Uncertainty Quantification and Error Propagation in the Enthalpy and Entropy of Surface Reactions Arising from a Single DFT Functional

Gerhard R. Wittreich, Geun Ho Gu, Daniel J. Robinson, Markos A. Katsoulakis, Dionisios G. VlachosThe Journal of Physical Chemistry C 2021 · DOI

Abstract: Accounting for parametric uncertainty in models is essential for quantifying the models’ predictive ability. Recently, approaches have been introduced to estimate parametric uncertainty in kinetic models while accounting for correlations among energy parameters. However, correlations have been estimated indirectly and correlations in entropies have not been accounted for. For surface-catalyzed microkinetic models of >C2 (more than two carbon-containing) molecules, which consist of thousands of reaction steps and intermediate surface species, first-principles density functional theory (DFT) is costly, and thus, estimation of thermochemistry and reaction barriers requires surrogate methods of DFT, such as group additivity and Brønsted–Evans–Polanyi relationships, respectively. For such parametrization, model uncertainty is unclear. This work develops a framework to overcome these gaps using group additivity and a single DFT functional. We estimate correlations in parameters of kinetic models and quantify uncertainty for thermochemistry, reaction barriers, reaction paths, and ultimately reaction rates, accounting also for the contribution of entropic uncertainty. The approach is illustrated on propane combustion and ethane oxidative dehydrogenation reactions.

Introduction

Predicting catalytic performance accurately requires microkinetic models (MKMs) built upon thousands of parameters, including reaction barriers and adsorption energies. Typically, these parameters are derived from Density Functional Theory (DFT). However, for molecules larger than two carbons (>C2), direct DFT calculations for every intermediate and transition state become computationally prohibitive. This necessitates the use of surrogate methods like Group Additivity (GA) and Brønsted–Evans–Polanyi (BEP) relationships.

While these surrogates enable the modeling of complex networks, they introduce "parametric uncertainty." Previous efforts to quantify this uncertainty often required ensembles of multiple DFT functionals—an expensive approach—or ignored critical factors like parameter correlations and entropy. This paper introduces a scalable statistical framework to quantify and propagate uncertainty in MKMs using a single DFT functional, specifically accounting for the often-overlooked effects of entropy.

Methodology

The core of the proposed framework is the integration of Group Additivity with rigorous statistical sampling. GA decomposes complex molecules into smaller, manageable chemical groups (e.g., a carbon atom bonded to three hydrogens).

Correlation and Uncertainty Estimation

The authors utilize a "configuration matrix" within the GA framework to track the frequency of specific groups across different molecules. This matrix allows for the calculation of correlations between parameters using linear algebra, preventing the "double-counting" of variance for groups that frequently co-occur.

Illustration of Group Additivity (GA) and the configuration matrix used to estimate parameter correlations.

To quantify uncertainty, a bootstrap resampling method was applied to a training set of 164 molecules (C1–C5 hydrocarbons). This process estimated the Root Mean Squared Error (RMSE) for both enthalpy and entropy. From these estimates, 5,000 "perturbed" versions of the thermodynamic database were generated by sampling from a multivariate normal distribution.

Propagation through Microkinetic Models

These perturbed datasets are fed into the MKM to observe how variations in thermodynamic inputs affect final Quantities of Interest (QoIs), such as turnover frequency (TOF) and reaction orders. The framework also accounts for lateral interactions and equilibrium constants as part of the iterative solving process.

Flowchart of the microkinetic modeling framework, including uncertainty propagation and lateral interaction adjustments.

Key Results and Insights

The Significance of Entropy

A major finding of this work is that entropic uncertainty contributes approximately 12% to the total error in Gibbs free energy. Neglecting entropy leads to an incomplete understanding of equilibrium and reverse reaction rates, particularly at the high temperatures common in industrial catalysis.

Stiff vs. Sloppy Parameters

The study reveals a distinction between "stiff" and "sloppy" metrics. While predicted reaction rates (TOF) can vary by over an order of magnitude due to parametric uncertainty, reaction orders remain remarkably stable. This low variance suggests that reaction orders are a highly reliable "stiff" signature for validating models against experimental data.

Bimodality in Reaction Pathways

Uncertainty analysis uncovered that minor shifts in energy parameters can lead to entirely different dominant reaction pathways. This phenomenon results in bimodal distributions for apparent activation energies. For example, ethane oxidative dehydrogenation showed distinct peaks at ~28 and ~42 kcal/mol. This suggests that conflicting reports in experimental literature may not be simple errors, but rather the activation of different pathways due to slight variations in conditions.

Conclusion

By leveraging Group Additivity, this framework provides a scalable solution for uncertainty quantification in massive reaction networks that were previously computationally inaccessible. Future work aims to expand this profile by incorporating uncertainties in BEP relationships and surface coverage effects, moving toward a truly comprehensive UQ profile for microkinetic modeling.

Summary updated

Abstract

Accounting for parametric uncertainty in models is essential for quantifying the models’ predictive ability. Recently, approaches have been introduced to estimate parametric uncertainty in kinetic models while accounting for correlations among energy parameters. However, correlations have been estimated indirectly and correlations in entropies have not been accounted for. For surface-catalyzed microkinetic models of >C2 (more than two carbon-containing) molecules, which consist of thousands of reaction steps and intermediate surface species, first-principles density functional theory (DFT) is costly, and thus, estimation of thermochemistry and reaction barriers requires surrogate methods of DFT, such as group additivity and Brønsted–Evans–Polanyi relationships, respectively. For such parametrization, model uncertainty is unclear. This work develops a framework to overcome these gaps using group additivity and a single DFT functional. We estimate correlations in parameters of kinetic models and quantify uncertainty for thermochemistry, reaction barriers, reaction paths, and ultimately reaction rates, accounting also for the contribution of entropic uncertainty. The approach is illustrated on propane combustion and ethane oxidative dehydrogenation reactions.
BibTeX
@article{wittreich2021uncertainty,
  author = {Gerhard R. Wittreich and Geun Ho Gu and Daniel J. Robinson and Markos A. Katsoulakis and Dionisios G. Vlachos},
  title = {{Uncertainty Quantification and Error Propagation in the Enthalpy and Entropy of Surface Reactions Arising from a Single DFT Functional}},
  journal = {The Journal of Physical Chemistry C},
  year = {2021},
  doi = {10.1021/acs.jpcc.1c04754},
}
Figure