Digital twinning for dynamically operated Power-to-X reactors
Please login to view abstract download link
While traditional chemical engineering processes are designed for steady-state operation, the status quo is increasingly shifting towards an environment where transient conditions are driven by external disturbances and fluctuating market demands. This dynamic operation is highly pronounced in renewable energy-integrated processes %energy-integrated processes , where reactor systems must respond continuously to fluctuating conditions. A representative example of such a system is the dynamically operated Power-to-X reactor considered here. We focus on a catalytic carbon dioxide methanation process within a single fixed-bed tube of industrial dimensions ("the physical twin"), acting as a representative process element for large-scale production plants. As core technology for renewable energy storage, this process converts hydrogen and carbon dioxide into methane. The end goal is to construct a digital twin (DT) \cite{P1} for the methanation reactor, i.e., a virtual representation of the physical twin, continuously updated with real-time data from its physical counterpart. This bidirectional communication between the two entities needs to be dynamic, allowing for the simulation, analysis, and monitoring of the physical asset, enabling better decision-making for maintenance and operational improvements. Digital twinning for complex chemical processes requires models that balance predictive accuracy with the ability of rapid execution of simulations, a requirement often unmet by first-principle models. To bridge this gap, we learn structured low-dimensional reduced-order models directly from reactor data using Operator Inference \cite{P2}. We demonstrate the approach using real-world data from a pilot-scale Power-to-X methanation reactor. Based on experimental sensor measurements, we identify the governing system dynamics under realistic operating conditions. To address challenges arising from measurement noise and partially observed spatial states, we extend the standard Operator Inference framework with a neural decoder for nonlinear field reconstruction and a hybrid physics-informed formulation for robust coefficient identification. This allows the exploration of critical operating regimes and demonstrates feasibility for real-time safety verification under idealized conditions.
