(693b) Analyzing the Uncertainty of Linear Scaling, Machine Learning, and DFT Derived Thermodynamics for the Catalytic Partial Oxidation of Methane on Rhodium 111

Microkinetic models (MKMs) have become an essential tool for understanding reaction mechanisms and designing new catalysts in heterogeneous catalysis. Accurate thermodynamic data for surface species is crucial for developing reliable MKMs. In this study, we explored different methods for generating thermodynamic data for species involved in catalytic methane partial oxidation (CMPO) on rhodium, an important reaction used in various industrial processes.

We generated a MKM for CMPO on rhodium with the Reaction Mechanism Generator (RMG), using linear scaling relationships (LSRs) to estimate the surface species’ thermodynamic parameters based on the species enthalpy for platinum. We then compared the thermochemical data generated by LSRs with parameters generated by two alternative methods: the Open Catalysis Project (OCP) pretrained model and density functional theory (DFT). The OCP is a machine learning model trained to predict the adsorption energies of molecules on transition metal surfaces. The thermodynamic data generated by the OCP and DFT were compared with LSR values, and the three sets of thermodynamic data were used to simulate the CMPO process on rhodium. The simulations were performed in Cantera, and the results were compared to the experimental data.

To quantify each method's uncertainty, we used the Parameter Estimation and Uncertainty Quantification for Science and Engineering (PEUQSE) package. PEUQSE was used to estimate the Maximum A Posteriori (MAP) values for binding energy and the High Probability Density (HPD) region for the LSR, OCP, and DFT models. This revealed a feasible set and uncertainties for binding energies calculated with each of these methods.

In conclusion, this study highlights the importance of accurate thermodynamic data and discusses the applicability of the pretrained neural network models in microkinetic modeling. These methods can be applied to other catalytic systems to generate more accurate microkinetic models and ultimately enable the design of more efficient and sustainable catalytic processes.