Embedding Physics into Neural ODEs to learn Kinetics from Integral Reactors

While the digitalization of chemical research and industry is vastly increasing the amountof data for developing kinetic models, model parametrization is not keeping up. To take advantage of the full potential of this data, machine learning tools are required that autonomously learn kinetics from reactor data. Previously, we introduced Global Reaction Neural Networks with embedded stoichiometry and thermodynamics for kinetic modelling. When trained as a neural ordinary differential equation (neural ODE), they discovered kinetics from integral reactor measurements of an equilibrium limited steam reforming reactor whereas conventional neural ODEs failed. We now extend their application to another industrially relevant case of reactors operating at full conversion. Using the preferential oxidation of CO in H2 rich streams as an example, we show that the physics-embedded neural network discovers kinetics from stiff systems containing cases of both full conversion and equilibrium limitation using integral reactor data