AI-Driven Optimization of Plaster Manufacturing Using Design of Experiments and Machine Learning

Abstract

This paper presents a two-stage framework for optimizing process control in plaster manufacturing under uncertain operating conditions, integrating Design of Experiments (DoE) and Machine Learning (ML) techniques. In the first stage, a full factorial 2^k DoE was employed to evaluate four input parameters—roller mill current (M), blower hot air flow (F), classifier speed (S) and temperature (T)—with combined water (CW) as the system output. The analysis identified temperature as the most influential factor, while several multi-factor interactions also significantly impacted system behavior. In the second stage, three ML models—Decision Tree (DT), Support Vector Regression (SVR) and Gaussian Process Regression (GPR)—were implemented to predict CW. All models achieved strong predictive accuracy (R² > 0.8), with mean absolute error (MAE) values ranging from 0.0860 to 0.0906. While SVR achieved the highest training R² (0.862), GPR demonstrated the most consistent performance and lowest overall error. These results support a hybrid control strategy that leverages DoE for initial feature screening and ML for robust data-driven prediction. GPR was identified as the most suitable model among those tested for applications requiring high precision and reliability, due to its superior generalization and integrated uncertainty quantification.