Challenges in data-based reactor modeling: A critical analysis of purely data-driven and hybrid models for a CSTR case study

Abstract

In this study, we critically examine the performance of hybrid and purely data-driven models in reactor systems using catalytic CO2 methanation in a continuously stirred tank reactor as a representative case. Our comparative analysis includes four models: one purely data-driven model and three hybrid models. These hybrid models blend data-driven and mechanistic approaches, using data-driven submodels for specific process parts and data correction for mechanistic inaccuracies. The models are evaluated on simulated data to assess their accuracy, training effort, and reliability. Our results show that hybrid models do not consistently outperform the purely data-driven model. This highlights the need for careful model selection, taking into account the specifics of the problem. The choice between hybrid and pure data-driven models requires a balanced evaluation of effort and potential benefits, emphasizing the importance of systematic analysis in model selection.