(372u) Learning a Data-Driven Reactor Model from Experimental Data of a Continuously Operated Fixed-Bed CO2 methanation Reactor on Pilot-Scale

Data-driven approaches are increasingly used in process systems engineering, e.g. for modelling, simulation and control of chemical reactors. Depending on the complexity of the problem, these methods require different amounts of data of sufficiently high quality. The aim of the present study is the systematic evaluation of different data-based approaches for modelling a pilot-scale CO₂ methanation reactor operated continuously under steady-state conditions.

In a first step, we used in-silico data generated by the numerical solution of a mechanistic model of a cooled CO₂ methanation reactor derived in our previous work [1]. This model produces a very good representation of the reactor behavior, but the derivation of any mechanistic reactor model is mostly based on simplifying assumptions, and thus such a model is never a perfect digital twin of its physical counterpart. In order to produce a highly accurate representation of the behavior of a real reactor at pilot-scale, one needs to train a suitable data-driven model with experimental data harvested at the real system.

For this purpose, we designed, largely automated and operated a tubular pilot scale fixed-bed reactor (2 m length), equipped with a thermo-optical sensor fiber able to record the axial temperature profile along the center line of the catalytic packed bed with high spatial resolution of the axial coordinate [2]. The product gas mixture was measured with an analytical system combining a mass spectrometer (MS), an infrared (IR) spectrometer and a thermal conductivity detector (TCD). The cooling jacket inlet temperature and the feed flow rate were used as input variables of the system. The total feed flow rate was precisely controlled by use of mass flow controllers, while the feeding ratio of the two reactants H₂ and CO₂ was kept constant at stoichiometric condition, i.e. H₂/CO₂ = 4, by use of mass flow controllers.

We developed a sampling strategy to generate a big set of experimental data required for precise steady state data-driven modeling of the methanation reactor. The upper and lower bounds of the variable process parameters were fixed according to a classical 2^k factorial design. By connecting these points, we defined pathways in the parameter space that we followed under quasi-steady state conditions to generate a total of more than 250 million data points (reactor outlet concentrations and associated temperature profiles at each point in the parameter space).

The data set obtained from our sampling strategy was used to train artificial neural networks (ANNs) of different complexity. ANNs were compared with simple linear models and also with the mechanistic model mentioned above. ANNs yielded the most precise representation of the training data, but they must by applied with great care to avoid overfitting [3]. This is why cross-validation is essential; otherwise the applicability of ANNs to assist optimal reactor operation is greatly reduced.

References

[1] Peterson, L., Bremer, J. and Sundmacher, K. (2024). Challenges in data-based reactor modeling: A critical analysis of purely data-driven and hybrid models for a CSTR case study. Computers & Chemical Engineering, 184, 108643.

[2] Bremer, J. (2020). Advanced operating strategies for non-isothermal fixed-bed reactors exemplified for CO₂ methanation. PhD dissertation, Otto-von-Guericke-Universität Magdeburg.

[3] Himmelblau, D.M. (2000). Applications of artificial neural networks in chemical engineering. Korean Journal of Chemical Engineering, 17(4), 373–392.