Mechanism Deduction from Noisy Chemical Reaction Networks

Abstract

We introduce KiNetX, a fully automated meta-algorithm for the kinetic analysis of complex chemical reaction networks derived from semi-accurate but efficient electronic structure calculations. It is designed to (i) accelerate the automated exploration of such networks and (ii) cope with model-inherent errors in electronic structure calculations on elementary reaction steps. We developed and implemented KiNetX to possess three features. First, KiNetX evaluates the kinetic relevance of every species in a (yet incomplete) reaction network to confine the search for new elementary reaction steps only to those species that are considered possibly relevant. Second, KiNetX identifies and eliminates all kinetically irrelevant species and elementary reactions to reduce a complex network graph to a comprehensible mechanism. Third, KiNetX estimates the sensitivity of species concentrations toward changes in individual rate constants (derived from relative free energies), which allows us to systematically select the most efficient electronic structure model for each elementary reaction given a predefined accuracy. The novelty of KiNetX consists in the rigorous propagation of correlated free-energy uncertainty through all steps of our kinetic analyis. To examine the performance of KiNetX, we developed AutoNetGen. It semirandomly generates chemistry-mimicking reaction networks by encoding chemical logic into their underlying graph structure. AutoNetGen allows us to consider a vast number of distinct chemistry-like scenarios and, hence, to discuss the importance of rigorous uncertainty propagation in a statistical context. Our results reveal that KiNetX reliably supports the deduction of product ratios, dominant reaction pathways, and possibly other network properties from semi-accurate electronic structure data.

Transforming Predictions into Testable Hypotheses: The Case of Polar Organic Reactivity

“…Only 17-31% of their predictions (with respect to reference species only) fall within their 95% confidence intervals, clearly indicating that their model underestimates parameter uncertainty. We observed this behavior of Gaussian processes in another context [18] and concluded to select kernel functions not only with respect to predictive power; they should also yield statistically significant results, i.e., pass a hypothesis test.…”

“…To explore the potential of UQ for chemical research, we build upon previous work by Proppe and Reiher [13], addressing Mössbauer spectroscopy [7,14], dispersion corrections to density functional theory [15,16], and reaction kinetics [17,18]. This foundation will support our endeavor to pave the way for a novel approach to determining reactivity parameters with steadily increasing accuracy.…”

Systematic microsolvation approach with a cluster‐continuum scheme and conformational sampling

“…As such extensions will require significant computational resources, a combination with approximate interaction models ranging from semiempirical methods to classical molecular-mechanics force-fields and machine-learning models is a natural extension (for the fast-growing literature on these schemes, we may refer to references in References [98][99][100][101]). However, we emphasize that reliability of such methods that trade accuracy for computational efficiency is best guaranteed if suitable uncertainty quantification schemes (such as those reported by us in References [102][103][104][105]) are in operation that inform about the range of applicability. Note also that such an approach should then not be considered as some general transferable model, but as a systemfocused baseline model whose suitability is quantitatively assessed by uncertainty quantification procedures in a rolling fashion through continuous benchmarking as discussed in References [102,105].…”

Automatic discovery of chemical reactions using imposed activation

“…Investigations are also underway to improve systematicity by, for example, automating the selection of activation coordinates using chemical and computational means. This kind of automation could enable the study of complex chemical reaction networks 16,73 by applying imposed activation computations recursively over obtained products. 23 Finally, the inclusion of automated (de)protonation and tautomerization capabilities 74,75 could increase the range of chemistry that can be studied systematically.…”