Explainable machine learning reveals viscosity–speed synergy governing thickness in pilot-scale NIPS water-treatment membrane manufacturing

Predicting membrane thickness during pilot-scale non-solvent induced phase separation (NIPS) manufacturing remains difficult because process variables interact under continuous roll-to-roll operating conditions. In this study, 760 pilot-scale manufacturing records were analyzed using machine learning models and physically motivated interaction features. Eighteen models were compared, and six interaction terms related to casting dynamics were introduced to improve process interpretation. Gradient boosting regression with 500 estimators (GBR-500) gave the best overall performance, with R² = 0.7355 and MAE = 6.70 μm under random 5-fold cross-validation. Additional grouped and time-aware validations were used to examine robustness under sequential production conditions. SHAP analysis indicated that the viscosity×speed interaction was the dominant engineered feature, suggesting that coupled casting dynamics contribute more strongly to thickness variation than individual variables alone. PDP/ICE and SHAP interaction analyses further indicated that solution concentration modulates the viscosity×speed response. A simplified five-feature model retained 95.7% of the full-model R², supporting practical use of a reduced process-variable set. These results suggest that interaction-aware machine learning can provide useful interpretation of pilot-scale NIPS manufacturing data and may support thickness-control strategies within the sampled operating range.