Comparing Surrogate Models for Real-Time Dynamic Reactor Simulations
Please login to view abstract download link
Chemical engineering systems often rely on computationally expensive PDEs, posing a challenge for real-time digital twin applications such as optimization and control. Surrogate models help by reducing computational costs, categorized as data-fit, reduced-order, or hierarchical models. We study the dynamic behavior of a catalytic CO2 methanation reactor, which is important for power-to-x systems that convert CO2 and green hydrogen to methane. Using data from a mechanistic model calibrated with pilot plant data, we simulate load changes with three surrogate models: Operator Inference (OpInf), Sparse Identification of Nonlinear Dynamics (SINDy), and Graph Neural Networks (GNN). OpInf shows the best balance of accuracy and interpretability with minimal error, while SINDy is the most interpretable but requires more computation. GNN is the fastest, but less interpretable. All models reduce computation time and are suitable for real-time operation.
