Uncertainty analysis of correlated parameters in automated reaction mechanism generation

Abstract

Uncertainty analysis is a useful tool for inspecting and improving detailed kinetic mechanisms because it can identify the greatest sources of model output error. Owing to the very nonlinear relationship between kinetic and thermodynamic parameters and computed concentrations, model predictions can be extremely sensitive to uncertainties in some parameters while uncertainties in other parameters can be irrelevant. Error propagation becomes even more convoluted in automatically generated kinetic models, where input uncertainties are correlated through kinetic rate rules and thermodynamic group values. Local and global uncertainty analyses were implemented and used to analyze error propagation in Reaction Mechanism Generator (RMG), an open‐source software for generating kinetic models. A framework for automatically assigning parameter uncertainties to estimated thermodynamics and kinetics was created, enabling tracking of correlated uncertainties. Local first‐order uncertainty propagation was implemented using sensitivities computed natively within RMG. Global uncertainty analysis was implemented using adaptive Smolyak pseudospectral approximations as implemented in the MIT Uncertainty Quantification Library to efficiently compute and construct polynomial chaos expansions to approximate the dependence of outputs on a subset of uncertain inputs. Cantera was used as a backend for simulating the reactor system in the global analysis. Analyses were performed for a phenyldodecane pyrolysis model. Local and global methods demonstrated similar trends; however, many uncertainties were significantly overestimated by the local analysis. Both local and global analyses show that correlated uncertainties based on kinetic rate rules and thermochemical groups drastically reduce a model's degrees of freedom and have a large impact on the determination of the most influential input parameters. These results highlight the necessity of incorporating uncertainty analysis in the mechanism generation workflow.

Updated thermochemistry for renewable transportation fuels: New groups and group values for acetals and ethers, their radicals, and peroxy species

“…Note that the same trend is observed for standard entropies and heat capacities, where the larger species are on the right‐hand side instead of the left‐hand side: The larger the species, the more negative are their standard enthalpies of formation and the larger are their standard entropies and heat capacities. Note that besides these “intrinsic” uncertainties in the group values, species thermochemistry estimated by the GA methods also comprises an “estimation” uncertainty rooted in the GA approximation itself, which cannot be reduced, even if all group values are known without uncertainties 46 …”

“…Note that besides these "intrinsic" uncertainties in the group values, species thermochemistry estimated by the GA methods also comprises an "estimation" uncertainty rooted in the GA approximation itself, which cannot be reduced, even if all group values are known without uncertainties. 46 For standard entropies, the total RMSD improves from 4.64 cal/(mol⋅K) to 2.36 cal/(mol⋅K) (cf Figure 1B), with a RMSD of 1.28 cal/(mol⋅K) for the fuel and fuel radical species. Again, the hydroperoxide and peroxy species are less accurately described compared to the fuel and fuel radical species.…”

Quantifying the Impact of Parametric Uncertainty on Automatic Mechanism Generation for CO2 Hydrogenation on Ni(111)

“…Global uncertainty analysis (GUA) was recently added to RMG as a post-processing tool exclusively for gas-phase mechanisms. 68 The effect of parametric uncertainty on the automatic mechanism generation process itself, in contrast, has not been previously investigated. A major component of the current work is to quantify how uncertainties in the model-generation parameters (e.g.…”

Moving from postdictive to predictive kinetics in reaction engineering

“…This problem is endemic in large kinetic models, because several parameters are estimated based on the same underlying functional group values or linear free energy relationships. Recent papers have highlighted the importance of accounting for this type of correlation when computing uncertainties in the predictions 93,94 …”

“…Often the errors in the conversion and the product selectivities are very different, since the conversion is sensitive to the absolute concentration of reactive species (e.g., radicals, active sites of a catalyst), but the selectivities are not. Sometimes this systematic error‐cancellation in certain observables can be detected by uncertainty analyses 93,94 . But of course there are other cases where the errors in individual parameters combine in ways that make it quite difficult to accurately predict the observable of greatest interest.…”