supports uncertainty-aware surrogate modeling for LDPE process analysis. The revised framework provides a practical basis for future optimization, monitoring, and decision support in high-pressure polymerization systems.