Assessing Dynamic Viscosity of Polyethylene Glycol (PEG): Sensitivity Analysis and AI-Based Modeling

In this study, the hyperparameters of a gradient boosting decision tree (GBDT) machine learning model are meticulously fine-tuned using four distinct evolutionary optimization approaches: Evolutionary Strategies (ES), Bayesian Probability Improvement (BPI), Batch Bayesian Optimization (BBO), and Self-Adaptive Differential Evolution (SADE) to predict (polyethylene glycol) PEG viscosity. Analysis of the correlation matrix indicates that pressure and molecular weight have influence on dynamic viscosity. Pressure displays a negligible positive correlation (0.17), while molecular weight shows a positive correlation (0.24). On the other hand, temperature exhibits an inverse relationship (−0.75) with dynamic viscosity, establishing it as the most critical factor affecting viscosity. Thus, temperature significantly impacts dynamic viscosity, whereas molecular weight and pressure contribute marginally. These observations are essential for comprehending system behavior and enhancing process efficiency. Among the optimization techniques, SADE outperforms the others, yielding the most accurate GBDT-based hybrid predictive model, as evidenced by performance metrics, including R-squared, mean squared error (MSE), and average absolute relative error (AARE). Despite its extended runtime, SADE's exceptional precision, reflected in the lowest MSE and AARE. Sensitivity analysis confirms that all input variables affect the target parameter, with SHapley Additive exPlanations (SHAP) analysis highlighting temperature as a dominant factor influencing PEG viscosity.