The RMG Database for Chemical Property
Prediction
Matthew S. Johnson,† Xiaorui Dong,† Alon Grinberg Dana,†,§ Yunsie Chung,†
David Farina, Jr.,‡ Ryan J. Gillis,† Mengjie Liu,† Nathan W. Yee,† Katrin
Blondal,¶ Emily Mazeau,‡ Colin A. Grambow,† A. Mark Payne,† Kevin
Spiekermann,† Hao-Wei Pang,† C. Franklin Goldsmith,¶ Richard H. West,‡ and
William H. Green∗,†
†Department of Chemical Engineering, Massachusetts Institute of Technology, Cambridge,
MA 02139, United States
‡Department of Chemical Engineering, Northeastern University, Boston, MA 02115,
United States
¶School of Engineering, Brown University, Providence, RI 02912, United States
§The Wolfson Department of Chemical Engineering, Grand Technion Energy Program
(GTEP), Technion – Israel Institute of Technology, Haifa 3200003, Israel
E-mail: whgreen@mit.edu
Abstract
The RMG database for chemical property prediction is presented. The RMG
database consists of curated datasets and estimators for accurately predicting param-
eters necessary for constructing a wide variety of chemical kinetic mechanisms. These
datasets and estimators are mostly published and enable prediction of thermodynamics,
kinetics, solvation effects, and transport properties. For thermochemistry prediction,
1
the RMG database contains 45 libraries of thermochemical parameters with a com-
bined 4564 entries, a group additivity scheme with nine types of corrections including
radical, polycyclic and surface absorption corrections with 1580 total curated groups
and parameters for a graph convolutional neural net trained using transfer learning
from a set of >130,000 DFT calculations to 10,000 high-quality values. Correction
schemes for solvent-solute effects, important for thermochemistry in the liquid phase,
are available. They include tabled values for 195 pure solvents and 152 common solutes
and a group additivity scheme for predicting the properties of arbitrary solutes. For
kinetics estimation the database contains 92 libraries of kinetic parameters contain-
ing a combined 21,000 reactions and contains rate rule schemes for 87 reaction classes
trained on 8655 curated training reactions. Additional libraries and estimators are
available for transport properties. All of this information is easily accessible through
the graphical user interface at https://rmg.mit.edu. Bulk or on-the-fly use can be fa-
cilitated by interfacing directly with the RMG Python package which can be installed
from Anaconda. The RMG database provides kineticists with easy access to estimates
of the many parameters they need to model and analyze kinetic systems. This helps
speed up and facilitate kinetic analysis by enabling easy hypothesis testing on path-
ways, by providing parameters for model construction and by providing information to
check other kinetic parameters against.
Introduction
Understanding and optimization of many important chemical processes such as combustion,
polymerization, electrochemistry, pyrolysis, and oxidation can benefit from detailed and
predictive chemical kinetic mechanisms. In many of these cases, to accurately represent the
chemistry involved mechanisms need to include hundreds to thousands of species and tens
to hundreds of thousands of reactions.
In these cases, we need to assign a rate coefficient to each reaction in the model and often
2
even to many reactions outside the model to determine whether they should remain outside
the model. Additionally, a reverse rate needs to be available for most reactions either by
explicitly specifying it or by defining the thermochemistry of the species involved. For larger
models the latter is typically preferred because it guarantees thermodynamic consistency,
because thermochemistry is easier to estimate than kinetics, and because there are typically
far fewer species in a model than reactions.
The stakes for these estimations are highly variable. Some reactions that are non-limiting
and don’t impact any important branching can tolerate several orders of magnitude of error,
while others may drastically change the chemistry when their rate coefficient is adjusted
by as little as 25%. Sensitivity to thermochemical parameters may also vary significantly,
although large underestimates of any species’ Gibbs free energy will almost always have a
drastic effect.
Typically databases, property estimators, and quantum chemistry methods are used to
estimate these parameters. While many advances have been made recently in on-the-fly
quantum chemistry calculations, in most cases these systems are too computationally ex-
pensive or not robust enough to use for every parameter in a mechanism. For this reason we
will focus our review on databases and property estimators.
Some of the most popular databases are those archived by NIST containing experimental
and quantum chemical rate coefficients and thermodynamic properties.1 The NIST databases
are very extensive although the data quality can be highly variable. While this makes NIST
a great place to search for parameters it also makes it not ideal for applications where
there is no human in the loop. The active thermochemical tables (ATcT) database provides
high accuracy values for a limited number of species.2 Many databases such as PriMe,3
ChemKED,4,5 CloudFlame6,7 and ReSpecTh exist for storing experimental measurements
that may be relevant. The MolSSI QC Archive project is a quantum chemistry specific
database intended to help supply homogeneous data sets.8
There are many ways of estimating thermochemical and kinetic parameters. By far the
3
most common way of estimating thermochemical parameters is the Benson type group ad-
ditivity method.9,10 Implementations of this method are used in THERM and Genesys.11,12
Graph convolutional neural net methods are beginning to emerge as an alternative.13–16 Ki-
netics estimators typically rely on either use of rate rules assigning specific rates to reactions
matching specific templates or on reaction group additivity.17–19 Recent machine learning
approaches using hierarchical decision trees have been found to be effective for predicting
rate coefficients and reactivity.20,21
The RMG database was primarily created to supply the Reaction Mechanism Generator
software (RMG) with good estimates for key species and kinetic properties.22 RMG is a
rate-based mechanism generation software. This means that RMG selects species based
on their computed production rates during simulations. During mechanism generation at
each iteration RMG simulates the set of species and reactions it has already chosen to
include in the model, referred to as the ”core”. During this simulation it calculates fluxes
through a set of reactions that are under consideration to enter the core, referred to as the
”edge”. If the flux toward a species in the edge is high enough it is added to the core.
This means that unlike kinetic models constructed by hand or using rule-based mechanism
generation, RMG needs to estimate properties for all the species and reactions in the edge,
which are typically 100x and 10x larger, respectively, than the core. Furthermore, these
estimations are much higher stakes. This is because unlike in other mechanism construction
methods, the decisions of whether or not species and reactions are included in the resulting
mechanism are directly based on their estimated properties. Poor thermochemistry and
kinetic estimation in RMG can therefore cause model generation to miss the real pathways
in a kinetic system, highlighting the important role the RMG database has in providing
reasonable data estimation schemes.
Accounting for pressure dependence is very important for many gas-phase chemical sys-
tems particularly when the system involves small molecules, low pressures or high temper-
atures.23 As far as the authors are aware, RMG is the only publicly available automatic
4
mechanism generation tool that can automatically handle pressure dependence including
chemical-activation. It does this by drawing on automatically formulating and solving the
master equation using Arkane24–26 for each pressure-dependent reaction network on the fly.
This requires estimation of E0 and frequencies for each species in the network and k(E) for
each reaction.
To facilitate automated model generation in RMG, the RMG database provides a robust
set of estimators for thermochemistry, rates, and many other relevant properties. It also
provides a database of reactions suitable for small scale machine learning applications and
includes many mechanisms and submechanisms from literature. These estimators, data,
and submechanisms have been tempered and refined during the creation of many RMG
mechanisms.27–32
Theory
Thermo Group Additivity
Figure 1: Example segment of a group additivity correction tree.
RMG’s group additivity scheme is broadly based on that developed by Benson et al.9
Hf298, Sf298 and Cp(T ) at a set of temperatures for a given molecule are defined by summing
5
a contribution from each heavy atom in the molecule. In the 20th century, most of these
group values were derived from experimental data, but in the last 20 years a much larger set
of group values has been derived from quantum chemical calculations.33–36 RMG’s scheme
also integrates several advanced corrections to the ordinary Benson groups:
• radical corrections using the hydrogen bond increment (HBI) method adopted from
Lay et al. 1995.37
• ring and polycyclic strain corrections using the method of Han et al. 2018.38,39
• long distance interaction corrections for non-cyclic gauche 1,4 and 1,5 interactions.
• long distance interaction corrections for halogenated molecules.36
• ortho, para, and meta cyclic non-nearest neighbor interactions using values from Ince
et al. 2015 and Ince et al. 2017.40,41
• ketene corrections.33
• surface adsorption corrections for absorbed species on nickel and platinum.42,43
Each correction is chosen using tree structures like those shown in Figure 1. Starting from
the top of the tree, the algorithm descends to matching child nodes until it finds the most
specific correction that matches the actual structure.
Readily available and reliable values for each group are critical to efficient and accurate
estimation of thermodynamic parameters using group additivity schemes. Unfortunately,
given the diversity of chemical space,44 it is challenging to construct a scheme that com-
prehensively covers all chemistries that may be of interest. With this in mind, it is vital
to have efficient workflows for continually extending this group additivity scheme. For this
purpose we provide scripts and notebooks for generating new groups from a set of species of
interest, proposing species to calculate thermochemistry for, and fitting the new groups to
the calculated species.
6
Thermodynamic Property Neural Network Estimator
The main thermodynamic property neural network estimator is based on the work from Li
et al. 2019 and Grambow et al. 2019.13,45 It uses a graph convolutional neural network to
predict thermochemistry parameters. The model was trained using transfer learning from
130,000 calculations at the B3LYP/6-31G(2df,p) level to around 10,000 high-quality data
points at CCSD(T)-F12a/cc-pVDZ-F12//B3LYP/6-31G(2df,p) or better (including accurate
experimental values). The low level calculations were drawn heavily from HCNO species in
the QM9 data set. For Hf298 predictions the high level calculations were selected species from
that same subset, plus some molecules with high accuracy experimental data. The high-level
training data for the entropy and heat capacity estimators were a mix of experimental data
and 900 ωB97X-D3/def2TZVP calculations from the same subset of molecules. The details
are available in Grambow et al. 2019.45
An additional neural network based thermochemistry estimator specific for halogenated
species was trained off of G4 data from Farina et al. 2021 recently.16,36,46 While not yet
added to the RMG database this neural network will be available in the near future.
Both the group additivity and neural net approaches give usefully-accurate estimates of
the thermochemistry of hundreds of thousands of molecules. But neither is perfectly reliable,
and there are no error bars on the estimates. An important challenge for the future is reliably
identifying the uncertainties in the estimates, even for strange molecules.
Rate Coefficient Estimation Rules
RMG’s rate rule based estimator uses tree structures similar to those used by group addi-
tivity. An example for the R Addition MultipleBond reaction family is shown in Figure 2.
The template for this reaction family is available in Figure 3. Much like traditional rate
rule schemes, when a reaction is proposed it descends down the tree to find the best corre-
sponding rate rule. The rule typically gives Arrhenius parameters as log(A) and Ea which
can be used to compute the desired k(T). Note that this is typically a two-dimensional (as
7
Figure 2: Portion of the RMG rate rule trees for the R Addition MultipleBond family, with
separate trees for the double bond and the attacking radical.
Figure 3: RMG reaction template for R Addition MultipleBond family.
8
shown in Figure 2) or three-dimensional tree space. If there is no estimation rule for the
rate coefficient where the reaction lands, the algorithm searches the tree upwards to find and
average the rules closest in Euclidean distance within the tree space.
Unlike traditional rate rule estimators, however, nearly all rules are defined by descending
a set of training reactions whose rate parameters are known down the tree and placing those
parameters as a rule in the tree. More general rules are then generated by averaging all of
the rules one Euclidean distance step below each point in the tree space and so on up the
tree. These trees are quite sparsely populated as a result of the combinatorically large tree
spaces.
These trees described in this section are designed to estimate ”high-pressure limit” gas-
phase rate coefficients k(T), the quantities typically computed using Transition State Theory
from quantum chemical calculations. Due to fall-off and chemical-activation the actual
pressure-dependent rate coefficients can be very different. Also, in liquid phase the rate
coefficients are different due to solvent effects. The way RMG implements the corrections
needed to obtain the actual rate coefficients in the environment of interest are discussed in
later sections.
Automatic Tree Generation
Automated tree-generation methods in the RMG database can automatically generate the
nodes and rules from the training data using the Subgraph Isomorphic Decision Tree (SIDT)
training algorithm developed by Johnson and Green 2021.20 Many of the families already
use these automatically generated trees. Work is in progress on switching all families to use
automatically generated trees.
Vibrational Frequency Estimation for Estimating ρ(E)
RMG estimates the vibrational frequencies of each molecule as one step in predicting the
density of states, ρ(E), for pressure-dependent rate calculations.25,47 For this purpose the
9
RMG database contains parameters for a group contribution scheme for estimating vibra-
tional frequencies of functional groups in a molecule. The details of this scheme are available
in Grinberg Dana et al. 2022.47 Note these frequencies are only intended for and validated
for estimation of ρ(E) for pressure-dependent rate calculations within this scheme and are
likely not suitable for other purposes. A different scheme, outside of RMG, is available for
those who are interested in predicting IR spectra.48
Liquid Phase Diffusivity Estimation
In liquid phase, it is common for some rate coefficients to be diffusion-controlled. To estimate
these rate coefficients, one needs estimates of molecular diffusivities. These diffusivities
are also needed in liquid-phase reaction-diffusion and reacting flow simulations. The RMG
database can estimate the liquid-phase diffusion coefficient D of a molecule using the Stokes-
Einstein equation based on the method of Zhao et al. 2003
kBT
D = (1)
6πηr
with
( ) 1
0.75V 3
r = (2)
πNA
N∑atoms
V = (V )− 6.56× 10−6i Nbonds (3)
i
where V 49i is the McGowan volume of a particular atom in m
3/mol, η is the viscosity of the
solvent and Nbonds is the number of bonds in the molecule.
50 The database also contains sets
of parameters for calculating the pure component viscosity of 150 solvents using the relation
B
ln(η) = A+ + C log(T ) +DTE (4)
T
10
where η is the viscosity, T is the temperature and the rest are solvent specific viscosity pa-
rameters. The majority of the viscosity parameters are obtained from the work by Viswanath
et al.51 and some are obtained from DIPPR.52
Estimating the Lennard-Jones Parameters
Gas-phase transport properties are important in many kinetics simulations and they can be
easily calculated by most simulation software from the Lennard-Jones parameters for the
associated species. These are calculated by first estimating the critical temperature and
pressure using the Joback group additivity method
∑
ΣT = GAVT ,i (5)c c
i
∑
ΣP = GAVP ,i (6)c c
i
Tb = 198.2 + ΣT (7)c
Tb
Tc = (8)
0.584 + 0.965ΣT − Σ
2
c Tc
1
Pc = (9)
(0.113 + 0.0032N 2atoms − ΣP )c
where the summations are over the groups associated with the heavy atoms in the molecule
Tb is the boiling point in K, Tc is the critical temperature in K, Pc is the critical pressure
in bar and Natoms is the number of atoms in the molecule.
53,54 For halogenated molecules, a
boiling point correction derived by Devotta and Pendyala55 is applied. We are then able to
estimate the Lennard-Jones parameters using empirical correlations from Tee et al.
(T )
1
c 3
σ = 2.44 (10)
Pc
 = 0.77kBTc (11)
11
where σ and  are the Lennard-Jones parameters
[(σ)12 (σ)6]
V (r) = 4 − (12)
r r
in Angstroms and Joules respectively, where kB is the Boltzmann constant.
56 As a fallback
the Lennard Jones parameters can be estimated simply based on the number of atoms in
the molecule.
Solvent and Solute Property Estimation
Within the RMG database liquids are largely assumed to be dilute. The Gibbs free energy
of a solute in solution (∆G∗liq) is defined using the standard state of a solute dissolved in an
ideal solution at a concentration of 1 mol/L and can be calculated from the relationship
∆G∗ = ∆Go + ∆G∗ o→∗liq f,gas solv + ∆G (13)
with
∗
∆Go→∗
V
= −RT ln = 1.9 kcal/mol at 298K (14)
V o
where ∆Gof,gas is derived from gas-phase values with the standard state of an ideal gas
at 1 bar, ∆G∗solv is the solvation free energy with the standard state of an ideal gas at a
concentration of 1 mol/L dissolving into an ideal solution at a concentration of 1 mol/L,
V ∗ = 1 `/mol and V o = RT . ∆Go→∗ corrects for the difference between the gas phase and
1 bar
liquid-phase standard states. The partition coefficient for a given solute and solvent can be
defined as
cliq
K = (15)
cgas
where c is the concentration of solute in the appropriate phase at equilibrium. This K is
directly related to ∆G∗solv by
∆G∗solv = −RT ln(K) (16)
12
thus ∆G∗solv,298K can be calculated using the Abraham linear solvent energy relationship
(LSER)
log10(K298K) = c+ eE + sS + aA+ bB + lL (17)
where the lower case parameters are Abraham solvent parameters and the upper case pa-
rameters are solute parameters.57 The solvent and solute parameters are independent of each
other. The similar Mintz LSER can be used to calculate the enthalpy of solvation at 298 K
∆H∗solv,298K (kJ/mol) = c
′ + e′E + s′S + a′A+ b′B + l′L (18)
where the lower case are Mintz solvent parameters and the upper case are the same solute
parameters as those used in Equation 17.58 The RMG database can then calculate ∆S∗solv,298K
using
∆H∗ ∗
∗ solv,298K
−∆Gsolv,298K
∆Ssolv,298K = (19)298 K
and then linearly extrapolate the temperature to estimate ∆G∗solv
∆G∗ = ∆H∗ − T∆S∗solv solv,298K solv,298K (20)
providing an algorithm to calculate ∆G∗solv for solute-solvent pairs with known parameters.
More accurate temperature-dependent ∆G∗solv prediction using the method of Chung et al.
59
is available within the RMG database for a limited number of solvents. This method can
estimate ∆G∗solv at elevated temperatures along the solvent’s saturation pressure based on
G∗solv,298K, H
∗
solv,298K and the solvent’s temperature-dependent density, and gives a mean
absolute error of approximately 0.4 kcal/mol.59 The density can be computed for 23 solvents
in the RMG database using an open source package CoolProp.60
The Abraham and Mintz solvent parameters are available for 195 and 66 common solvents
respectively in the RMG database. The solute parameters are calculable for neutral species
using the group additivity method from Chung et al. 2022.61 The group additivity scheme
13
for the solute parameters largely follows that of gas-phase thermo but uses a slightly different
approach for halogenated species. If a compound contains any halogens, RMG first computes
the solute parameters for a molecule with all the halogens replaced by hydrogen atoms and
then adds corrections for each halogen atom. This approach is used because it allows RMG
to pull more accurate values from libraries for the de-halogenated species improving accuracy
on the halogenated compounds after correction.
Forbidden Structures
The RMG database tabulates types of species and also specific subtemplates for reaction
families that are forbidden. These groups of species are generally either species that RMG’s
templates may try to create but don’t really exist, or particular chemistries that RMG can’t
represent well yet such as ozonides. The subtemplates are most typically reactions that
shouldn’t be able to happen or unimportant reactions that RMG severely overestimates. As
a result of these forbidden structures the RMG database may occasionally not predict the
existence of certain reactions that a user might expect it to based on the reaction templates.
Atom Energy Corrections and Bond Additivity Corrections
The RMG database includes Atom Energy Corrections (AECs) and Bond Additivity Cor-
rections (BACs). AECs and BACs are empirical corrections to systematic errors in quantum
chemistry calculations that can be used to improve the accuracy of species thermochem-
istry calculations. AECs are energy corrections to the atomization energies associated with
each atom in a molecule. Most AECs in the RMG database are obtained from fitting to
a high-accuracy experimental dataset in the database containing 16 small molecules. The
experimental atomization energies come from CCCBDB,62 and all have uncertainty values
less than 0.2 kcal mol−1. However, at most levels of theory, AECs alone are insufficient to
estimate enthalpies of formation accurately enough for most kinetics applications. BACs
add an additional correction associated with each bond. They are fit using a dataset of
14
about 400 species with well-known heats of formation, primarily drawn from ATcT63 and
CCCBDB.62 Figure 4 shows a histogram of the residuals before and after applying BACs;
errors are relative to the reliable set of experimental enthalpies used for fitting. After ap-
plying BACs the distribution of errors both centers much more closely to zero and notice-
ably tightens, representing a significant reduction in both the bias and variance of errors.
The RMG database stores two types of BACs. Petersson-type BACs64 apply a correction
a) B97X-D3/def2-TZVP b) CCSD(T)-F12a/cc-pVDZ-F12// B97X-D3/def2-TZVP
100
80
80
60
60
40
40
20 20
0 0
10 5 0 5 10 15 20 25 6 4 2 0 2 4 6
Error (kcal/mol) Error (kcal/mol)
Figure 4: Histogram of fitting errors (corrected minus experimental enthalpies) for Petersson-
type BACs for a) ωB97X-D3/def2-TZVP and b) CCSD(T)-F12a/cc-pVDZ-F12//ωB97X-
D3/def2-TZVP. Using a higher level of theory, such as coupled cluster calculations, give
lower errors both before and after applying the fitted BACs.
to ∆fH(298) for each bond type defined by the two atoms and the bond order. Melius-
type BACs65 apply a correction based on bond length integrating information about the
3D geometry. Our BAC procedure and resulting fitted parameters generalize well to new
molecules. For example, Grambow et al.45 applied fitted Melius-type BACs at CCSD(T)-
F12a/cc-pVDZ-F12//B3LYP/6-31G(2df,p) to a test set of about 400 molecules from NIST.
Although the experimental uncertainties were unknown, the resulting RMSE of 1.31 kcal
mol−1 relative to the experimental values is impressive. Similarly, Spiekermann et al.66 ap-
plied fitted Petersson-type BACs at CCSD(T)-F12a/cc-pVDZ-F12//ωB97X-D3/def2-TZVP
to molecules from the Pedley compilation set67 that have experimental uncertainty of less
than 1 kcal mol−1. The resulting RMSE of 1.24 kcal mol−1 demonstrates the accuracy of
our BACs.
15
Count
Count
Content
Thermodynamic parameters
Broadly, RMG’s thermochemical parameter estimation has been designed to accurately han-
dle chemistry containing the elements H,C,O,N, and S. Performance on halogens and metal
adsorbates is rapidly improving as well. However H,C,O molecules are by far the best cov-
ered.
Libraries
RMG has many libraries of accurate experimental and quantum chemistry calcula-
tions for important species. The libraries DFT QCI thermo, CBS QB3 1dHR, and
thermo DFT CCSDTF12 BAC are curated calculated libraries of thermochemistry at the
particular level of theory. For H2 combustion chemistry BurkeH2O2 is recommended while
for C2 chemistry Klippenstein Glarborg2016 is recommended for general purpose use. These
libraries cover common C3 and smaller species quite well.68,69 Many nitrogen species are avail-
able within primaryNS, NitrogenCurran and NOx2018.70,71 Many sulfur species are available
in primaryNS, SulfurGlarborgH2S and SulfurLibrary.72 Species involving both Nitrogen and
Sulfur are available in primaryNS and many important aromatic species are available in
SABIC aromatics. For halogen chemistry, the CHOX G4 (X=F,Cl,Br,FCl,ClBr,FClBr) li-
braries are recommended as they cover a significant portion of their respective chemical
spaces with up to 4 carbons and oxygen atoms.36 Catalyst adsorbates are available in Sur-
faceThermoPt111 and SurfaceThermoNi111.42,43 Many other specialized thermo libraries are
also available within the RMG database.
Group Additivity
The RMG database’s group additivity method for estimating gas-phase thermochemistry
performs very well on C/H/O species. With some exceptions for heavily oxygenated species,
16
the overall accuracy is around on par with that of B3LYP DFT calculations. Nitrogen and
sulfur species are represented within the groups and estimates are reasonable, but accuracy is
variable. For halogens, the groups perform well for sparsely halogenated non-cyclic species,
but often underpredict enthalpies for cyclics and more heavily halogenated molecules. For
C/H/O cyclics the mean absolute error of Hf298 for small cyclics is about 3 kcal/mol while
it is about 5 kcal/mol for large linear cyclics and 10 kcal/mol for fused cyclics.38
Neural Network
The neural network thermochemistry estimator performs well on polycyclics and CHON
species. In practice it is usually superior to the group additivity method on polycyclics
and nitrogen species. Sulfur species were not included in the training set. Aliphatic and
monocyclic oxygenated and hydrocarbon species tend to be better estimated by RMG’s
group additivity method.
Kinetics
Libraries
The RMG database includes a wide variety of literature mechanisms for different
purposes. For general purpose unfitted chemistries we recommend BurkeH2O2inN2
or BurkeH2O2inArHe and Klippenstein Glarborg2016 for C/H/O molecules, pri-
maryNitrogenLibrary and Nitrogen Dean and Bozzelli for nitrogen species and
First to Second Aromatic Ring for aromatic species. The database also includes the
CurranPentane, JetSurF1.0, JetSurF2.0, and FFCM1(-) libraries. The content of each of
RMG’s kinetic libraries is specified and kept up to date in the RMG documentation online.
Families
RMG contains an extendable set of 87 reaction families. Reaction types common to com-
bustion and pyrolysis are very well covered. Common types important for low temperature
17
chemistry and gas-surface chemistry on metal catalysts are covered.
Training Reactions
About half of the RMG database families have four or fewer training reactions. The num-
ber of training reactions for each of the larger families, with more than 20 training reac-
tions, is available in Table 1. Except for H Abstraction and R Addition MultipleBond the
training reactions for families without loose transition states tend to be mostly CBS-QB3
or similar level of theory calculations. However, there are experimental values and esti-
mates mixed in for most of the families (Table 1). The largest families H Abstraction and
R Addition MultipleBond have about 834 and 430 reactions respectively at CBS-QB3 or
similar levels of theory and are supplemented by 2274 and 2521 reactions respectively es-
timated from reaction group additivity schemes inherited from RMG’s original set of rate
rules.
Table 1: Number of training reactions in families with more than 20 training reactions.
Reaction Family Number of Reactions
H Abstraction 3107
R Addition MultipleBond 2894
Intra R Add Endocyclic 843
intra H migration 431
Intra R Add Exocyclic 371
Cl Abstraction 238
SubstitutionS 148
R Recombination 145
Disproportionation 137
Substitution O 128
Br Abstraction 100
F Abstraction 90
Retroene 65
Disproportionation-Y 42
Cyclic Ether Formation 37
XY Addition MultipleBond 33
1,3 Insertion ROR 22
18
Transport
The RMG database contains four literature libraries of transport parameters: GRI-Mech,
NOx2018, OneDMinN2, and NIST Fluorine and a curated library of species not well ac-
counted for by the groups called primaryTransportLibrary. The Joback groups cover
C/H/O/N/S and halogens fairly well.53,55 In cases where the Joback groups are missing,
RMG makes very rough estimates based on number of atoms in the molecule.
Solvent and Solute Properties
RMG’s solvent library contains 195 pure solvents with known Abraham solvent parameters,
66 of which have Mintz solvent parameters and 150 of which have viscosity parameters.
It also has Abraham solvent parameters for four co-solvents. RMG’s solute library holds
parameters for 310 common solutes and the groups for estimating non-radical neutral solute
parameters cover elements H,C,O,N,S,P,F,Cl,Br, and I. The solute parameters predicted from
the group additivity scheme and the solvent parameters in the library together give a mean
absolute error of approximately 0.6 kcal/mol for both solvation free energy and solvation
enthalpy estimates at 298 K when evaluated on a 10 % test set containing randomly chosen
out-of-sample solute compounds.61 The groups needed to estimate solvation of radicals are
a bit more limited and do not include any groups containing S, P, or halogen atoms.
Atom Energy Corrections and Bond Additivity Corrections
RMG’s database of AECs and BACs covers the elements H,C,O,N and S quite consis-
tently. Newer and more recently updated corrections also cover halogens. Some AECs
in the database that are drawn from literature cover phosphorous.
19
Discussion
Combining Data From Many Different Sources
One major challenge when constructing chemical kinetic mechanisms is combining data from
many different sources. Carelessly combining parameters from independently accurate mech-
anisms in many cases can result in an inaccurate mechanism. This occurs primarily for two
different reasons. The first is that many literature mechanism parameters are fit in bulk
to experimental data. Doing this can greatly improve accuracy on a set of targets for a
specific kinetic model. However, these fits are often not unique. In many cases this results
in parameters that are far from their physical values and are strongly correlated with each
other. When these correlated parameters are used out of context they often contribute to
significant inaccuracies in constructed mechanisms. Apart from libraries based on popu-
lar literature mechanisms and named appropriately, the thermochemistry and kinetic data
within the RMG Database is entirely free from fitted parameters. Great care should be taken
when combining fitted parameters with any other data sources.
The second challenge is a result of a convenient cancellation of errors that occurs in the
thermochemical parameters of chemical kinetic mechanisms. Kinetic mechanism simulations
are only dependent on the heats of reaction and Gibbs free energies of reaction, not the ther-
modynamic properties of individual species. Quantum chemical calculations in particular
have many correlated errors that cancel out conveniently in these differences. However, when
mixed with other thermochemistry data from other sources these convenient error cancella-
tions can disappear, resulting in inaccurate estimates of reaction thermochemical properties.
The RMG-database handles this situation by ensuring all ab initio thermochemical param-
eters in curated libraries are corrected using AECs and either BACs or other cancellation
schemes that reference to high accuracy experimental values. All of RMG’s thermochemical
estimators are trained on corrected ab initio or experimental data.
20
What is the RMG Database Good for Doing?
The RMG database has predictors for many properties as discussed above, but it is worth
keeping in mind that the RMG database has been primarily developed to further property
prediction that improves the accuracy of reaction mechanisms produced by the RMG soft-
ware. This means that one should expect the RMG database to be particularly good at
predicting reaction rates and thermochemistry, and any properties they are sensitive to. In
general the existence of a published, experimentally-validated RMG mechanism involving a
specific chemistry means some care has been taken inside the RMG database with regards
to those chemistries and that at least the most relevant corrections for those reaction condi-
tions are present. For example Grinberg Dana et al. 2018 demonstrated RMG’s capabilities
on Nitrogen compounds, Class et al. 2019 demonstrated RMG’s capabilities on some Sul-
fur chemistries and Chu et al. 2019 demonstrated RMG’s capabilities for small aromatic
species.28,30,73 However, the user should keep in mind that all large databases, including the
RMG database, likely include some erroneous values - perfection is hard to achieve, and it
is challenging to check tens of thousands of numbers.
The RMG database has been repeatedly shown to be sufficiently accurate that it can be
used to build satisfactory higher temperature (> 650 K) models on non-aromatic C/H/O
chemistries for combustion and pyrolysis. It can also handle up to two ring aromatic
chemistries under pyrolysis conditions. In the 400 − 650 K range rate predictions tend
to be less accurate due to higher sensitivity to the errors in estimates of activation barriers
and enthalpies. The lower accuracy makes prediction of product yields and overall reaction
rates much less robust. Most small nitrogen and sulfur functional groups are covered in
the database. RMG can handle their small molecule chemistries quite well, including some
nitrogen – sulfur interactions. However, for more complex hetero-molecules there are gaps
in the rate rules and thermodynamic data groups.
RMG’s methods for estimating solvation energies for small neutral molecules have been
well-tested against experimental data, and are particularly good for common solvents whose
21
properties are well-known. RMG’s ability to predict solvent effects on rate coefficients,
transport properties, etc. have not been very extensively tested yet. At present RMG
cannot handle ions.
There are some successful examples of RMG building satisfactory kinetic models for
reactions over metal catalysts, implying that the RMG estimates are accurate enough for
those cases.74 But only a few systems have been studied so far, and it will be some time
before RMG could estimate that parameters needed to predict catalysts over all the metals
of interest.
How can one efficiently use the RMG database?
The easiest way to access information from the RMG database is the RMG website, located
at https://rmg.mit.edu/. The web pages provide a graphic user interface (GUI) for browsing
through the database entries and searching for specific properties of a molecule or reaction of
interest, displaying data in both numerical and corresponding graphical form. The Molecule
Search tool enables users to search the database for all thermochemistry and transport
sources for a given species, as well as exploring the resonance structures thereof.73 The
solvation search tool does the same for solvent and solvation parameters for given solute
solvent pairs. The kinetic search tool allows users to search the database for reactions
based on just reactants or reactants and products. It is worth noting that in many cases in
addition to available library data the website will return rate coefficient estimates from both
rate rules and group additivity. The latter are generated using a reaction group additivity
scheme based on the rate tree and training reactions. This group additivity scheme is no
longer maintained as it was found to be appreciably less accurate than rate rules.19
When estimates for many species or reactions are required it becomes much more practical
to directly interface with the RMG software. The RMG Python package can be easily
installed from Anaconda and can be used to generate bulk or on-the-fly estimates from the
RMG database.
22
Conclusions
We have presented the RMG database for chemical property prediction. The RMG database
is an ever-expanding database designed for estimating properties important to chemical
kinetics. Over the years, it has been expanded and curated to accurately predict properties
for many chemical kinetic mechanisms.28,30,32,42,75,76 Currently the database is a very solid
foundation for building gas-phase mechanisms involving CHONSFClBr species. Property
prediction in the liquid and catalyst phases is state of the art and advancing rapidly. This
suite includes estimators for thermochemistry including solvent and adsorption corrections,
kinetics for 87 different classes of reactions, solvent viscosity, and species diffusivity in both
gas and liquid phases. Every addition to the RMG database is tested in an automated
workflow to ensure estimator performance is maintained or improved as the RMG database
is expanded. While there are many databases available today, the RMG database’s focus
on chemical mechanism generation makes it uniquely suited to for providing properties for
construction of chemical kinetic mechanisms.
Data and Software Availability
Most of the data presented is easily accessible through graphical user interfaces provided
through the RMG website at https://rmg.mit.edu. The database in its entirety is avail-
able on Github at https://github.com/ReactionMechanismGenerator/RMG-database. The
software used to access and manipulate the database is also available on Github in a sepa-
rate repository at https://github.com/ReactionMechanismGenerator/RMG-Py. These two
packages can be easily installed using Anaconda from the rmg channel.
23
Acknowledgments
Funding from the Gas Phase Chemical Physics Program of the US Department of Energy,
Office of Basic Energy Sciences, Division of Chemical Sciences, Geosciences, and Biosciences
(under award number DESC0014901) is appreciated. The work was also supported by the
U.S. Department of Energy, Office of Science, Basic Energy Sciences, through the Exas-
cale Catalytic Chemistry (ECC) Project as part of the Computational Chemical Sciences
Program, and by the National Science Foundation under Grant No. 1751720. A.G.D. was
supported by The George J. Elbaum Scholarship in Engineering, The Ed Satell Foundation,
and The Zuckerman STEM Leadership Program. H.P. was supported by the Think Global
Education Trust Scholarship.
References
(1) Manion, J. A.; Huie, R. E.; Levin, R. D.; Jr., D. R. B.; Orkin, V. L.; Tsang, W.;
McGivern, W. S.; Hudgens, J. W.; Knyazev, V. D.; Atkinson, D. B.; Chai, E.;
Tereza, A. M.; Lin, C.-Y.; Allison, T. C.; Mallard, W. G.; Westley, F.; Herron, J. T.;
Hampson, R. F.; Frizzell, D. H. NIST Chemical Kinetics Database, NIST Standard
Reference Database. https://kinetics.nist.gov/kinetics/welcome.jsp.
(2) Ruscic, B.; Pinzon, R. E.; Von Laszewski, G.; Kodeboyina, D.; Burcat, A.; Leahy, D.;
Montoy, D.; Wagner, A. F. Active Thermochemical Tables: Thermochemistry for the
21st century. Journal of Physics: Conference Series. 2005; pp 26–30.
(3) PrIMe Kinetics. http://primekinetics.org/.
(4) Chemked. http://www.chemked.com/.
(5) Weber, B. W.; Niemeyer, K. E. ChemKED: A Human- and Machine-Readable Data
Standard for Chemical Kinetics Experiments. International Journal of Chemical Ki-
netics 2018, 50, 135–148.
24
(6) CloudFlame. https://cloudflame.kaust.edu.sa/.
(7) Goteng, G. L.; Nettyam, N.; Sarathy, S. M. CloudFlame: Cyberinfrastructure for com-
bustion research. Proceedings - 2013 International Conference on Information Science
and Cloud Computing Companion, ISCC-C 2013. 2014; pp 294–299.
(8) Smith, D. G. A.; Altarawy, D.; Burns, L. A.; Welborn, M.; Naden, L. N.; Ward, L.;
Ellis, S.; Pritchard, B. P.; Crawford, T. D. The MolSSI QC Archive project: An open-
source platform to compute, organize, and share quantum chemistry data. WIREs
Computational Molecular Science 2020, 11, 1491.
(9) Benson, S. W.; Golden, D. M.; Haugen, G. R.; Shaw, R.; Cruickshank, F. R.;
Rodgers, A. S.; O’neal, H. E.; Walsh, R. Additivity rules for the estimation of thermo-
chemical properties. Chemical Reviews 1969, 69, 279–324.
(10) Eigenmann, H. K.; Golden, D. M.; Benson, S. W. Revised group additivity parameters
for the enthalpies of formation of oxygen-containing organic compounds. Journal of
Physical Chemistry 1973, 77, 1687–1691.
(11) Vandewiele, N. M.; Van Geem, K. M.; Reyniers, M. F.; Marin, G. B. Genesys: Kinetic
model construction using chemo-informatics. Chemical Engineering Journal 2012, 207-
208, 526–538.
(12) Ritter, E. R.; Bozzelli, J. W. THERM: Thermodynamic property estimation for gas
phase radicals and molecules. International Journal of Chemical Kinetics 1991, 23,
767–778.
(13) Li, Y. P.; Han, K.; Grambow, C. A.; Green, W. H. Self-Evolving Machine: A Continu-
ously Improving Model for Molecular Thermochemistry. Journal of Physical Chemistry
A 2019, 123, 2142–2152.
25
(14) Grambow, C. A.; Pattanaik, L.; Green, W. H. Deep Learning of Activation Energies.
The journal of physical chemistry letters 2020, 11, 2992–2997.
(15) Han, K. Enabling Automatic Generation of Accurate Kinetic Models for Complicated
Chemical Systems. Ph.D. thesis, Massachusetts Institute of Technology, 2018.
(16) Sirumalla, S. K. Graph neural networks and high throughput quantum chemistry work-
flows for detailed kinetic modeling. Ph.D. thesis, 2021.
(17) Van de Vijver, R.; Sabbe, M. K.; Reyniers, M.-F.; Van Geem, K. M.; Marin, G. B. Ab
initio derived group additivity model for intramolecular hydrogen abstraction reactions.
Physical Chemistry Chemical Physics 2018, 20, 10877–10894.
(18) West, R. H.; Allen, J. W.; Green, W. H. ChemInform Abstract: Automatic Reaction
Mechanism Generation with Group Additive Kinetics. ChemInform 2012, 43 .
(19) Allen, J. W. Predictive Chemical Kinetics : Enabling Automatic Mechanism Generation
and Evaluation. Massachusetts Institute of Technology 2013, 02139, 59–64.
(20) Johnson, M. S.; Green, W. H. A Machine Learning Based Approach to Reaction Rate
Estimation. ChemRxiv 2022,
(21) Heid, E.; Goldman, S.; Sankaranarayanan, K.; Coley, C. W.; Flamm, C.; Green, W. H.
EHreact: Extended Hasse Diagrams for the Extraction and Scoring of Enzymatic Reac-
tion Templates. Journal of Chemical Information and Modeling 2021, 61, 4949–4961.
(22) Gao, C. W.; Allen, J. W.; Green, W. H.; West, R. H. Reaction Mechanism Generator:
Automatic construction of chemical kinetic mechanisms. Computer Physics Communi-
cations 2016, 203, 212–225.
(23) Wong, B. M.; Matheu, D. M.; Green, W. H. Temperature and molecular size dependence
of the high-pressure limit. Journal of Physical Chemistry A 2003,
26
(24) Dana, A. G.; Johnson, M.; Allen, J.; Sharma, S.; Raman, S.; Liu, M.; Gao, C.; Gram-
bow, C.; Goldman, M.; Ranasinghe, D.; Gillis, R.; Payne, A. M.; Li, Y.-P.; Dames, E.;
Buras, Z.; Vandewiele, N.; Yee, N.; Merchant, S.; Buesser, B.; Class, C.; Goldsmith, F.;
West, R.; Green, W. Automated Reaction Kinetics and Network Exploration (Arkane):
A Statistical Mechanics, Thermodynamics, Transition State Theory, and Master Equa-
tion Software. ChemRxiv 2022,
(25) Allen, J. W.; Goldsmith, C. F.; Green, W. H. Automatic estimation of pressure-
dependent rate coefficients. Phys. Chem. Chem. Phys. 2012, 14, 1131–1155.
(26) Johnson, M. S.; Green, W. H. Examining the Accuracy of Methods for Obtaining
Pressure Dependent Rate Coefficients. Faraday Discussions 2022,
(27) Class, C. A.; Liu, M.; Vandeputte, A. G.; Green, W. H. Automatic mechanism gener-
ation for pyrolysis of di-tert-butyl sulfide. Physical Chemistry Chemical Physics 2016,
18, 21651–21658.
(28) Class, C. A.; Vasiliou, A. K.; Kida, Y.; Timko, M. T.; Green, W. H. Detailed kinetic
model for hexyl sulfide pyrolysis and its desulfurization by supercritical water. Physical
Chemistry Chemical Physics 2019, 21, 10311–10324.
(29) Vandewiele, N. M.; Magoon, G. R.; Van Geem, K. M.; Reyniers, M.-F.; Green, W. H.;
Marin, G. B. Kinetic Modeling of Jet Propellant-10 Pyrolysis. Energy & Fuels 2015,
29, 413–427.
(30) Chu, T. C.; Buras, Z. J.; Oßwald, P.; Liu, M.; Goldman, M. J.; Green, W. H. Modeling
of aromatics formation in fuel-rich methane oxy-combustion with an automatically
generated pressure-dependent mechanism. Physical Chemistry Chemical Physics 2019,
21, 813–832.
(31) Gillis, R. J.; Green, W. H. Thermochemistry Prediction and Automatic Reaction Mech-
27
anism Generation for Oxygenated Sulfur Systems: A Case Study of Dimethyl Sulfide
Oxidation. ChemSystemsChem 2020, 2, e1900051.
(32) Johnson, M. S.; Nimlos, M. R.; Ninnemann, E.; Laich, A.; Fioroni, G. M.; Kang, D.;
Bu, L.; Ranasinghe, D.; Khanniche, S.; Goldsborough, S. S.; Vasu, S. S.; Green, W. H.
Oxidation and pyrolysis of methyl propyl ether. International Journal of Chemical
Kinetics 2021, 53, 915–938.
(33) Sumathi, R.; Green, W. H. Thermodynamic properties of ketenes: Group additivity
values from quantum chemical calculations. Journal of Physical Chemistry A 2002,
106, 7937–7949.
(34) Sumathi, R.; Green, W. H. Missing thermochemical groups for large unsaturated hy-
drocarbons: Contrasting predictions of G2 and CBS-Q. Journal of Physical Chemistry
A 2002, 106, 11141–11149.
(35) Sumathi, R.; Green, W. H. Oxygenate, oxyalkyl and alkoxycarbonyl thermochem-
istry and rates for hydrogen abstraction from oxygenates. Physical Chemistry Chemical
Physics 2003, 5, 3402–3417.
(36) Farina, D. S.; Sirumalla, S. K.; Mazeau, E. J.; West, R. H. Extensive High-Accuracy
Thermochemistry and Group Additivity Values for Halocarbon Combustion Modeling.
Industrial and Engineering Chemistry Research 2021, 60, 15492–15501.
(37) Lay, T. H.; Bozzelli, J. W.; Dean, A. M.; Ritter, E. R. Hydrogen atom bond increments
for calculation of thermodynamic properties of hydrocarbon radical species. Journal of
Physical Chemistry 1995, 99, 14514–14527.
(38) Han, K.; Jamal, A.; Grambow, C. A.; Buras, Z. J.; Green, W. H. An Extended Group
Additivity Method for Polycyclic Thermochemistry Estimation. International Journal
of Chemical Kinetics 2018, 50, 294–303.
28
(39) Lai, L.; Khanniche, S.; Green, W. H. Thermochemistry and Group Additivity Values
for Fused Two-Ring Species and Radicals. The Journal of Physical Chemistry A 2019,
123, 3418–3428.
(40) Ince, A.; Carstensen, H.-H.; Reyniers, M.-F.; Marin, G. B. First-principles based group
additivity values for thermochemical properties of substituted aromatic compounds.
AIChE Journal 2015, 61, 3858–3870.
(41) Ince, A.; Carstensen, H.; Sabbe, M.; Reyniers, M.; Marin, G. B. Group additive model-
ing of substituent effects in monocyclic aromatic hydrocarbon radicals. AIChE Journal
2017, 63, 2089–2106.
(42) Blondal, K.; Jelic, J.; Mazeau, E.; Studt, F.; West, R. H.; Goldsmith, C. F. Computer-
Generated Kinetics for Coupled Heterogeneous/Homogeneous Systems: A Case Study
in Catalytic Combustion of Methane on Platinum. Industrial and Engineering Chem-
istry Research 2019, 58, 17682–17691.
(43) Goldsmith, C. F.; West, R. H. Automatic Generation of Microkinetic Mechanisms for
Heterogeneous Catalysis. Journal of Physical Chemistry C 2017, 121, 9970–9981.
(44) Reymond, J.-L.; Ruddigkeit, L.; Blum, L.; van Deursen, R. The enumeration of chemical
space. Wiley Interdisciplinary Reviews: Computational Molecular Science 2012, 2, 717–
733.
(45) Grambow, C. A.; Li, Y.-P.; Green, W. H. Accurate Thermochemistry with Small Data
Sets: A Bond Additivity Correction and Transfer Learning Approach. J. Phys. Chem.
A 2019, 123, 5826–5835.
(46) Curtiss, L. A.; Redfern, P. C.; Raghavachari, K. Gaussian-4 theory. Journal of Chemical
Physics 2007, 126, 084108.
29
(47) Dana, A. G.; Johnson, M.; Allen, J.; Sharma, S.; Raman, S.; Liu, M.; Gao, C.; Gram-
bow, C.; Goldman, M.; Ranasinghe, D.; Gillis, R.; Payne, A. M.; Li, Y.-P.; Dames, E.;
Buras, Z.; Vandewiele, N.; Yee, N.; Merchant, S.; Buesser, B.; Class, C.; Goldsmith, F.;
West, R.; Green, W. Automated Reaction Kinetics and Network Exploration (Arkane):
A Statistical Mechanics, Thermodynamics, Transition State Theory, and Master Equa-
tion Software. 2022,
(48) McGill, C.; Forsuelo, M.; Guan, Y.; Green, W. H. Predicting Infrared Spectra with
Message Passing Neural Networks. 2021, 61, 2594–2609.
(49) Abraham, M. H.; McGowan, J. C. The use of characteristic volumes to measure cavity
terms in reversed phase liquid chromatography. Chromatographia 1987, 23, 243–246.
(50) Zhao, Y. H.; Abraham, M. H.; Zissimos, A. M. Determination of McGowan Volumes
for Ions and Correlation with van der Waals Volumes. Journal of Chemical Information
and Computer Sciences 2003, 43, 1848–1854.
(51) Viswanath, D. S.; Ghosh, T. K.; Prasad, D. H.; Dutt, N. V.; Rani, K. Y. Viscosity of
Liquids: Theory, Estimation, Experiment, and Data; Springer Netherlands, 2007; pp
1–660.
(52) Bloxham, J. C.; Redd, M. E.; Giles, N. F.; Knotts, T. A.; Wilding, W. V. Proper Use
of the DIPPR 801 Database for Creation of Models, Methods, and Processes. Journal
of Chemical and Engineering Data 2021, 66, 3–10.
(53) Joback, K. G.; Reid, R. C. Estimation of Pure-Component Properties from Group-
Contributions. Chemical Engineering Communications 1987, 57, 233–243.
(54) Joback, K. G. A unified approach to physical property estimation using multivariate
statistical techniques. Ph.D. thesis, Massachusetts Institute of Technology, 1984.
30
(55) Devotta, S.; Pendyala, V. R. Modified Joback Group Contribution Method for Nor-
mal Boiling Point of Aliphatic Halogenated Compounds. Industrial and Engineering
Chemistry Research 1992, 31, 2042–2046.
(56) Tee, L. S.; Gotoh, S.; Stewart, W. E. Molecular parameters for normal fluids: Lennard-
Jones 12-6 Potential. Industrial and Engineering Chemistry Fundamentals 1966, 5,
356–363.
(57) Abraham, M. H.; Platts, J. A.; Hersey, A.; Leo, A. J.; Taft, R. W. Correlation and
estimation of gas-chloroform and water-chloroform partition coefficients by a linear free
energy relationship method. Journal of Pharmaceutical Sciences 1999, 88, 670–679.
(58) Mintz, C.; Clark, M.; Acree, W. E.; Abraham, M. H. Enthalpy of Solvation Correlations
for Gaseous Solutes Dissolved in Water and in 1-Octanol Based on the Abraham Model.
Journal of Chemical Information and Modeling 2007, 47, 115–121.
(59) Chung, Y.; Gillis, R. J.; Green, W. H. Temperature-dependent vapor–liquid equilibria
and solvation free energy estimation from minimal data. AIChE Journal 2020, 66 .
(60) Bell, I. H.; Wronski, J.; Quoilin, S.; Lemort, V. Pure and pseudo-pure fluid thermophys-
ical property evaluation and the open-source thermophysical property library coolprop.
Industrial and Engineering Chemistry Research 2014, 53, 2498–2508.
(61) Chung, Y.; Vermeire, F. H.; Wu, H.; Walker, P. J.; Abraham, M. H.; Green, W. H.
Group Contribution and Machine Learning Approaches to Predict Abraham Solute
Parameters, Solvation Free Energy, and Solvation Enthalpy. 2022, 62, 433–446.
(62) Johnson III, R. D. NIST computational chemistry comparison and benchmark database,
NIST standard reference database number 101. http://cccbdb.nist.gov/ 2020,
(63) Ruscic, B.; Bross, D. Active Thermochemical Tables (ATcT) values based on ver. 1.122d
31
of the Thermochemical. https://atct.anl.gov/Thermochemical\%20Data/version\
%201.122d/index.php, 2018.
(64) Petersson, G. A.; Malick, D. K.; Wilson, W. G.; Ochterski, J. W.; Montgomery Jr, J. A.;
Frisch, M. J. Calibration and comparison of the Gaussian-2, complete basis set, and
density functional methods for computational thermochemistry. Journal of Chemical
Physics 1998, 109, 10570–10579.
(65) Anantharaman, B.; Melius, C. F. Bond additivity corrections for G3B3 and G3MP2B3
quantum chemistry methods. Journal of Physical Chemistry A 2005, 109, 1734–1747.
(66) Spiekermann, K.; Pattanaik, L.; Green, W. H. High accuracy barrier heights, enthalpies,
and rate coefficients for chemical reactions.
(67) Pedley, J. Thermochemical data and structures of organic compounds ; CRC Press, 1994;
Vol. 1.
(68) Burke, M. P.; Chaos, M.; Ju, Y.; Dryer, F. L.; Klippenstein, S. J. Comprehensive
H2/O2 kinetic model for high-pressure combustion. International Journal of Chemical
Kinetics 2012, 44, 444–474.
(69) Hashemi, H.; Christensen, J. M.; Gersen, S.; Levinsky, H.; Klippenstein, S. J.; Glar-
borg, P. High-pressure oxidation of methane. Combustion and Flame 2016, 172, 349–
364.
(70) Bugler, J.; Somers, K. P.; Simmie, J. M.; Güthe, F.; Curran, H. J. Modeling Nitrogen
Species as Pollutants: Thermochemical Influences. Journal of Physical Chemistry A
2016, 120, 7192–7197.
(71) Glarborg, P.; Miller, J. A.; Ruscic, B.; Klippenstein, S. J. Modeling nitrogen chemistry
in combustion. 2018, 67, 31–68.
32
(72) Song, Y.; Hashemi, H.; Christensen, J. M.; Zou, C.; Haynes, B. S.; Marshall, P.; Glar-
borg, P. An Exploratory Flow Reactor Study of H <sub>2</sub> S Oxidation at
30-100 Bar. International Journal of Chemical Kinetics 2017, 49, 37–52.
(73) Dana, A. G.; Buesser, B.; Merchant, S. S.; Green, W. H. Automated Reaction Mecha-
nism Generation Including Nitrogen as a Heteroatom. International Journal of Chem-
ical Kinetics 2018, 50 .
(74) Mazeau, E. J.; Satpute, P.; Blöndal, K.; Goldsmith, C. F.; West, R. H. Automated
Mechanism Generation Using Linear Scaling Relationships and Sensitivity Analyses
Applied to Catalytic Partial Oxidation of Methane. ACS Catalysis 2021, 11, 7114–
7125.
(75) Class, C. A.; Liu, M.; Vandeputte, A. G.; Green, W. H. Automatic mechanism gener-
ation for pyrolysis of di-tert-butyl sulfide. Physical Chemistry Chemical Physics 2016,
18 .
(76) Dames, E. E.; Rosen, A. S.; Weber, B. W.; Gao, C. W.; Sung, C. J.; Green, W. H.
A detailed combined experimental and theoretical study on dimethyl ether/propane
blended oxidation. Combustion and Flame 2016, 168, 310–330.
33