Robust extraction of thermodynamically consistent neural-network-based kinetic models from reactor data: Methanation, methanol synthesis, and dimethyl ether synthesis

Kinetic models are central to the design, analysis, and control of catalytic reaction systems. Current kinetic modeling approaches rely on mechanistic rate expressions derived and parameterized using chemical intuition, ab initio calculations, and laboratory reactor measurements. However, this process is time-intensive and requires substantial system-specific expert knowledge. These challenges motivate the development of alternative frameworks to efficiently extract kinetic models directly from experimental reactor datasets. Here, we demonstrate robust extraction of stoichiometrically and thermodynamically consistent neural-network-based kinetic models from laboratory reactor data for three industrially relevant catalytic systems: methanation, methanol synthesis, and dimethyl ether synthesis. We first identify failure modes that hinder training and then introduce a bounded Global Reaction Neural Network architecture (bGRNN) that limits the contribution of the thermodynamic driving force to enable stable optimization. Using this architecture, we extract accurate kinetic models for all three systems that generalize to previously unseen operating conditions. For dimethyl ether synthesis over a mixed-catalyst system, the framework correctly attributes methanol synthesis to the Cu/ZnO/Al 2 O 3 catalyst and methanol dehydration to the γ -Al 2 O 3 catalyst without imposing a priori reaction-catalyst assignments. Taken together, these results support neural-network-based kinetic model extraction as a practical route to efficiently derive physically consistent kinetics from reactor data across catalytic reaction systems.