Data-Driven and Expert-Based Comparative Feature Analysis for Polyurethane Foam Formulation

Accurately identifying the most influential chemical constituents is essential for optimizing the quality of polyurethane foams (PUFs), particularly in terms of density, cell morphology, and overall performance. This study integrates statistical, machine learning, and expert-informed techniques to evaluate feature importance for PU foam density prediction. A commercial dataset containing over 20,000 observations from a batch-slab stock PU foam facility was analyzed across seven chemical input variables. Feature selection methods included variance inflation factor (VIF) to assess multicollinearity, recursive feature elimination (RFE) integrated with support vector regression (SVR) using a radial basis function (RBF) kernel to rank features based on predictive power, and opinion discriminative analysis (ODA) to incorporate expert judgment. Fifty (50) industry professionals evaluated the features using a 5-point Likert scale, and their responses were analyzed to compute discriminative power (Dp) scores. Spearman’s rank correlation coefficient (ρ) was employed to assess alignment among the models. RFE+SVR and ODA, RFE+SVR and VIF, all showed weak agreement (ρ = 0.178), while VIF strongly correlated positively with ODA (ρ = 1). This study highlights the value of combining statistical methods, algorithm-driven techniques, and expert insights to achieve more dependable feature selection in the production of polyurethane (PU) foam.