(731c) Towards Fully Automated Generation of Predictive Detailed Kinetic Models: 2,3-Dimethylbut-2-Ene Oxidation Modeling As an Example

Generating a predictive detailed mechanism model remains a challenge, despite the availability of automated reaction mechanism generation packages (e.g., Reaction Mechanism Generator (RMG)). One major bottleneck is supplying high-accuracy sensitive thermodynamic and kinetic parameters, usually from first-principal calculations, to yield reasonable qualitative and quantitative predictions from the resulting model. The large number of never-before-seen intermediates and reactions involved in a new chemical system and the complexity of calculating their properties pose a high barrier regarding the requirement of expert knowledge and labor force.

In our previous work, we showed that calculating a large number of thermodynamic properties is achievable by implementing a workflow management software, Automatic Rate Calculator (ARC), to coordinate required calculations; coupling RMG and ARC allows generating an accurate detailed mechanism with little human intervention. Here, we further extend our capability to calculate kinetic parameters in a self-guided manner, using a newly-developed machine learning-based algorithm to predict transition state (TS) geometries. The algorithm, taking 3D geometries of reactants and products as inputs, will align those geometries and feed them into a graph neural network to produce a high-quality TS guess; ARC will then utilize the guess and conduct required calculations to obtain reaction rates. The effectiveness of this improvement is demonstrated by 2,3-dimethylbut-2-ene (DMBENE) oxidation modeling. DMBENE exhibits the octane hyper-boosting phenomenon, where its mixture with another blendstock has a higher Research Octane Number (RON) than the RON of individual components. Although the relative simplicity and symmetry of DMBENE makes it a good example to study the phenomenon, the lack of sufficient thermodynamic and kinetic parameters greatly hinders the modeling effort. The effectiveness of the improved workflow not only fills the gap of DMBENE modeling but also brings the prospect of conducting fully automated model generation within reach.