Uncertainty Quantification of Linear Scaling, Machine Learning, and DFT Derived Thermodynamics for the Catalytic Partial Oxidation of Methane on Rhodium

Accurate and complete microkinetic models (MKMs) are powerful for anticipating the behavior of complex chemical systems at different operating conditions. In heterogeneous catalysis, they can be further used for the rapid development and screening of new catalysts. Density functional theory (DFT) is often used to calculate the parameters used in MKMs with relatively high fidelity. However, given the high cost of DFT calculations for adsorbates in heterogeneous catalysis, linear scaling relations (LSRs) and machine learning (ML) models were developed to give rapid estimates of the parameters in MKM. Regardless of the method, few studies have attempted to quantify the uncertainty in catalytic MKMs, as the uncertainties are often orders of magnitude larger than those for gas phase models. This study explores uncertainty quantification and Bayesian Parameter Estimation (BPE) for thermodynamic parameters calculated by DFT, LSRs, and GEMNET, a ML model developed under the Open Catalyst Project. A model for catalytic partial oxidation of methane on Rhodium was chosen as a case study, in which the model’s thermodynamic parameters and their associated uncertainties were determined using DFT, LSR, and GEMNET. Markov Chain Monte Carlo coupled with Ensemble Slice Sampling was used to sample the highest probability density (HPD) region of the posterior and determine the maximum of the a posteriori (MAP) for each thermodynamic parameter included. The optimized microkinetic models for each of the three estimation methods had quite similar mechanisms and agreed well with the experimental data for gas phase mole fractions. Exploration of the HPD region of the posterior further revealed that adsorbed hydroxide and oxygen likely bind on facets other than Rhodium 111. The demonstrated workflow addresses the issue of inaccuracies arising from the integration of data from multiple sources by considering both experimental and computational uncertainties, and further reveals information about the active site that would not have been discovered without considering the posterior.