An Explainable Ensemble Learning Approach with CMPI Bridging SBSp Process Parameters and EVOH Microfiltration Membrane Performance

Abstract Solution blow spraying (SBSp) provides a strategy for fabricating nanofibrous membranes. However, SBSp membranes involve coupled fabrication variables, making trial-and-error optimization inefficient. In this study, a data set comprising fiber diameter, dispersion concentration, spray volume, spray flow rate, nozzle-substrate distance, microsphere retention rate, and membrane flux was established. Then, a comprehensive membrane performance index (CMPI) was proposed by equally weighting normalized retention and flux. Subsequently, a Bayesian-optimized weighted ensemble model (BO-RF-SVR) was developed and achieved high predictive accuracy, with test-set R2 values of 0.9734 and 0.9894 for microsphere retention rate and membrane flux, respectively. SHAP and 3D response-surface analyses identified fiber diameter as the most important feature, with SHAP-based feature contributions of 73%, 61%, and 66% to retention, flux, and CMPI predictions, respectively. These findings reveal the retention–permeability trade-off effect and provide an explainable ensemble learning approach for interpretable prediction of the process optimization of PP-supported EVOH microfiltration membranes.