Dynamic grey box modeling of decanter centrifuges via continual learning

Decanter centrifuges are essential for solid–liquid separation across multiple industrial sectors, yet their operation still depends heavily on expert knowledge and time-intensive experimentation. Existing predictive models, including white box, black box and hybrid grey box models, typically require static offline calibration on empirical data. This limits their ability to adapt to changing material properties during operation, which leads to diminished prediction accuracy under dynamic operating conditions. This study addresses this limitation by integrating a Continual Learning (CL) framework into a grey box model for decanter centrifuge operation, enabling the model to adapt online to changing material properties and process behavior. While CL has shown promise in machine learning, its application in process and chemical engineering remains limited. To mitigate catastrophic forgetting, defined as the gradual loss of previously acquired knowledge during sequential learning, the Online Elastic Weight Consolidation (EWC) method is employed. The proposed approach is evaluated through an application-driven case study of a dynamic mill-decanter process circuit. Results demonstrate that the proposed model adapts reliably online while maintaining memory stability and generalization, outperforming a statically calibrated baseline under changing operating conditions. This adaptability enhances operational stability and enables more responsive process control, highlighting strong potential for real-time deployment in complex and dynamic industrial processes.