Advanced evaluation of performance of machine learning models for soapstock splitting optimisation under uncertainty

This study proposes a computational framework for the prediction and optimisation of soapstock splitting under conditions of limited measurement data and input uncertainty. The objective was to evaluate and select the modeling approaches based on (i) data availability, (ii) model complexity, (iii) predictive accuracy, and (iv) sensitivity to input uncertainty. Machine learning algorithms - Extreme Gradient Boosting (XGBoost) and Support Vector Machines (SVM) - were assessed in comparison with Response Surface Methodology (RSM). XGBoost provided the most accurate predictions for chemical oxygen demand (COD) and organic phosphorus (P_org), while SVM performed best for acid number (AN). K-means clustering identified specific input domains where RSM models could effectively substitute for XGBoost, offering a balance between simplicity and performance. GSA showed that the key influence on P_org (organic phosphorus), COD (chemical oxygen demand) and AN (acid number) was the phosphorus content of the oil, and less important were the operational parameters of the soapstock splitting system. Multi-criteria optimisation under uncertainty using a genetic algorithm (NSGA II) showed a significant influence of phosphorus content uncertainty on the choice of soapstock splitting operating conditions. These findings underscore the importance of accurate phosphorus quantification and support the development of robust, data-efficient computational tools for the monitoring, prediction, and optimisation of complex industrial processes such as soapstock splitting.