New Graph Network Models Industrial Process Dynamics with Learned Routing.

Industrial process networks, such as those found in manufacturing or chemical plants, frequently change their operational topology due to throttling, bypassing, or unit state changes. Traditional models often struggle to capture these dynamic shifts accurately, sometimes fitting data without assigning physically meaningful routing. This research proposes Conservative Hybrid Graph Networks (CHGNs) to overcome these limitations. CHGNs are designed to learn the routing, operational regime assignments, and removal rates directly from data. Crucially, these learned components are integrated into a fixed transport equation, ensuring that the fundamental principle of mass balance is maintained by construction for any predicted routing. Evaluations show that CHGNs can transfer effectively to larger, unseen graph topologies without retraining, significantly outperforming baseline graph neural networks. While excelling at detecting physical faults, the model's ability to predict manual interventions is limited when valve actions are unobserved, highlighting an area for future data collection and model refinement.