Creating a Reaction Data Set Labeled with Reaction Class for Automated Reaction Classification for ReaxFF Molecular Dynamics Simulations of Realistic Fuel Pyrolysis

Pyrolysis chemistry is important
in both engine combustion
and
industrial utilization of various fuels. Understanding pyrolysis chemistry
is challenging due to the large number of reactions involved and the
explosion of intermediate species structures in the radical-driven
process. Since the bond changes reflect the very core information
on a reaction, automatic reaction classification based on reaction
centers can be useful to peak at a simplified reaction view of a complex
pyrolysis process. This work proposes and implements a scheme to build
a reaction data set labeled with the reaction class for reactions
from reactive molecular dynamics simulations using ReaxFF (ReaxFF
MD) in generating global reactions in pyrolysis of realistic fuel
mixtures. The major steps include the automated conversion of reactions
into elementary-like reactions with a pseudosingle reaction center,
automatic extraction of extended reaction centers, reaction class
defining, and manual labeling. There are 46 reaction classes defined
in total based on both pyrolysis reaction knowledge and reaction observations
from ReaxFF MD simulations of realistic hydrocarbon fuel pyrolysis.
With the effort to have as adequate number of reactions as possible
labeled for each reaction class defined, 7862 reactions were manually
labeled with reaction classes for the data set of 26,881 elementary-like
reactions that cover major pyrolysis reaction classes of typical hydrocarbon
fuel components of n-paraffins, iso-paraffins, olefins, cycloparaffins, and aromatics. The reaction data
set has been used in the scheme of SRG-Reax to build a semisupervised
machine learning model of tri-training to predict the reaction classes
of pyrolysis reactions. Through automated reaction classification,
30 major reaction classes involved in a total of 3479 pyrolysis reactions
of real RP-3 fuel containing 45 components unravel the overall pyrolysis
reaction characteristics of the fuel system. With additional reaction
classes defined and reaction data labeled, the approach can be used
for various fuels.