Introduction To Chemical Kinetics

Chapter 2
Introduction to Chemical Kinetics
Abstract After introducing the basis of the Arrhenius equation and its relationship
to transition-state theory, forms of global chemical kinetic models are summarized,
including shrinking core and pseudo nth-order reactions; sigmoidal reactions such
as sequential, random scission, autocatalytic, logistic, and nucleation-growth
model; and distributed reactivity models, including continuous and discrete acti-
vation energy distribution models. Isoconversional and model fitting methods for
deriving chemical kinetic models are described, including how to use simple kinetic
analyses to derive initial guesses for nonlinear regression of complex models.
Common errors that lead to erroneous Arrhenius parameters are outlined.

Keywords Isoconversional Model fitting Prout-tompkins model Sigmoidal
reactions Autocatalytic reactions
Random scission reactions
Distributed

reactivity model DAEM
Chemical kinetics describe of the rate at which chemical structures change.

Chemical kinetics can be fundamental (involving quantum mechanics, statistical
mechanics, or both) or phenomenological (semi-empirical rate laws derived from
simplified physical models). The latter are often also called global kinetics or
engineering kinetics, because they are derived primarily for application purposes
instead of a truly fundamental understanding. They usually deal with lumped
species and reactions, that is, reactions that represent the net progress of a set of
related reactions. This chapter and book address phenomenological kinetics. An
important warning up front is that there are probably more bad published papers on
phenomenological (global) kinetics than good ones, so an understanding of this
field is necessary to separate the wheat from the chaff. Results are presented for
non-fossil fuels to demonstrate various reaction characteristics that should be
considered for fossil fuels.
© Springer International Publishing AG 2017 25

A.K. Burnham, Global Chemical Kinetics of Fossil Fuels,
DOI 10.1007/978-3-319-49634-4_2
26 2 Introduction to Chemical Kinetics
2.1 The Activation Energy and Frequency Factor
Most reactions must traverse an energy barrier to change from one chemical
structure to another, as shown schematically in Fig. 2.1. Ef and Er are the forward
and reverse activation energies. A reaction can be unimolecular (one chemical
breaking down) or multi-molecular. It can either absorb heat (endothermic) by
forming products with weaker bonds or emit heat (exothermic) by forming products
with stronger bonds. The amount of heat absorbed or emitted is the enthalpy, which
equals the difference between the forward and reverse activation energies, ΔE, at
least within the simplified view of Fig. 2.1.
The reaction rate typically increases exponentially with increasing temperature
in a fashion first parameterized by Arrhenius in 1889 [1]. Specifically, the rate
constant k increases according to the formula
k ¼ AeE=RT ; ð2:1Þ
where A is a frequency factor, as it has units of reciprocal time in addition to other

units, depending on the kind of reaction, E is the activation energy for the reaction,
R is the universal gas constant, and T is the absolute temperature. R has a value of
1.987 cal/(mol K) or 8.314 J/(mol K), depending on the units chosen for energy. In
Fig. 2.1 Schematic

representation of the chemical k
reaction barrier
reactant(s) products
k is the reaction rate constant

k has units of reciprocal time
The reaction rate has units of quantity per unit time
Endothermic reactions
Er
Ef
products
reactant(s) ΔE positive
Exothermic reactions
Ef Er
reactant(s)
ΔE negative products
2.1 The Activation Energy and Frequency Factor 27
practice, it is far simpler to use activation energies in Kelvins to avoid a set of unit
conversions and multiplications during usage.
The reason for this temperature dependence is shown simplistically in Fig. 2.1.
At any given temperature, molecules have a statistical distribution of energies,
commonly a Boltzmann distribution [2]. For most reactions, only a small fraction of
the molecules have enough energy to get over the energy barrier of the reaction.
That fraction increases with temperature. For a bimolecular reaction, the fraction
depends on the fraction of molecules having a velocity greater than some threshold
value in the Maxwell-Boltzmann distribution [3]. For unimolecular decomposition,
the Boltzmann distribution describes the probability that a given bond will be in an
energetic-enough state to dissociate. Some reactions have no reaction barrier—
notably free-radical recombination reactions, which are important in hydrocarbon
cracking.
Multi-molecular gas-phase reactions typically have power-law temperature
dependence in addition to the exponential dependence to account for collision
frequency
k ¼ AT b eE=RT ; ð2:2Þ
where b is ranges from 0.5 to 1.5. However, one has to traverse several hundred
degrees before pre-exponential temperature dependence has any practical signifi-
cance. Burnham and Braun [4] give the example of an ideal linear temperature
coefficient absorbed into an effective activation energy. Over a 50 °C range, the
effective activation energy predicts the true temperature dependence to an accuracy
of one part in 104. Extending that example, the reaction rate predicted by the
effective rate constant is within 10 % of the ideal value over a 700 °C range. Given
that the error is only in one direction, the overall error could be reduced to no more
than 5 % by not requiring exact agreement at the calibration temperature. Using a
pre-exponential temperature dependence makes sense only when it is known from
first principles, and that is commonly the case for detailed mechanistic modeling of
hydrocarbon pyrolysis and combustion reactions.
The reaction rate constant is often characterized by transition-state theory [5],
which hypothesizes an activated complex at the peak of the reaction barrier. The
enthalpy of formation of that activated complex, ΔH, is the activation energy. For
unimolecular decomposition reactions, the frequency factor is related to the
vibrational frequency of the bond to be broken in the active complex and is a
function of Boltzmann’s constant, kb, Planck’s constant, h, the absolute tempera-
ture, and the entropy of formation of the activated complex, ΔS. This gives a rate
constant

kB T DG=RT kB T DS=R DH=RT
k¼ e ¼ e e : ð2:3Þ
h h
For typical pyrolysis temperatures, the kBT/h is a little over 1013 s−1, and the
entropy term is often ignored.
While some people studying complex reactions in a phenomenological or global

manner refer to transition-state theory and activated complexes, such references are
usually ill-placed and more often give an indication that the person is naïve con-
cerning the huge simplifications involved in their work relative to fundamental
reaction mechanisms.
For a simple unimolecular decomposition reaction, the activation energy is equal
to the energy of the bond to be broken. However, such simple reactions are rarely of
practical interest. Instead, reactions of interest, especially for fossil fuels, involve
chains of reactions. For example, if a normal hydrocarbon splits into two fragments,
the activation energy of that initiation reaction is equal to the strength of a
carbon-carbon bond, or about 82 kcal/mol. However, the molecular fragments are
unstable and can either split an ethylene molecule off their end or start attacking
their neighbors by abstracting a hydrogen atom from the middle of a neighboring
chain, thereby converting a primary free radical to a more stable secondary free
radical. In the latter case, the secondary free radical can decompose into an alkene
and another primary free radical. These latter steps are called propagation reactions.
Sometimes two free radicals run into each other and form a stable hydrocarbon,
often branched. This is called a termination reaction and usually has zero activation
energy. Since the breaking of one strong bond results through a chain reaction to
break many bonds through propagation reactions, which have lower activation
energies, the overall activation energy for hydrocarbon cracking is in the range of
50–60 kcal/mol, dependent upon reaction conditions.
2.2 Introduction to Conversion and Pressure Dependence
It is commonly assumed for practical purposes that the dependences on conversion,

temperature, and pressure can be separated [6]. Before proceeding, it should be
understood that chemical reaction networks are expressed as differential equations
describing the disappearance and appearance of reactants and products. Many
reactions of interest are expressed as a single differential equation giving the rate
law for reactants going to products, and in this case, some write the rate law in
terms of disappearance of reactants, x or some other symbol, and some use the
appearance of products, α. Obviously, x = 1 − α.
The resulting differential equations for either x or α are given in equations
dx=dt ¼ kðTÞf ð1 xÞhðPÞ ð2:4Þ
and
da=dt ¼ kðTÞf ðaÞhðPÞ: ð2:5Þ
where P is pressure. Functions for f(α) and the integral form, g(α), are commonly
tabulated in the thermal analysis literature for solid-state reactions, where
2.2 Introduction to Conversion and Pressure Dependence 29
Z a
da
gðaÞ ¼ : ð2:6Þ
0 f ðaÞ
One selection of such functions is given in Table 2.1. These equations deal with
situations involving nucleation-growth characteristics, shrinking cores, intraparticle
diffusion resistance, and distributed reactivity (e.g., power law in time). The
Avrami-Erofe’ev and Prout-Tompkins models are alternative mathematical
approaches to treat sigmoidal reactions (nucleation-growth), and they can also be
used for generalized initiation-propagation reactions. The Jander and
Ginstling-Brounshtein (G-B) equations are included for completeness and are
reasonably accurate up to about 50 % conversion, but more accurate equations for
planar, cylindrical and spherical particles have been derived [7].
The h(P) function can be empirical or fundamentally based, and separating the
pressure dependence is an approximation that is not always valid. The most obvious
example of pressure dependence is for a gas-solid reaction, for which the reaction
rate depends on the gaseous reactant partial pressure. The pressure dependence could
be, for example, a power law or some type of adsorption isotherm [8]. Another
common example where it does work is a reversible mineral decomposition, e.g.,
CaCO3 ↔ CaO + CO2, where h(P) = 1 − P/Peq, P is the gaseous product partial
pressure, and Peq is the equilibrium vapor pressure. However, there are other cases
where the pressure dependence is not truly separable. In hydrocarbon cracking, a
change in pressure changes relative amounts of initiation, propagation, and termi-
nation reactions, which leads to a change in the global activation energy [9]. Another
example is where increasing pressure modestly accelerates the reaction rate due to
increased autocatalysis [10], whereas increasing pressure greatly decelerates the
Table 2.1 Selection of commonly used kinetic models

Reaction model Code Differential form, f(α) Integral form, g(α)
1st order F1 1−α −ln(1 − α)
Pseudo nth order Fn (1 − α)n 1/(n−1)[(1 − α)(1−n) − 1]
Contracting cylinder R2 2(1 − α)1/2 1 − (1 − α)1/2
Contracting sphere R3 3(1 − α) 2/3
1 − (1 − α)1/3
Power law Pν να (ν−1)/ν
α1/ν
Avrami-Erofe’ev Ap p(1 − α)[ −ln(1 − α)] (p−1)/p
[−ln(1 − α)]1/p
Extended ePT or (1 − α) αn m
No analytical solution
Prout-Tompkins Bna
1D diffusion D1 ½ α−1 α2
2D diffusion D2 [ −ln(1 − α)]−1 (1 − α)ln(1 − α) + α
1/3 −1
3D diffusion (Jander) D3 3/2(1 − α) [1 − (1 − α) ]
2/3
[1 − (1 − α)1/3]2
−1/3
3D diffusion (G-B) D4 3/2[(1 − α) − 1] 1 − 2α/3 − (1 − α)2/3
There is some variation in the code nomenclature and location of the constants in the integral and
differential forms
reaction rate due to increased diffusion resistance [11]. Pressure can also appear in an
activation volume term, which can be interpreted as an entropic contribution to the
transition-state free energy, but it is very hard to uniquely identify such contribution
for a complicated reaction system [12]. Pressure dependence will be taken up again
in later chapters.
Equations (2.4) and (2.5) also assume that one starts with a material that is
homogeneous. Materials of interest often have moieties with different reactivity.
Consequently, it is often appropriate to divide the reaction into a set of independent
parallel reactions, with or without the same pressure dependence:
dxi =dt ¼ ki ðTÞf ð1 xi Þ: ð2:7Þ
The sum of the initial xi equals one. The parallel reaction model can be extended
from a set of discrete reactions to a continuous distribution. Three common con-
tinuous distributions are Gaussian [13], Weibull [14] and Gamma distributions [15].
The Gaussian distribution has the disadvantage of being symmetric, but a Gaussian
distribution of nth-order reactions can achieve similar fits to the Weibull distribu-
tion [4].
Reactions are often grouped into three classes: acceleratory, deceleratory, and
sigmoidal. Sigmoidal reactions have an acceleratory phase followed by a deceler-
atory phase. In fact, chemical reactions with purely acceleratory character are rare.
A solid-state reaction that starts in the center of a sphere and then propagates
outward is an ideal example. A chain-branching reaction with a branching ratio
greater than one is another example. Runaway nuclear fission is a non-chemical
example of such an acceleratory reaction. Decomposition of nitro energetic mate-
rials initially accelerates due to autocatalysis, but if the released heat and pressure
are allowed to dissipate, they turn deceleratory and are then sigmoidal.
Most, if not all, reactions of interest in fossil fuels are either sigmoidal or
deceleratory. In contrast to the usual classification, it is useful to divide the
deceleratory reactions into three types: first-order reactions, whose reaction rates
decay exponentially, those that decay faster than an exponential, and those that
decay slower than an exponential. A plot of the logarithm of the reaction rate versus
time distinguishes the three types, as shown in Fig. 2.2, which is adapted from
Burnham and Braun [4].
In such a semi-log plot, nth-order reactions with n < 1 are concave downward,
which means that the reactant disappears faster than predicted from an initial
first-order reaction rate at low conversion. Pseudo nth-order reactions, with n > 1
are concave upward, which gives the opposite prediction. The adjective pseudo is
introduced, because the reactions are not nth-order in the sense of traditional
gas-phase or solution reaction kinetics. In these cases, reaction order is expressed in
terms of partial pressure or concentration. For pseudo nth-order reactions, the
reaction is parameterized in terms of reaction progress of a distinct phase, which is
normalized to zero or one, depending on whether one is using fraction reacted or
fraction remaining.
2.2 Introduction to Conversion and Pressure Dependence 31
0.1
Gaussian Distribution in E
nth-order
sigma = 2%
ReacƟon rate, s -1
0.01
n=2
sigma = 4%
0.001 n = 0.67
sigma = 0%
n=1
n = 0.5
0.0001
0 25 50 75 100 0 25 50 75 100 125
Time, s Time, s
Fig. 2.2 Plot of the logarithm of the reaction rate versus time showing that some models
decelerate faster than a first-order model and some decelerate slower. Adapted from Burnham and
Braun [4] courtesy of LLNL
A Gaussian distribution in activation energy, which is not the same as a

Gaussian distribution in k, is also concave upward, so it is similar in some respects
to a pseudo nth-order reaction with n > 1. This is because a pseudo nth-order
reaction is mathematically equivalent to having a Gamma distribution of frequency
factors [4]. Consequently, pseudo nth-order reactions are a legitimate way of
treating a reactivity distribution. Given that this book does not deal with traditional
nth-order reactions, the pseudo adjective will usually be dropped for the duration of
the book.
2.3 Integrating the Reaction Rate
While some reaction models can be integrated exactly for isothermal conditions,
most experiments and applications of practical interest involve either reaction
models or thermal histories for which an exact analytical solution is not available.
This section discusses the mathematical methods needed to deal with this situation.
This section closely follows Burnham and Braun [4].
The first concept is that there are generally two types of modeling of interest.
The first is exploratory modeling, and the second is application modeling.
Exploratory modeling explores various concepts using fast and flexible tools where
the ultimate in accuracy is not required. Application modeling uses more mature
models that have been programmed with more sophisticated and accurate numerical
methods. Often, the application modeling is done by someone other than the
computer code developer, so changes are more difficult. Numerical methods for
both types of activities are presented in this section.
A second concept is that neither isothermal nor constant-heating-rate conditions
occur in real-life applications. In fact, it is even hard to do truly isothermal
experiments over the range of temperatures of interest, particularly considering that
it is important to achieve high conversions over the entire range of temperatures to

fully characterize the reaction mechanism. Consequently, most of the literature
developing equations and algorithms for strictly isothermal or strictly constant
heating rates must be viewed as too limiting for modern computers. Once the
simple scoping calculations are done, it is just as easy to program highly accurate
algorithms that work for any arbitrary thermal history.
The starting point for all algorithms involving either a single reaction or a set of
parallel independent reactions is the integration Eq. (2.4) or (2.5) over time. This
approach does not use the assumption that k is constant, as does Eq. (2.6). The
following expression uses fraction remaining (2.4) rather than fraction reacted (2.5):
Z t
x¼ kðTðtÞÞf ð1 xÞdt: ð2:8Þ
0
For first and nth-order reactions, respectively, this equation evolves to

Z t
x ¼ exp kðTðtÞÞdt ð2:9Þ
0
and
Z t 1=ð1nÞ
x ¼ 1 ð1 nÞ kðTðtÞÞdt : ð2:10Þ
0
For isothermal conditions, k is constant, so one obtains the familiar equations
x ¼ exp½kt ð2:11Þ
and
x ¼ ½1 ð1 nÞkt1=ð1nÞ : ð2:12Þ
In the middle of the 20th century, it became fashionable to conduct kinetic

experiments using a constant heating rate to avoid the transient heat-up time
characteristic of isothermal experiments. At a constant heating rate, dT/dt = β, and
defining y = E/RT, Eq. (2.10) becomes
Z
AE 1 expðyÞ
x ¼ exp dy : ð2:13Þ
bR y y2
The inside integral over y, often called the “Temperature Integral,” does not have
an exact analytical solution. With a change of variables, it can be converted to the
well-known exponential integral that has many approximate solutions to any
desired accuracy. Numerous papers in the thermal analysis literature have explored
various approximations, and Flynn [16] gives a fairly complete review of that work.
2.3 Integrating the Reaction Rate 33
He also addresses alternate forms of the rate law. The most interesting alternative to
the conventional approach is that adding T2 to the pre-exponential factor leads to an
exact solution for a constant heating rate, which obviously has some justification as
discussed in the context of the previous section, but the case is not strong enough to
include it, given that dealing with the integral is not very difficult. Its solution can
be expressed as an infinite series as follows:

ART 2 2RT RT
E
x ¼ exp 1 þ e : ð2:14Þ
bE E
Although some German coal workers [17] and early LLNL oil shale papers [18]
take only the unity term, a more common approximation, generally attributed to
Coats and Redfern [19], is to include the 2RT/E term as well. This gets one most of
the way to the exact solution for large values of E/RT. Even so, Flynn criticizes this
approximation and says that it leads to publication of incorrect activation energies.
However, while better, yet still simple, approximations are available [20], e.g.,
(1 + 2RT/E)−1 and (1 + 4RT/E)−0.5, there are usually far more serious problems
causing inaccurate activation energies. Nevertheless, programming one of the
simple alternatives for scoping calculations is just as easy, so it might as well be
done.
The more important issue is the single-minded advocacy of higher level, more
complicated approximations to the Temperature Integral. Rational approximations
up to the eighth degree achieve essentially perfect accuracy [21, 22]. However,
concentrating on only the constant-heating-rate numerical solution misses the real
issue for deriving chemical kinetics—it is best to use a variety of heating schedules,
including nominally isothermal, constant heating rate, and even oscillatory ramps
[16] to get the most stringent test of any kinetic model. For this purpose, it is far
superior to adopt the approach of Braun and Burnham [23] that divides any
complex thermal history into a series of constant-heating-rate segments. Consider
the integration of the rate constant in Eqs. (2.9) and (2.10) as being broken into a
series of such steps:
Z t Xn Z ti
kðTðtÞÞdt ¼ i¼0
k ðTðtÞÞdt; ð2:15Þ
0 ti1
where
Z ( 2 " 2 # )
ti
1 yi þ a1 yi þ a2 yi1 þ a1 yi1 þ a2
k ðTðtÞÞdt ¼ Ti ki 1 2 Ti1 ki1 1 2
ti1 b ðyi þ b1 yi þ b2 Þ yi1 þ b1 yi1 þ b2
ð2:16Þ
and y = E/RT, a1 = 2.334733, a2 = 0.250621, b1 = 3.330657, and b2 = 1.681534.

Increased accuracy can be obtained using a five-term rational approximation [24],
but such accuracy is a matter of diminishing returns in the case of applied kinetics
where the true mechanism is not known well. If the heating rate, β, is zero
(isothermal), one can use either the isothermal solution or an extremely low heating
rate so that the change in temperature over the longest isothermal segment is a small
fraction of a degree.
For a first-order reaction, there is a very simple method to get a good estimate for
the extent of reaction over any thermal history. It takes advantage of the fact that the
amount of reactant over each increment of time is simply the exponential of the
time interval times the average rate constant over the integral, specifically
xi h i
¼ exp ðti ti1 Þ Ai eEi =RTi þ Ai1 eEi1 =RTi1 =2 : ð2:17Þ
xi1
Here a provision is made for a variable frequency factor as could be derived

from isoconversional kinetic analysis methods, in which A and E pairs are deter-
mined as a function of conversion as described in the next section. That makes the
calculation of the average rate constant over the interval a little more complicated—
an iterative process for best accuracy—but the principle is simple. Regardless of
whether A and E are independent of conversion or not, the total extent of reaction is
simply

h X i
xi ¼ Pi exp Dti kavg;i ¼ exp i
Dti kavg;i ð2:18Þ
There are many other ways to integrate differential equations, ranging from a
1-step explicit integration that does surprising well for small step sizes to multistep
implicit methods. A review of and instruction in numerical integration is outside the
scope of this book, and there are numerous texts on the subject. An advantage of
numerical integration over any of the methods discussed here is that any rate law
can be integrated without consideration of whether an analytical solution exists, and
systems of kinetic equations can be integrated as well. Various numerical inte-
gration packages are available, and LSODE [25] is the package used in many
applications developed at LLNL and elsewhere.
2.4 Methods of Kinetic Analysis
There are basically two types of methods for deriving a kinetic expression for a
condensed phase reaction—isoconversional analysis and model fitting. Both have
advocates, but since both have advantages and disadvantages, they are best used in
concert. Inconsistent activation energies from the two methods should raise a red
flag.
2.4 Methods of Kinetic Analysis 35
2.4.1 Isoconversional Methods
Isoconversional methods start with Eq. (2.4) or (2.5) with h(P) assumed to be unity.
Isoconversional methods are sometimes called model-free kinetics, because they do
not involving fitting to a phenomenological model, but they are not assumption
free. The basic assumption is that at any given extent of reaction, the same reactions
occur in the same ratio independent of temperature. It is easy to construct reaction
networks that do not follow the isoconversional principle. One example is a set of
parallel independent reactions that change in relative reactivity as a function of
temperature due to different activation energies [26]. Another example is the case of
competitive reactions with different activation energies, which causes the overall
reaction pathway to be different at different temperatures [27]. Nevertheless, some
isoconversional methods usually work pretty well and should be part of the toolkit
for developing a robust kinetic model. See Vyazovkin [28] for a thorough dis-
cussion of various approaches and applications.
Isoconversional methods were developed about 50 years ago. Friedman first
developed the differential method [29]. Ozawa and Flynn and Wall independently
developed the integral method [30, 31]. Flynn gives a couple short but educational
discussions of development of the integral methods [32, 33]. The general concept is
shown graphically in Fig. 2.3.
Friedman assumed no pressure dependence and differentiated Eq. (2.4) to obtain

dx
ln ¼ lnA þ lnf ðxÞ E=RT ð2:19Þ
dt
He then stated that f(x) can be assumed to be constant for constant values of x,
the definition of isoconversional, which he states is true if the chemical pathway is
Fig. 2.3 Graphical

representation of the
isoconversional kinetic
analysis concept from
Vyazovkin et al. [6].
Reprinted with permission.
© Elsevier
independent of temperature. Plotting the natural log of the reaction rate versus
1/T gives a slope of E/R and an intercept of ln[Af(x)]. He further noted that if the
reaction is nth order, a plot of ln[Af(x)] versus ln(x) would be linear with a slope of
n and an intercept of ln(A). In practice, E/R is frequently not constant over x, and the
plot to determine n is rarely linear, so the value of the method is different than
originally envisioned, as will be explained below.
Friedman originally derived his isoconversional method using fraction remain-
ing, x, but the normal convention in thermal analysis is to use fraction reacted, α.
Consequently, the intercept is normally given as ln[Af(α)]. Regardless, the function
is the same, and one can find f(x) from Table 2.1 by merely substituting x for 1 − α.
For a first-order reaction, f(x) = x = 1 − α, so A is merely the intercept divided by
the fraction remaining. If the conversion steps are small enough, those effective
first-order frequency factors can be used for each instantaneous reaction rate and
incremental fraction reacted.
Ozawa and Flynn and Wall independently observed that the approximation of
the temperature integral proposed by Doyle [34] leads to the equation
lnðbi Þ ¼ const 1:052Ea =RTa;i : ð2:20Þ
This is known as the OFW equation. Plotting ln(βi) versus 1/Ti at conversion α
gives a slope of E/R at that α.
A more accurate simplification of the temperature integral leads to the KAS
(Kissinger-Akahira-Sunose) equation [35]:
Aa R
ln bi =Ta;i
2
¼ ln½ Ea =RTa;i : ð2:21Þ
Ea gðaÞ
Slightly more accurate activation energies can be obtained by replacing the

square of the temperature with a power of 1.92 and adding a coefficient of 1.0008 in
front of the right-hand term [36], but this still does not address the issue that the
equation assumes that Eα is constant, and it rarely is. Vyazovkin [37] addresses this
issue by integrating the time or temperature integral in a way that the activation
energy is assumed to be constant only over small steps.
Z
ta
Ea
J½Ea ; Tðta Þ ¼ exp dt: ð2:22Þ
taDa RTðtÞ
Vyazovkin’s advanced integral method obtains activation energies practically

identical to those obtained by Friedman’s method [6].
For completeness and historical perspective, it is important to recognize the
Coats-Redfern equation and its adaptation to an isoconversional method. The
original Coats-Redfern method [19] fitted data from a single heating rate to the
equation

lnðxÞ E 2RT
ln ¼ þ ln AR 1 bE : ð2:23Þ
T2 RT E
This method is highly inaccurate and has been strongly discouraged by recent
ICTAC papers on kinetic analysis [6, 38]. Even so, this equation can be converted
into an isoconversional method by rearrangement [6]:

2RT
ln bi = Ta;i
2
1 ¼ ln½Aa R=ðEa lnðxÞÞ Ea =RTa;i : ð2:24Þ
E
This must be solved by iteration until convergence by substituting previous

values of Eα into the 1 − E/RT term. Further improvement in the method could be
obtained by using more accurate approximations for the 1 − E/RT term, but again,
this is a matter of diminishing returns, as the equation is fundamentally limited by
the assumption of constant activation energy when the purpose is often to determine
the conversion dependence of the activation energy.
One final isoconversional method to be mentioned only briefly at this point is the
model of Miura et al. [39, 40] He follows much earlier work by Vand [41] arguing
that due to overlapping parallel reactions that make up the reactivity distribution, an
effective fraction reacted rather than the actual fraction reacted should be used in
isoconversional methods. This method is discussed further in a later section.
Once the fundamental limitations were recognized for all integral isoconver-
sional methods except the advanced integral method, a recurring debate in the
thermal analysis community is whether the rate method or advanced integral
method is superior in view of the fact that real data have noise, both in terms of
random instrument noise and uncertainty in baselines.
In principle, integral data can be differentiated and rate data can be integrated,
but errors can be introduced by both differentiation and integration if not done
properly. A minimum requirement of any method is that successive differentiation
and integration of integral data must recover the original integral data (except, of
course, the initial value, as it corresponds to a constant baseline), and the corre-
sponding recovery of rate data from cyclic integration and differentiation.
Next, one can smooth rate data to reduce noise, but only to the extent that one
wants to resolve variations in A and E as a function of conversion. Resolution down
to 10 % conversion increments is adequate for some qualitative purposes, but
resolution down to 1 % increments is desirable for application modeling purposes.
Obviously, the original data must have at least 100 points over the conversion
interval for the latter analysis, and that is commonly available from thermal analysis
instruments.
One suggestion is that it is best to use integral methods on integral data (e.g.,
weight loss and product accumulation) and differential methods for rate data (e.g.,
differential scanning calorimetry and gas evolution). However, I have experience
only using Friedman’s methods to derive kinetics for application purposes from
both integral and rate data (using both the LLNL Kinetics2015 [42] and AKTS
Thermokinetics [43] programs) and do not see any fundamental limitations in using
Friedman’s method for both as long as the data are of high quality.
One final note is that when using isoconversional kinetics for applications,
isoconversional kinetics should really be classified as an infinitely sequential model.
One can assume, for example, that each small increment of conversion is governed
by a first-order reaction and derive the first-order frequency factor. Then the extent
of conversion over any thermal history is merely a series of first-order reactions
over the conversion intervals. This is a direct outgrowth of the isoconversional
assumption—that the same reaction progress occurs, independent of temperature,
and over each successive reaction interval.
Kissinger’s method [44] is a very simple method that is not universally an
isoconversional method but sometimes is and otherwise is usually close to it. It
involves using the maximum of the reaction rate as a function of heating rate to
derive an activation energy and pre-exponential factor. It is rigorously correct and
isoconversional for first-order reactions. It is an excellent approximation for
nth-order [45], nucleation-growth [46], and distributed reactivity reactions [23],
although Vyazovkin et al. [6] discuss some cases where it does not work. It is
important not to confuse Kissinger’s Tmax method with the KAS method, which is
rigorously isoconversional. Kissinger’s method is also known in the vacuum
community as Redhead’s method [47].
At the peak reaction rate, the derivative of the reaction rate (second derivative of
conversion) is equal to zero. Differentiating Eq. (2.4) with h(P) = 1, setting it equal
to zero, and eliminating a term outside the brackets yields

bE E df ðxÞ
0¼ 2
þ Aexp : ð2:25Þ
RTmax RTmax dx
For a first order reaction, df(x)/dx = − 1, so

b E AR
ln 2 ¼ þ ln : ð2:26Þ
Tmax RTmax E
A few final comments are appropriate concerning the strengths and limitations of
the isoconversional methods for fossil fuels. By their nature, isoconversional
methods recognize the possibility of reaction heterogeneity, which implies that
different products may be formed at different stages of the global conversion. If one
merely wants to calculate to global conversion of a starting material, they are
frequently as good as or better than model fitting, because they can more easily
accommodate conversion dependence of both A and E. Burnham and Braun [4]
discuss their use for coal pyrolysis, and Dieckmann [48] discusses their use for
modeling petroleum formation. Section 4.4.2 shows how they help diagnose that
dependence for typical oil-prone kerogens, and Sect. 7.1 shows how they predict a
broader range of temperature for petroleum generation.
However, they do have limitations. If one wants to model differences in gen-
erated products as a function of conversion, it is not straightforward with
isoconversional methods. If one wants to model two different measures of reaction

progress, the isoconversional kinetics for those measure can be completely differ-
ent, e.g., weight loss and heat evolution for a material with overlapping
endothermic and exothermic stages. Finally, isoconversional methods may not be
appropriate for reaction intermediates, because the combination of generation and
consumption reactions makes a simple definition of conversion difficult or
impossible, depending on the complexity of the reaction network.
2.4.2 Model Fitting Methods
Model fitting can take several forms, but in the end, all methods involve minimizing
the difference between a calculated and a measured conversion or rate as a function
of some time and temperature.
Linear regression [49] involves minimizing the residual sum of squared differ-
ences for the simple equation, y = mx + b. The solution for m is simply

P P P
i x i yi ð i xi i yi Þ=n
m¼ P : ð2:27Þ
P ð x Þ
2
i xi
2 i i
n
The example of Kissinger’s equation in Eq. (2.26) means plotting ln(β/T2max)

versus 1/T2max gives the slope, m, of E/R. The same linear regression, of course, is
the basis of all isoconversional methods.
The first measure of the quality of fit of a linear model is the correlation
coefficient, r:

P P P
ð i xi i yi Þ=n
i xi yi
r ¼
2
P P : ð2:28Þ
P 2 ð i xi Þ2 P 2 ð i yi Þ2
x
i i n y
i i n
For a perfectly linear data set, r2 = 1. Two realizations are shown in Fig. 2.4 for
a Kissinger plot using Arrhenius parameters of 1.0 × 1014 s−1 and
E = 54,000 cal/mol and maximum random error of ±3 °C for each point for a
heating range of 1–16 °C/min.
The frequency factor, A, changes in concert with E according to a compensation
relationship. A more complete study of random errors on A and E is given else-
where [50], and a rather detailed discussion of the respective contributions of
random errors and mechanistic effects to the compensation effect has been pre-
sented recently [51, 52]. Rather than engaging is a detailed discussion of the
compensation effect at this point, I merely mention that values of r2 less than 0.98
indicate that the results should be used with caution outside the range of calibration.
For reliable extrapolation from lab experiments to geological time and temperature
-14 -14
-15 -15
ln(Hr/RT2)
-16 -16
2.48e14 s-1 2.09e13 s-1

-17 -17
55363 cal/mol 51809 cal/mol
r2=0.9875 r2=0.9941
-18 -18
1.34 1.36 1.38 1.4 1.42 1.44 1.46 1.34 1.36 1.38 1.4 1.42 1.44 1.46
1000/T, K-1 1000/T, K-1
Fig. 2.4 Two realizations of a Kissinger plot including measurement noise
conditions, the r2 must be greater than 0.997. The geological extrapolation issue is
discussed in greater detail in Sect. 7.1. In addition, it is best to plot the residuals
(differences between measured and calculated) versus the abscissa variable to
inspect for non-random trends that would indicate the need for a higher order
model.
Any model that can be transformed into a linear equation can be analyzed by
linear regression. The most common example is a first-order reaction, for which the
logarithms of the reaction rate and fraction remaining are linear with time. Such a
transformation, however, introduces a problem that depends on the nature of the
detection.
In most measurements of rates and fractions reacted (or remaining), there are
both random errors in the measurement and an uncertainty in the baseline. Random
errors can have components proportional to the maximum instrument sensitivity
and the measured signal. Model fitting by linear transformation and regression
magnifies error associated with baseline errors and noise that is proportional to
maximum instrument sensitivity. Consequently, it effectively raises the relative
importance (weighting) of the least reliable data. Taking the example of a first-order
reaction with random errors of 2 % of maximum plus 2 % of the value plus a 2 %
error in the baseline, the error in the rate constant from linear regression is 2.5 %,
while the error in the nonlinear fit is 1.4 %. More details on linearization methods
for optimizing kinetic models are given elsewhere [6].
Nonlinear regression [53] also involves minimizing some function of the dif-
ference between calculated and measured values, ordinarily the residual sum of
squares (RSS):
X
RSS ¼ ðy
i exp;i
ycalc;i Þ2 : ð2:29Þ
Unlike the linear regression case, there is no analytical solution. Instead, the
minimum value must be found by some type of search algorithm. One way is to
merely map out a surface with variations in the variables and look for the minimum
by a grid search. Another method is to look at the derivative of the RSS with respect
to changes in the different variables and move continuously toward the lowest
value. One popular method of doing the latter is the Levenberg-Marquardt
Algorithm [54], which is also known as damped least squares, alternating between
gradient decent and Gauss-Newton methods.
Kinetics2015 and predecessors dating back to 1987 use the Levenberg-
Marquardt algorithm for optimization. In one case, the discrete energy distribution
model, optimization uses iterative linear regression and constrained non-linear
regression optimization based on the Levenberg-Marquardt algorithm. Lakshmaman
et al. [55] use a grid-search method, and Hillier et al. [56] use a simulated annealing
method that has aspects similar to the grid-search method.
All nonlinear regression methods have the danger of falling into a local mini-
mum rather than the global minimum. One method to check for such a problem is to
start the minimization from a different set of initial conditions. Another method is to
restart the minimization a few times to be sure that it cannot find a lower minimum.
The latter is used by Kinetics2015, and the improvement is usually minor. More
important is to start the non-linear regression with good initial guesses. The phi-
losophy adopted by Kinetics2015 is to provide those initial guesses from simpler
analyses such as isoconversional analysis. How that is done is explained in the
following sections.
Another issue is the ability to fit multiple reactions to reaction data. This is
relatively straightforward, for example, if for a constant heating rate, two fairly well
resolved peaks are observed. Even so, sometimes it is useful to do successive
optimizations with subsets of the parameters prior to a final optimization with
parameters that are close to the final values. More serious is when some system of
parallel, sequential, and competing reactions does not give distinctive features to
insure a unique fit. The Netzsch Thermokinetics [57] software has the capability to
construct such complex models. If some reaction parameters can be fixed reliably
by other considerations, the probability of achieving a defendable model improves.
A final consideration is that many people want a universally applicable measure
of goodness of fit like r2. It does not exist for non-linear regression. While lin-
earization methods are sometimes used to estimate parameter uncertainties in
nonlinear regression, they may be grossly in error unless the model is close to linear
[58]. When running a set of similar optimizations, one can compare the normalized
residual sum of squares for a comparison within that set. Another indicator would
be to change one parameter with the others fixed to cause the residual sum of
squares to change by some amount deemed to be within the experimental repro-
ducibility of the minimum, then re-optimize the other variables with that variable
fixed [59]. Roduit et al. [60] present a statistical test for whether adding another
parameter is justifiable in terms of the improvement in the residual sum of squares.
2.5 Sigmoidal Reaction Models
As mentioned in the previous section, the isothermal reaction rate of sigmoidal

reactions is initially small, gradually accelerates with time as some intermediate
property controlling the reaction rate increases, then gradually decreases to zero as
the reactants are consumed. In integrated form it looks like a horizontally stretched
“S”, with a maximum slope at some intermediate conversion. Because there is
always a finite heat-up time for a pyrolysis reaction, the reaction rate and conver-
sion have some sigmoidal characteristics from the acceleration of the reaction with
temperature, and a careless analyst may dismiss this acceleratory phase as sample
heat-up, but it is truly a fundamental characteristic of many chemical reactions.
Sigmoidal reactions were introduced in the previous section, with the extended
Prout-Tompkins (ePT) and Avrami-Erofe’ev equations as two examples. In fact, the
variations are far more extensive. The first appearances of these equations were
actually prior to the papers from which they are commonly named.
An introduction to the various nucleation-growth models in the solid state lit-
erature is given by Brown et al. [61]. The simplest nucleation-growth model does
not recognize the overlapping consumption of reactants and leads to the power law
in Table 2.1. Avrami [62], Johnson and Mehl [63], and Kolmogorov [64] and
Erofe’ev [65] then independently considered the merging of reaction centers,
obtaining
a ¼ 1 expðktp Þ: ð2:30Þ
This equation is known variously as the Avrami equation, the JMA equation, the
JMAK, and the JMAEK equation after its originators. It assumes that the reaction
nucleates (initiates) randomly and homogeneously over the entire unconverted
volume and grows uniformly in all directions. Solving for kt, it is evident that this
equation is the g(α) given in Table 2.1. The value of p was once considered to be
the sum of the number of steps in the nucleation process and number of growth
dimensions, but it is now generally treated only as an empirical constant.
Prout and Tompkins [66] used a different approach to model the thermal
decomposition of potassium permanganate. They found that the decomposition
followed the logistic equation
h a i
ln ¼ kt þ c; ð2:31Þ
1a
which corresponds to the rate form
da=dt ¼ kð1 aÞa: ð2:32Þ
The Prout-Tompkins model is more closely related to autocatalytic reactions

than nucleation-growth reactions with overlapping volumes [67].
2.5 Sigmoidal Reaction Models 43
Šesták and Berggren [68] proposed an empirical equation that combines aspects
of the Prout-Tompkins and JMAEK equations:
da=dt ¼ kam ð1 aÞn ½lnð1 aÞs : ð2:33Þ
This equation has all the limits of the nth-order, ePT, and Avrami-Erofe’ev
(JMAEK) models, although s in this equation equals 1 − 1/p in the Avrami
equations in Table 2.1.
What my coworkers and I call the ePT model (s = 0 in Eq. 2.33) has an inter-
esting history. Although the name for this class of models comes from the 1944
paper of Prout and Tompkins on decomposition of KMnO4, it actually appears
earlier in the literature. Austin and Rickett [69] considered both the normal auto-
catalytic equation (Eq. 2.31) and the logarithmic analog
ln½a=ð1 aÞ ¼ klnt þ c; ð2:34Þ
for analyzing decomposition kinetics of austenite, an allotrope of iron or steel. The

latter worked better for them, and Prout and Tompkins [70] also found in a later
paper that the logarithmic form worked better for AgMnO4. Avrami refers to the
Austin and Rickett paper part I of his series [62] as a limit of his equation, and
Erofe’ev [71] showed by series expansion and grouping that ePT model is equiv-
alent with the Avrami-Erofe’ev (JMAEK) model with certain ordered pairs of n and
m. A more extensive numerical comparison appeared recently [72]. Finally, the ePT
model (n and m ≠ 1) actually appears in a 1940 Russian publication [73] that only
recently has been recognized in the English literature [60].
More recently, a generalized form of the logistic equation [74, 75] was intro-
duced to chemical kinetics analysis:
1 a ¼ 1=ð1 þ sebðtcÞ Þ1=s ð2:35Þ
where b is a reaction rate parameter, t is either time or temperature, and c is the

midpoint of the reaction in either time or temperature for a constant heating rate.
(Note in contrast that c in Eq. 2.31 is dimensionless.) When differentiated this
equation gives
da=dt ¼ bebðtcÞ =ð1 þ sebðtcÞ Þs þ 1

1
ð2:36Þ
Comparing Eqs. (2.35) and (2.36) to Eqs. (2.31) and (2.32), respectively, indi-
cates that if t is time, c is an arbitrary shift in time, b is the negative of the
conventional rate constant, and τ is a new shape parameter. Numerical studies [76]
show that the parameter τ has basically the same effect as reaction order in the ePT
model, with τ = 0.1 comparable to n = 1.7 and τ = 3.0 comparable to n = 0.65.
Although the implementation of the generalized logistic equation by that group is
not optimal and is less flexible than the ePT model, it is a potentially useful method
if properly implemented.
An issue with some simple forms of sigmoidal reaction models is that the initial
reaction rate is zero and the reaction can never begin. This can be rectified in several
ways, but it is very instructive to first explore one more complete version of an
autocatalytic model commonly used for energetic materials [10]. Consider the
reaction sequence
A ! B; B þ A ! 2B
which can be described mathematically by the relation dA/dt = –k1A − k2AB. The
differential equation cast in terms of fraction reacted in the thermal analysis style is
da=dt ¼ k1 ð1 aÞn1 þ k2 am ð1 aÞn2 : ð2:37Þ
The empirical reaction orders are added for flexibility. If we assume that n1 = n2,
then
da=dt ¼ k2 ð1 aÞn ðam þ zÞ k2 ð1 aÞn ½1 qð1 aÞm ð2:38Þ
where z = k1/k2 ≈ 1 − q. The left-hand approximation is used in CISP software

[77], and the right-hand approximation is used in Kinetics2015. The z and q factors
are basically ways of starting numerical integration with a non-zero reaction rate in
the absence of an explicit initiation reaction, and they can be optimized to fit any
data set. While no analytical solution is available for these equations, a solution is
available for n and m equal to 1 (basic Prout-Tompkins limit):
a ¼ 1=ðekt ð1 a0 Þ=a0 þ 1Þ ð2:39Þ
where α0 is the integration constant that provides an initial fraction reacted to start
the reaction.
Brown et al. [78] observe that it is difficult to distinguish whether the ePT or
JMAEK approaches fit any given data set better. That has been my experience, also,
so I have used the ePT approach almost exclusively. A qualitative comparison [79]
of the two approaches is reproduced in Fig. 2.5. The two models could be matched
even closer by optimizing the factor q if desired.
A further comparison of the logarithmic (Eq. 2.34) and ePT approaches is also
instructive. (Recall that Prout and Tompkins found that AgMnO4 was sigmoidal in
logarithmic time while KMnO4 was sigmoidal in linear time.) A conversion versus
time curve was generated using Eq. (2.34) solved for α and then fitted by nonlinear
regression to the right-hand approximation of Eq. (2.34). The results are shown in
Fig. 2.6. There is a hint of an acceleratory phase when plotted on a linear scale, but
the sigmoidal character is clearly evident on a logarithmic scale. The agreement is
as good as needed for any practical purpose.
In discussions with researchers doing pyrolysis of cellulose and kerogens, some
resisted the adoption of a sigmoidal reaction model despite clear evidence that it
was necessary. One doubted that models developed for inorganic materials were
Fig. 2.5 Comparison of extended Prout-Tompkins and Avrami-Erofe’ev approaches to modeling

acceleratory reaction rates from Burnham et al. [79]. The ePT calculations used q = 0.99, and the
frequency factor had to be renormalized by 1/(1 − m) for Tmax to be independent of m. The symbol
m′ was used in the Avrami equation in that paper instead of the symbol p in Table 2.1. Reproduced
courtesy of LLNL
1.0 1.0
0.8 0.8
FracƟon reacted
0.6 0.6
0.4 0.4
0.2 m=0.34, n=1.67, q=0.999 0.2

m=0.34, n=1.67, q=0.999
logarithmic PT logarithmic PT
0.0 0.0
0 100 200 300 1 10 100 1,000
Time, arbitrary units Time, arbitrary units
Fig. 2.6 Demonstration of the ability of the ePT model to fit logarithmic sigmoidal reaction
characteristics [76]. © Springer
relevant to organic pyrolysis. Another thought it was fine to allow the frequency
factor to depend on heating rate and thereby increase the activation energy to
narrow the reaction profile enough to match experiment. Both positions are without
merit. The following paragraphs give published examples to support that statement.
It has been known for a 100 years or so that organic pyrolysis reactions occur by
chain reactions. By the early 1930s, detailed mechanisms [80] had been worked out
for initiation, propagation, and termination of hydrocarbon cracking, including how
to calculate the overall activation energy from the activation energies of the various
steps.
The essence of a chain reaction is that there is an active center that initiates the
chain, a chain propagation step that generates some number of new active centers,
and a quenching reaction that ends the chain. Although the chemical and physical
processes of each step may differ, the mathematics is the same unless one develops
a detailed mechanistic model, which is beyond the objectives of this book.
The most directly relevant analog to kerogen pyrolysis comes from synthetic
polymer pyrolysis, which was explored in great detail about 50 years ago. Flynn
and Florin give [81] a good review of that work, including a lengthy report from the
National Bureau of Standards and multiple books. They summarize the main
mechanisms as random main-chain scission, depolymerization, side-group reac-
tions, and carbonization.
Random chain scission is the breaking of the main chain to form smaller
molecules of all sizes. The molecular weight falls, eventually forming molecules
small enough to be volatile. There is no preference for monomer formation. For
example, polyethylene and polypropylene decomposition do not preferentially form
ethylene and propylene. Depolymerization is the unzipping of the polymer after an
occasional break. Monomer formation is primary. Tetrafluoroethylene, α-methyl
styrene, and methyl methacrylate are examples. Side-group reactions involve the
elimination of a labile group off the main chain. Examples include elimination of
hydrogen chloride from polyvinyl chloride to leave a polyene structure and elim-
ination of isobutylene from poly(t-butyl methacrylate) to leave polyacrylic acid.
Carbonization is the formation of a carbonaceous residue and involves a variety
of processes, usually in parallel with random scission. It may include cross-linking,
polyene formation by elimination of side chains, cyclization, and aromatization of
cyclics. In the coal literature, “retrograde reactions” is sometimes used as a
descriptor. Lyon et al. [82] deduced group formation contributions to char forma-
tion and found that poly-aromatic-heterocyclic rings were the strongest char former,
while simple aromatic groups were far less char-prone—e.g., polystyrene forms
little char. Formation of carbon fibers from polyacrylonitrile involves cyclization of
the nitrile side chains and aromatization, enhanced by mild oxidation.
Flynn and Florin [81] review depolymerization theory and present theoretical
results for various situations. Depending on the initiation location and unzipping
length in the absence of chain transfer, the reaction rate versus conversion can be
either deceleratory or autocatalytic, as shown in Fig. 2.7, which is adapted from
Wall [83]. Note that the end initiation with small unzipping length leads to a flat
reaction curve, because each end is generates volatiles at a constant rate and the
Random iniƟaƟon End iniƟaƟon

large zip large zip
RelaƟve reacƟon rate
small zip small zip
Conversion, %
Fig. 2.7 Theoretical calculations of the depolymerization reaction rate versus conversion for
various assumptions of initiation location and unzipping lengths (zip). Adapted from Wall [83]
with permission. © 1962 Wiley Interscience
number of ends remains constant until most of the material is consumed. In con-
trast, random initiation with a small unzipping length results in an acceleratory
phase, as the number of ends generating volatiles increases with time during the first
phase of pyrolysis. For very large unzipping lengths, the overall rate decreases with
time, because a single initiation consumes the entire chain and the number of chains
decreases with time.
Flynn and Florin [81] then review random scission theory and, as noted previ-
ously by Burnham et al. [10], summarize equations that have similar functional
form to the Prout-Tompkins model. As shown by Flynn and Florin for polyethylene
and in the original work by Wall and Straus [84] for a wider range of polyolefins
(see Fig. 2.8 for the polyethylene example), the rate of devolatilization is always
autocatalytic in shape (sigmoidal in fraction reacted) for linear polyolefins but
strictly deceleratory for branched polyolefins. This is because the probability of
Fig. 2.8 Effect of chain

branching on volatilization
rate versus conversion for
polyethylene pyrolysis at
constant temperature, adapted
from Wall and Straus [84].
Reprinted with permission. ©
1960 Wiley
generating a volatile fragment for a given scission is proportional to the amount of

material present for a branched material, while random scission of linear material
generates primarily shorter, non-volatile chains initially. As the chains get shorter,
the probability of a volatile product from a given scission increases with time until
the amount of material has decreased enough that the rate becomes deceleratory.
After success fitting a variety of kerogen samples with first-order and distributed
activation energy models, my first measurements of polymer pyrolysis provided a
rude awakening—the reaction profiles of several of the polymers provided by a
colleague, Richard Lyon, were narrower than could be calculated by a first-order
model using an activation energy derived by Kissinger’s method [85]. Using a
reaction order less than one matched the profile width but didn’t get the shape right.
We noticed the same character for one kerogen sample, and we correlated the
characteristic of a narrow reaction profile with an acceleratory phase during
isothermal pyrolysis [86].
It was another 6 years before these characteristics were fully connected to the
polymer and solid state kinetic model literature [79]. During that time, we found
that a sequential reaction gave fits comparable to the nucleation-growth approach,
but the non-linear regression was not as robust. Consequently, from the practical
perspective of this book, there is little reason to explore models other than the ePT
model. A few examples of its utility for polymers follow.
Prior publications [4, 87, 88] showed that polyethylene, poly(1,4-dimethylene
naphthalene), polystyrene, polysulfone, polycarbonate, polyether-ether-ketone
(PEEK), Kel-F 800, and Viton-A have essentially single decomposition reactions
that are narrower than predicted from first-order kinetic parameters derived from
Kissinger’s equation. The ratio of measured to calculated profile widths ranged
from 0.36 to 0.86, with an average of 0.62. At the same time, the peak shapes were
different, with asymmetry factors ranging from 0.6 (slightly less asymmetric than a
first-order reaction) to 1.1 (skewed to high temperature significantly more than a
first-order reaction. In addition, two polymers with two clear decomposition steps,
Estane and poly(vinyl acetate), had one of the two steps being narrower than a
first-order reaction. The resulting ePT model parameters for these polymers ranged
from 0.23 to 0.90 for m, with an average value of 0.62, and from 1.0 to 2.2 for n,
with an average value of 1.3. Comparing these averages to the parameters in
Fig. 2.6, linear polymers are generally between the linear and logarithmic forms of
the Prout-Tompkins model.
A recurring question in the literature is whether kinetics derived from constant
heating rate experiments agree with those derived from isothermal experiments. If
the kinetic model is correct, they should and can agree. Such agreement is shown in
Fig. 2.9 for the decomposition of PEEK [89]. In this case, the kinetic model was
optimized simultaneously on the heating rate and isothermal data. For three other
cases [88], the ePT kinetic model was calibrated on heating rate data and then
compared with isothermal data. The agreement was not as close but was qualita-
tively correct. Kel-F showed the poorest agreement of the three polymers, as shown
in the left side of Fig. 2.10. A refit of the Kel-F data to both constant heating rate
and temperature data is shown in the right side of Fig. 2.10. The agreement with the
Fig. 2.9 Fit of the ePT model 1.0

0.67
to PEEK decomposition data 0.96
from Nam and Sefaris [89]. 1.92
0.8
The indicated heating rates 4.83
FracƟon reacted
are in °C/min and the 19.5
temperatures are in °C. The 0.6
kinetic parameters are
A = 2.4 × 1014 s−1, 0.4
E/R = 28,940 K, m = 0.86,
and n = 0.89. Reproduced 0.2
courtesy of LLNL
0.0
500 550 600
o
Temperature, C
500
490
480
0.8
470
FracƟon reacted
460
0.6
0.4
0.2
0.0
0 200 400 600 800 1000 1200 1400
Time, min
isothermal data is better, as would be expected, but the fit to the constant

heating-rate data is not as good. E/R changes from 15,450 to 13,217 K. The
isothermal data was taken 2 years after the constant heating rate data, so it is an
open question whether the discrepancy is an experimental or model problem. This
example shows the importance of fitting a diverse set of thermal histories to be sure
one has the correct model and model parameters. We return to this issue later in
Sect. 2.7 on how to select a model to fit.
Cellulose also has a narrow reaction profile and acceleratory phase, and con-
sequently requires a sigmoidal reaction model [2, 90, 91]. The ePT growth
parameter m was between 0.4 and 0.5 for three different samples, and activation
energies were in the mid 40 kcal/mol range. These results are qualitatively con-
sistent with earlier work by Dollimore and Holt [92] that showed an acceleratory
phase on a reduced time plot, from which they concluded that a nucleation-growth
model was most appropriate and that the data was consistent with an Avrami
exponent p of 2. That corresponds to m ≈ 0.75 according to an extrapolation of
Fig. 2.5.
1.0 1.0
2.5 °C/min 400 °C 2.5 °C/min
400 °C
0.8 0.8
FracƟon reacted
FracƟon reacted
5 °C/min 5 °C/min
0.6 0.6
10 °C/min 10 °C/min
0.4 0.4
0.2 0.2
0.0 0.0
0 40 80 120 160 0 40 80 120 160
Time, min Time, min
Fig. 2.10 Fit of the ePT model to Kel-F decomposition data from Burnham and Weese [88].
Heating rates were 2.5, 5, 10, 20 and 40 °C/min. The left figure shows a fit to the heating-rate data
only with a comparison to the 400 °C isothermal data. The right figure shows a fit to all data
During this time period, a leader in the biomass community recommended

against the publication of Reynolds and Burnham [90] because he believed cellu-
lose pyrolysis was first order. He subsequently led a round robin exercise that fitted
a first-order model independently to pyrolysis data at two different heating rates
[93], in direct conflict with the conclusions in Reynolds and Burnham [90] and the
best practices outlined in the two subsequent ICTAC papers on kinetic analysis [6,
38]. The eight independent measurements of Tmax from different labs give a fairly
reliable estimate of the average Tmax value, and A and E calculated from Kissinger’s
method are A = 1.75 × 1015 s−1 and 48.04 kcal/mol, in excellent agreement with
Capart et al. [91]. In contrast, the average activation energies from single heating
rate fits were 58.32 and 53.06 kcal/mol, respectively, at 5 and 40 °C/min, which do
not make chemical sense in terms of the bond strengths.
All of the first-order Arrhenius parameter pairs derived in the round robin are
plotted in a compensation-law plot in Fig. 2.11 [94]. They fall on a line with the
parameters from Kissinger’s method. Calculating the average profile widths from
the round robin and Kissinger method kinetics, one finds that the measured
pyrolysis profile is only 84 % of the first-order width. That corresponds to a value
of m of about 0.3, which is qualitatively consistent with Capart et al. [91].
Fortunately, others in the cellulose pyrolysis field eventually understood the need
for a sigmoidal reaction model. Kim and Eom [95] used TGA and reduced-time plots
to choose a reaction model and determine an Avrami exponent of 3.7, which is
highly acceleratory. Barud et al. [96] used a nonlinear isoconversional approach to
confirm that cellulose pyrolysis followed the Avrami-Erofe’ev model with an
exponent p of 1.63 and an activation energy of about 44 kcal/mol. The Avrami
exponent corresponds to an ePT value for m of 0.45. These parameters are very close
to those of Reynolds-Burnham. Similarly, Sanchez-Jimenez et al. [97] used a master
plot approach discussed further in a following section to show that the appropriate
Fig. 2.11 Arrhenius 40 °C/min

parameters determined in a
1x10 21 avg 40 °C/min
round robin study of cellulose
5 °C/min
pyrolysis [93]
avg 5 °C/min
1x10 19 Kissinger method
A, s-1
1x10 17
1x10 15
46 50 54 58 62
E, kcal/mol
model was either a random-scission kinetic model mathematically similar to the ePT
model or an Avrami-Erofe’ev model with p < 1.5. Figure 2.5 puts their equivalent
m value for the ePT model at *0.4, and their activation energy of 46 kcal/mol is
between those of Capart et al. and Reynolds-Burnham. Ultimately, comparison of
global kinetic modeling to mechanistic modeling showed that cellulose pyrolysis is
not first order but is consistent with a random-scission or nucleation-growth model
[94]. Even so, there are still questions about how much of the sigmoidal character is
due to the chemical reaction mechanism and how much is due to delayed vapor-
ization of pyrolysis products, but that does not excuse the practice of fitting a single
heating-rate experiment to a first-order reaction model.
The lesson here is that leaders in a field are sometimes slow to believe and adopt
the obvious need to move on from commonly used historical pyrolysis models. We
will also see that in the kerogen pyrolysis literature.
2.6 Distributed Reactivity Models
The concept that a complex reaction proceeds by a distribution of independent

parallel reactions is very old, dating back at least to 1925 to the work of Constable,
who noticed a compensation effect between A and E for surface catalyzed reactions
and proposed that the surface had a variety of sites that led to a distribution of
activation energies governing the reaction [98]. In 1943, Vand [41] developed a
reactivity distribution model for irreversible electrical resistance changes upon
annealing of evaporated metal films. The parallel reaction approach got greater
exposure from the work of Pitt [99] and Hanbaba [100] in the coal pyrolysis
community in the 1960s. Probably the next influential step was the Gaussian energy
distribution model popularized by Anthony and Howard [13] for coal pyrolysis,
which continues to be common today despite its shortcomings. Finally, workers at

the Institut Français du Pétrole developed the discrete energy distribution model for
petroleum generation, and the first version having reasonable kinetic properties was
published by Ungerer and Pelet [101]. A concurrent approach to reactivity distri-
butions originates from the isoconversional methods of Friedman [29] and Ozawa
[30], which were originally designed for determining a single activation energy and
reaction order but have been used more often to determine how the activation
energy varies with conversion. A variety of other historical developments are
described by Burnham and Braun [4].
The distributed reactivity models all have their basis in Eq. (2.7). To generalize
to an arbitrary number of parallel reaction channels, we first define xi as the fraction
remaining of the ith reaction potential, giving
X X X
1 ¼ x0 ¼ ax ;
i i i0
a ¼ 1; xðtÞ ¼
i i
a x ðtÞ
i i i
ð2:40Þ
Differentiating the right-hand equation yields
dx X dxi X
¼ a
i i dt
¼ a k ðTÞf ð1 xi Þ;
i i i
ð2:41Þ
dt
which is an explicit summation of Eq. (2.7), where the f(1 − xi) are given in
Table 2.1. Obviously, Eq. (2.41) can be used in this discrete summation for any
reaction for which f(1 − xi) is known. Now consider replacing the discrete sum with
a continuous mathematical distribution, but restrict the derivation to a first-order
reaction, for which the integration over any thermal history is given in Eq. (2.9). If
the distribution in reactivity is given by a distribution in activation energy only,
Z 1 Z t
x¼ exp kðTðtÞÞdt DðEÞdE ð2:42Þ
0 0
and
Z 1
DðEÞdE ¼ 1: ð2:43Þ
0
Two commonly used distributions are the Gaussian distribution:

h i
DðEÞ ¼ ð2pÞ1=2 r1 exp ðE E0 Þ2 =2r2 ð2:44Þ
where σ is the normal distribution parameter, and the Weibull distribution:

" # 8 " # 9
2 b1 < 2 b=
b ð E cÞ ð E cÞ
DðEÞ ¼ exp ð2:45Þ
g g : g ;
2.6 Distributed Reactivity Models 53
where η is a width parameter, β is a shape parameter (not the heating rate in this
section), and γ is the activation energy threshold. The mean activation energy is
given by

1
E0 ¼ c þ gC þ1 ð2:46Þ
b
where Γ() is the Gamma function. The Weibull distribution is much more flexible,
being able to change in both shape and width, but only a few workers (as described
by Burnham and Braun [4]) have explored its use.
Burnham and Braun [4] provide a variety of comparisons between various
activation energy distribution models. As the Weibull distribution parameter η
increases, so do the values of σ and n when simulated Weibull data are fitted to a
Gaussian distribution or nth-order reaction model. A Weibull energy distribution
having η = 3 and β = 1 corresponds very closely to an nth-order reaction with
n = 2.7. Any number of such comparisons could be constructed.
It is important to understand that a distribution in activation energy is not the
only way to introduce a reactivity distribution. From a different perspective,
Kolar-Anić and Veljković [102] explored the use of Gamma and Weibull distri-
butions of reactivity to understand the Avrami exponent p in a variety of systems.
Ho and Aris [103] show that a Gamma distribution of rate constants for isothermal
conditions leads to apparent nth-order kinetics, with second order corresponding to
the limit of an exponential distribution of reactivity, which at isothermal conditions
is equivalent to a distribution of pre-exponential factors. Burnham and Braun [4]
use that concept to justify nth-order kinetics in combination with a Gaussian dis-
tribution of activation energies to gain the shape flexibility of a Weibull distribu-
tion, at least to the extent that it is needed to model pyrolysis reactions, and that
approach was quite useful for modeling the process of sintering [104].
It is also important to understand that a distribution in reactivity is not the same
as using a logistical distribution function to fit reaction profile shapes, as was done
by Cai and Liu [105]. While the latter might be a useful method for preprocessing
reaction profiles to eliminate noise, it does not result in fundamental kinetic
parameters that can be used for modeling conditions other than those fitted.
Some have looked for analytical equations to simplify the use of continuous
reactivity distributions for specific situations, e.g., a Gaussian energy model for a
constant heating rate [106], but we have found it more practical to break any
continuous distribution of activation energy into a discrete distribution of evenly
spaced energies. The reason is twofold. First, we find the numerical integration of
independent parallel first- and nth-order reactions is quite fast compared to having
to evaluate complex exponential functions. Second, we prefer to write software for
arbitrary thermal histories, because experimental data and industrial applications
rarely follow the ideal thermal histories people usually analyze.
One must consider how the energy spacing between reaction channels affects the
validity of the model. Our original work [23] used 0.25 kcal/mol spacing, but
subsequent numerical studies found that 0.5 kcal/mol spacing was adequate. For
systems that are characteristically broad in terms of reactivity distribution, using

nth-order reactions with n = 1.5 (approximately symmetrical for a constant heating
rate) or 2.0 (skewed to high temperature) in the distribution allows one to cut the
number of reaction channels by 1.5–2 times without compromising accuracy,
thereby speeding computational time. Furthermore, an nth-order model with
n = 1.001 is insignificantly different from a true first-order model, so programming
only one type of reaction would be required for each reaction channel. If not needed
to adjust the reaction peak shape, one could have the reaction order change in
concert with the energy distribution width in some predetermined manner to
minimize the additional channels needed for a wide reactivity distribution.
In a complex material such as a fossil fuel, different products are evolved at
different stages of maturity, so it is likely that a simple mathematical distribution
will not have flexibility to fit the reaction profile of some reactions. Discrete acti-
vation energy distributions were developed to provide that flexibility. Initially, the
model had a single pre-exponential factor and weighted first-order reaction channels
placed at discrete intervals. The first implementation by the IFP [101] used 20
channels spaced at 2 kcal/mol from 40 to 80 kcal/mol and optimized the
pre-exponential factor and weighting factors by nonlinear regression. We subse-
quently developed [106] a much faster, iterative constrained linear/nonlinear
regression method for the same optimization, including algorithms to select the
correct number of reaction channels constrained by the data and a user-selected
channel spacing, which we recommended to be 1 kcal/mol for most materials. This
general approach is the most commonly used method in the petroleum exploration
geochemistry community.
The additional flexibility of the discrete model compared to a single distribution
model enables it to easily fit the reaction profiles of coal and other materials that
have multiple overlapping chemical reactions. Neither the Weibull nor the nth-order
Gaussian models were able to simultaneously fit the relative sharp profile from 10
to 80 % conversion and the broader reaction profiles for the initial 10 % and final
20 % of the reaction, as shown in Fig. 2.12. Although fitting three parallel Weibull
or Gaussian-nth models could do considerably better than a single such model, the
discrete model fits the profiles with ease.
An issue that has been explored at some length is the proper energy spacing
between reaction channels. Recall that we found that 0.5 kcal/mol spacing was
adequate to discretize an analytical distribution with negligible error. Increasing the
reaction channel spacing to 1 kcal/mol still introduces negligible error compared to
the limitations of an analytical form. The discrete model optimization is a little
different, in that too wide of a spacing causes a modulation of the minimization
function corresponding to shifting the distribution by the channel spacing and
compensating with a change in the optimum pre-exponential factor. This is shown
in Fig. 2.13, which is reproduced from Sundararaman et al. [107]. Clearly, when
most of the reaction occurs in one reaction channel, the energy spacing must be less
than 1 kcal/mol to be sure of a good optimization, and Kinetics2015 allows them to
be 0.5 kcal/mol. Second, when the reactivity spans an activation energy range of
several kcal/mol, 1 kcal/mol spacing is needed.
1.0 1.0
Gaussian-nth model Discrete model
0.8 0.8
Fraction reacted
0.6 0.6
0.4 0.4
0.2 0.2
0.0 0.0
300 400 500 600 700 800 300 400 500 600 700 800
Temperature, °C Temperature, °C
Fig. 2.12 Comparison of the ability of the nth-order Gaussian distribution and discrete
distribution models to fit coal pyrolysis data at 5, 10, and 20 °C/min. A Weibull distribution fit
was comparable to the Gaussian-nth fit
Fig. 2.13 Graphics from Sundararaman et al. [107] showing the effect of too-widely-spaced
energy channels on the nonlinear regression minimization surface and the probability of finding a
false minimum. RMS = Root-mean-squared deviation. Reprinted with permission. © 1977
American Chemical Society
Kinetics2015 allows the user to pick a larger spacing, and the program improves
the probability of finding the right minimum even if the user chooses too large of
spacing by using the characteristics of the compensation law. The minima are
spaced by the relationship that a 1 kcal/mol shift in the activation energy distri-
bution is compensated by a factor-of-two change in the pre-exponential factor.
Once a minimum is found, Kinetics2015 searches on both sides of the minimum by
that spacing to assure that the correct minimum is found.
A second important issue is whether it is appropriate to assume the same
pre-exponential factor for all energy channels. Such an assumption simplifies
programming of industrial applications, but it is certainly not necessary. It is pos-

sible, and Kinetics2015 allows the pre-exponential factor to be a linear function of
the activation energy, which is consistent with the concept of a compensation law
that is a function of the chemical reactions and not merely a reflection of experi-
mental error.
The effect of such a compensation relationship is to spread the activation
energies over a wider range, which may or may not be chemically realistic [4]. As a
result, an extrapolation outside the temperature range of calibration will spread the
reaction over a wider temperature range. Such an example is given in Fig. 2.14,
which is calculated from kinetic parameters in Table 9 of Burnham and Braun [4].
Whether or not the variable A kinetics are more correct, the difference in predictions
is not too large, and it is doubtful that it would make a difference in any exploration
drilling decision given all the other uncertainties in paleo thermal history recon-
struction. Consequently, there has been no compelling reason for those involved in
oil and gas exploration to incorporate the complication of a variable A into their
petroleum generation calculations, and it is rarely, if ever, done.
Isoconversional methods were originally envisioned to determine a best single
activation energy, but they are currently used more to determine the dependence of
activation energy on conversion. The Friedman model does that without approxi-
mation, but all integral isoconversional methods except the advanced integral model
of Vyazovkin [37] assume that the activation energy is constant up to the point of
analysis, which is self-contradictory, and Vyazovkin’s method obtains values
essentially the same as Friedman’s method.
There are two philosophies on how to deal with a distribution of activation
energies from isoconversional methods. One is to consider that the activation
energy at the ith conversion point is simply the average activation energy at that
conversion. However, Miura [39, 40] argues that reaction profiles are not delta
functions, so any given level conversion involves overlapping of reactions with
different activation energies. He then derives a relationship for the effective
boundary between energies that have and have not significantly contributed to the
conversion to that point. Next, he replaces the equation
Fig. 2.14 Predictions of oil 1.0

and gas generation from a PredicƟons for 2 C/Myr
Ventura Basin Monterey shale 0.8 Variable A
sample using discrete
FracƟon reacted
Single A
activation energy models with 0.6
a single and variable
pre-exponential factor 0.4
(Myr = million years)
0.2
0.0
50 100 150 200 250
Temperature, C

ART 2 E
x ¼ exp e RT ð2:47Þ
bE
(which drops all the series-expansion terms in Eq. 2.14) by the equation

0:545ART 2 Es
x ¼ exp e RT : ð2:48Þ
bEs
He then calculates A and E pairs by an isoconversional method so that the

measured and calculated reaction curves agree.
Cai et al. [108] criticize Miura’s method by saying that it does not recover a
constant frequency factor when simulated data from Gaussian distribution models
using a constant frequency factor are fitted to Miura’s method. In fact, that char-
acteristic was pointed out far earlier by Burnham and Braun [4], and Friedman’s
method has a similar characteristic. However, one should realize that the derived
parameters compensate for each other in a way mandated by the mathematics, and
the real issue is how well they work for thermal histories besides those used for
calibration. In that respect, both the Friedman and Miura methods work adequately,
as shown by Burnham and Braun. Which method to use is a matter of personal
taste.
Miura [40] presents A-E distributions determined in this way for the Argonne
premium coals. Burnham and Braun [4] found good correspondence between those
distributions those determined by the extended discrete model in Kinetics2015.
Whether it provides a practical advantage over Friedman’s method or merely
provides an alternative way to define the effective activation energy at any extent of
conversion is another matter.
Some insights on these issues are provided by the following example. Simulated
data at 1.0 and 10 °C/min were created using a parallel reaction model having the
properties in Table 2.2. The simulated data were fitted to two different discrete
parallel reaction models: one having a single A factor and one where A depends on
activation energy. The results of those fits are also given in Table 2.2. Despite, the
fewer number of parameters in the extended discrete model (ln(A) = a + bE), it fits
the data significantly better, as shown in Fig. 2.15. The reaction rates rather than
fraction reacted are shown because they are more sensitive to the quality of the fit.
Both fits recover the properties principal reaction channel quite well.
Isoconversional analysis by Friedman’s method yields the parameters shown in
Fig. 2.16. They are closely related to but not identical to the input model param-
eters, in large part because of the fundamental differences in the nature of the
models themselves. Friedman’s method recovers the parameters for the effective
mix of reactions at any conversion level, while the discrete model deconvolves that
overlap. If one wants to compare activation energy distributions from isoconver-
sional analysis to a parallel reaction model, one must do something along the lines
of Miura’s analysis. Good correspondence was shown earlier [4], but it is not
obvious that such a comparison is warranted for ordinary purposes.
Table 2.2 Input and derived parallel reaction models having discrete distributions of activation
energies, one with a single A and one where A depends on E
Input model Extended discrete model fit Ordinary discrete model fit
% A, s−1 E, cal/mol % A, s−1 E, cal/mol % A, s−1 E, cal/mol
1 1 × 1012 40,000 0.01 1.88 × 1018 51,000
2 1 × 1013 44,000 2.06 1.18 × 1013 43,093 0.01 1.88 × 1018 52,000
4 1 × 1014 48,000 0.03 1.88 × 1018 53,000
6 1 × 1015 52,000 7.12 4.87 × 1014 49,943 0.07 1.88 × 1018 54,000
8 1 × 1016 56,000 0.15 1.88 × 1018 55,000
16 1 × 1017 60,000 17.61 2.01 × 1016 56,794 0.37 1.88 × 1018 56,000
26 1 × 1018 64,000 44.45 8.32 × 1017 63,644 0.66 1.88 × 1018 57,000
16 1 × 1019 68,000 20.49 3.44 × 1019 70,495 1.08 1.88 × 1018 58,000
8 1 × 1020 72,000 1.64 1.88 × 1018 59,000
6 1 × 1021 76,000 7.23 1.42 × 1021 77,345 2.62 1.88 × 1018 60,000
4 1 × 1022 80,000 3.74 1.88 × 1018 61,000
2 1 × 1023 84,000 1.04 5.87 × 1022 84,195 5.44 1.88 × 1018 62,000
1 1 × 1024 88,000 7.76 1.88 × 1018 63,000
16.84 1.88 × 1018 64,000
27.44 1.88 × 1018 65,000
14.99 1.88 × 1018 66,000
9.15 1.88 × 1018 67,000
7.04 1.88 × 1018 68,000
0.8 1.88 × 1018 69,000
0.12 1.88 × 1018 70,000
0.04 1.88 × 1018 71,000
1.0 1.0
Normalized reacƟon rate
extended discrete ordinary discrete

0.8 0.8
1 °C/min
0.6 0.6
10 °C/min
0.4 0.4
0.2 0.2
0.0 0.0
260 280 300 320 340 360 380 400 420 440 460 480 260 280 300 320 340 360 380 400 420 440 460 480
Temperature, ° C Temperature,°C
Fig. 2.15 Comparison of the extended and ordinary discrete model fits (lines) to simulated data
(points) at 1 and 10 °C/min. The ordinary discrete model has a single pre-exponential factor, and
the extended discrete model has one that depends on activation energy
90 50 1024
Friedman
1022
80 45
Model
1020
E, kcal/mol
70 40
A, s-1
ln(A)
1018
60 35
1016
Model E
50 30
Friedman E 1014
Friedman ln(A)
40 25 1012
0 0.2 0.4 0.6 0.8 1 40 50 60 70 80 90
FracƟon reacted E, kcal/mol
Fig. 2.16 Comparison of isoconversional kinetic parameters with the input model parameters
used to generate the simulated data set
As stated previously [27], Friedman’s method is fundamentally an infinitely

sequential model in which the reaction sequence must by definition maintain the
same order. Ordinarily, as for fossil fuels, the remaining reactivity in the reactivity
distribution decreases with conversion, but for energetic material, the activation
energy can decrease with conversion, thereby leading to the generation of highly
acceleratory reaction characteristics under some circumstances. Using isoconver-
sional kinetic parameters in an infinitely sequential mode works quite well in either
case. One merely generates an effective first-order pre-exponential factor for each
arbitrarily small fraction reacted interval and performs a sequence of first-order
reaction calculations. A comparison of the isoconversional parameter calculations
with the input model conversions is shown in Fig. 2.17. The agreement is excellent.
The lesson is that there are a variety of well-explored approaches to treat
reactivity distributions, and some work better than others. Single, continuous
activation energy distributions such as Gaussian, nth-order Gaussian, and Weibull
Fig. 2.17 Comparison of 1.0

fractions reacted from the
simulated data set with those
0.8
calculated (lines) from
FracƟon reacted
Friedman isoconversional 1 C/min

parameters derived from the 0.6
simulated data set (points) 10 C/min
0.4
0.2
0.0
260 280 300 320 340 360 380 400 420 440 460 480
Temperature, C
are not sufficiently flexible to achieve an excellent fit to real data from complex
systems. In those cases, either the discrete model, the extended discrete model, or
Friedman’s isoconversional model are far superior, although multiple continuous
distributions sometimes do well. The extended discrete and isoconversional
methods may be advantageous in some circumstances (e.g., energetic materials)
where it is important to account for the variability of the pre-exponential factor with
activation energy, but the difference is ordinarily negligible within the range of
calibration and small even when extrapolated to substantially different
time-temperature regimes. The one situation where continuous distribution models
are superior is for sparse data sets where there is insufficient information to calibrate
either the discrete or isoconversional models. Regardless of which models are used,
one must be careful not to over-interpret the model parameters.
2.7 Identifying and Fitting the Right Kinetic Model
People involved in kinetics can be categorized as users or musers [50]. Users use
chemical kinetics for process design, petroleum exploration, or lifetime and safety
evaluations. They need reliable algorithms. Theoretical rigor is important only to
the extent it improves reliability. Musers muse over experimental and theoretical
concepts. Many often consider only qualitative issues. Some musers come up with
conceptual and algorithmic advances that are useful to users.
This section is directed towards users, not musers. Many of the initial steps of
deriving a reliable model involve the same conceptual evaluation used by musers.
However, once the correct conceptual model is chosen, users proceed to calibrate
the conceptual model parameters so it can be useful. This often requires more
sophisticated software than musers are willing to acquire. Even so, just because a
person has powerful software does not mean they know how to use it effectively to
derive practical and reliable models. Learning how to be effective is the objective
here.
Guidance on how to collect meaningful data is given by Vyazovkin et al. [109]
and only a few points are covered here. First and foremost, equipment must be
properly calibrated to achieve accurate temperatures, and sample sizes must be
small enough that thermal gradients during thermal transients do not invalidate the
temperature measurement [110]. For ordinary materials, the product of sample size
and heating rate should not exceed 100 mg °C/min [111]. For particularly
endothermic or exothermic materials (e.g., explosives), that product should be 20–
100 times lower, depending on the reaction enthalpy.
The next step in calibrating a model is to acquire the relevant data. That means
gathering rate or conversion data over a wide range of temperatures and time
(actually logarithmic time) to adequate decouple the pre-exponential and expo-
nential dependences in the Arrhenius equation. The width of the required temper-
ature or heating rate range depends on the reproducibility of a single determination
[50, 111]. As is obvious from the isoconversional principle, one needs longer time
2.7 Identifying and Fitting the Right Kinetic Model 61
data at lower temperatures to achieve conversion levels that are helpful for
decoupling conversion and temperature dependences.
For thermal analysis experiments, a properly conceived experimental plan will
have experiments at multiple temperatures, multiple heating rates, or both. This
provides a robust data set. Nonlinear temperature profiles, such as changing the
temperature during the run to maintain an approximately constant reaction rate or a
slow sinusoidal oscillation of the temperature superimposed upon a constant ramp,
can also provide a good test of reaction models. In fact, the oscillating ramp (e.g.,
*20 oscillations of 5–10 °C amplitude per ramp) samples both the activation
energy and frequency factor at various extents of conversion, so a single experiment
can derive kinetic parameters under ideal conditions.
Batch experiments are more problematic, in that often conversion is known only
at the end of the experiment. Consequently, it is highly desirable to conduct mul-
tiple times at a given temperature, including longer times at lower temperatures.
Sometimes batch experiments are conducted only at fixed time for various tem-
peratures (e.g., 72 h for hydrous pyrolysis). Such data are useful for checking a
model based on other information but are inadequate by themselves to calibrate a
model because they cannot decouple temperature and conversion dependences.
The general concept for identifying and calibrating applied chemical kinetic
models for condensed phase pyrolysis is shown in Fig. 2.18. First, one uses a
variety of simple inspection methods to narrow the range of model possibilities. The
first and most obvious step is to look at the data. This means plotting it in various
ways, most likely as reaction rate or fraction converted as a function of time (or
temperature in the case of a constant heating rate). Distinct humps or separate peaks
at a constant heating rate indicate multiple processes, which should be modeled
separately. Isothermal experiments are particularly helpful for identifying
Fig. 2.18 General

Preliminary analysis
methodology for deriving
-inspect reaction profiles for multiple reactions
reliable applied chemical
kinetic models -check constancy of E by isoconversional analysis
-examine profile shape for acceleratory,
deceleratory, or auto-catalytic character
Choose one or more

probable models
Single Complex
Reaction reaction
Linear Nonlinear Nonlinear

model model model
fitting fitting fitting
acceleratory or sigmoidal characteristics that are harder to detect in rising tem-

perature experiments. Recall that the acceleratory characteristics may not be evident
on a linear time scale, and conversion versus logarithmic time is also important to
examine. It is also important to know the exact thermal history for nominally
isothermal experiments in order to separate acceleration during heat-up from true
autocatalytic or related acceleration.
For the preliminary analysis, three methods are now explored in more depth.
They are isoconversional methods, LLNL approximate (extended Kissinger) anal-
yses, and reduced scale plots. The first two are useful only for thermal-analysis-type
data, i.e., many data points.
Isoconversional methods were described in some detail earlier, but specific
features are relevant at this stage of analysis. First, it is highly desirable to use both
the Friedman method and one of the advanced integral methods. The ICTAC
kinetics report [6] recommends either the Kissinger-Akahira-Sunose (KAS) method
or one of Vyazovkin’s methods. Kinetics2015 uses an iterative multi-heating rate
version of the Coats-Redfern equation [4], which is slightly more accurate than the
KAS method. Although it is discouraged to use many isoconversional methods
(including the obsolete OFW method) and compare them in detail, it is useful to use
both differential and integral methods, as they have different weaknesses and tend to
fail under different circumstances.
The first question to address with isoconversional methods is whether A and
E are roughly constant versus conversion or vary significantly. One must take into
consideration problems with noise in the data that tend to make the values for the
first and last 10 % unreliable, but they should not be discounted completely. If there
is a change, is it gradual or can the timing of it be related to some feature in the
data?
The LLNL approximate analyses were developed between 1986 and 1998 and
are based on an extended Kissinger analysis of constant heating rate data.
1. The first step is to determine an initial estimate of the characteristic A and E for
the reaction system using one of the Kissinger methods to derive first-order
parameters for something around mid-conversion. Equation (2.24) is used for
rate data, and Eq. (2.19) is used for cumulative reacted.
2. The next step is to calculate the first-order rate curve from these parameters and
compare it to the measured curves. Two characteristics are monitored: profile
width and profile asymmetry. If the profile is narrower than a first-order reaction
(ρ < 1), it is either autocatalytic or nth-order with n less than one. If it is wider
(ρ > 1), it is some type of distributed reactivity reaction.
3. Profile asymmetry relative to a first-order reaction is then used in combination
with relative width to estimate parameter values for various models, including
the extended Prout-Tomkins model, nth-order, and Gaussian and Weibull dis-
tribution models.
Figure 2.19 shows the fields of relevance for the various models, and Table 2.3
shows the algorithms used for the model parameter estimates for the Gaussian
2.5 2.0
2.0 1.6
Weibull
Asymmetry
Asymmetry
Weibull nth-order nth-order
1.5 1.2
Gaussian
1.0 0.8 nucleation
Gaussian
nucleation
0.5 0.4 1st-order

1st-order
nth-order nucleaƟon
0.0 0.0
0.5 1.0 1.5 2.0 2.5 0.5 1.0 1.5 2.0 2.5
FWHH T80%-T20%
Fig. 2.19 Plots of asymmetry versus relative reaction profile width to help determine the relevant
model needed for further refinement. For rate data, the asymmetry is (Thigh − Tmax)/(Tmax − Tlow),
where Thigh and Tlow are the temperatures at 25 % of the maximum rate. For fraction reacted data,
asymmetry is (T0.9 − T0.5)/(T0.5 − T0.1). In both cases, the relative profile width is the
experimental value divided by the calculated width of a first-order reaction having the same
A and E values. Reproduced courtesy of LLNL
Table 2.3 Examples of equations used to estimate selected model parameters in Kinetics2015,
plus the simpler original algorithm used to estimate σ
Parameter Rate data
First-order ln(β/T2max = − E/RTmax + ln(AR/E)
A and E
Gaussian σ σ = − 1.1/ρ3 − 0.66/ρ30 + 2.88ρ − 1.12
(orig)
Gaussian σ σ = {6.41 × 106[(ρ − 1)β3000/ρE/E1.2 − 3.7 × 109]}0.54 + 7.14 × 109
[(ρ − 1)2β12000/ρE/E2.2
Gaussian A Aσ = A{1 − 0.4[1 − exp(σ/2.5)]}
Reaction order n = (asym/0.64)0.78
n
Growth order m m = 1 − ρ1.92/n
ePT A AePT = A/(1 − m + 0.17m2)
distribution and ePT models. Comparable correlations are used for the Weibull
model and for all models using fraction reacted data.
Reduced scale plots originated as a way to collapse conversion data for different
temperatures onto a common plot on which their shapes can be compared to each
other and to common reaction models [61]. Commonly, times for 50 or 90 %
conversion are used for the scaling parameter. Recently, there has been more
emphasis on plotting certain rate and conversion functions versus conversion.
Several examples are given by Vyazovkin [6].
A simple starting point is to consider plotting the reaction rate for constant
heating as a function of conversion. Such a plot is shown in Fig. 2.20 for three
Fig. 2.20 Plot of normalized

1.0
reaction rate versus
conversion, which removes
profile width effects to reflect 0.8
one aspect of the appropriate
dα /dt (normalized)
reaction model. Reproduced
courtesy of LLNL 0.6
0.5 order
0.4
1st order
2nd order
0.2 polystyrene
LD polyethylene
HD polyethylene
0.0
0.0 0.2 0.4 0.6 0.8 1.0
α
polymers with different reaction shapes [6]. Such a plot is an alternative to the
approach in Fig. 2.19 for evaluating the profile shape against selected model
shapes, shown here for nth-order reactions. The decrease of nucleation-growth
character is not captured in such a plot, as it affects profile width but not profile
shape for a constant heating rate.
A different approach is to normalize the reaction rate or fraction reacted to unity
at 50 % conversion. This works well when one has a good way to determine
complete conversion, which is not always the case for mixtures reacted at constant
temperature or where the amount of residue depends on the heating conditions.
Nevertheless, it is often a good approach for identifying plausible models. Such
normalized master curves for various models are shown in Fig. 2.21.
A limitation of the approaches in Figs. 2.19 and 2.20 is that they work only for
constant heating rates. Similarly, the approaches shown in Figs. 2.2, 2.7 and 2.8
work only for isothermal conditions. The curves in Fig. 2.21 can be used directly
only for isothermal data, but a method has been developed for their use for arbitrary
thermal histories [112]. For the rate master curve, the method is conceptually
simple. One first does an isoconversional or Kissinger kinetic analysis to generate
an activation energy near mid conversion and then uses it to calculate what the
reaction rate would have been had the temperature been the same as at
mid-conversion. This gives the relation

ðda=dtÞi E 1 1
Generalized Reduced Reaction Rate ¼ exp ð2:49Þ
ðda=dtÞ0:5 R T i T0:5
where the subscript i indicates the rate and temperature at the ith data point and the
subscript 0.5 indicates the rate and temperature at 50 % conversion. Details on the
integral method are given by Gotor et al. [112] and an example is given by
Sánchez-Jiménez et al. [97] for cellulose heated at a constant heating rate showing
its similarity to nucleation-growth and random scission models.
Fig. 2.21 Master plots of 4.0

reaction rates and fractions n=1/2
Reduced fraction reacted

reacted for various reaction n=1
models 3.0 n=2
AE with p=2
AE with p=4
2.0 1D diffusion
3D diffusion
1.0
0.0
0.0 0.2 0.4 0.6 0.8 1.0
4.0
n=1/2
Reduced reaction rate
n=1
3.0 n=2
AE with p=2
AE with p=4
1D diffusion
2.0
3D diffusion
1.0
0.0
0.0 0.2 0.4 0.6 0.8 1.0
Fraction reacted
A comparable plot is shown in Fig. 2.22 for the sample Kel-F data shown in
Fig. 2.10. (In this case, the activation energy is essentially constant, so the calcu-
lation of the reduced reaction rate is simple.) The constant heating rate and
isothermal data collapse on the same general curve. For the models shown in
Fig. 2.21, the match is best with the Avrami-Erofe’ev model with p = 2, but the
peak is clearly skewed to lower conversion. Comparing these plots to the cellulose
data of Sánchez-Jiménez et al. [97] the agreement is actually better for a random
scission kinetic model or an ePT model, although the two are similar in many
respects.
Many synthetic polymers are likely to follow a nucleation-growth model.
Figure 2.23 shows the variety of profile shapes available via the ePT model, all of
which have an acceleratory phase. The initial rate is determined by the quantity
1 − q, which is proportional to the ratio of initiation to propagation rates. Plotting
the experimental data on this type of plot can help identify an appropriate model
and initial guesses for non-linear regression. However, it is equally important to
consider the profile width as well as shape. That is particularly easy using a plot like
that in Fig. 2.19.
Sanchez-Jimenez et al. [113] recently used a simple random scission model
adapted from the work of Simha and Wall [114] to fit polymer decomposition data.
Fig. 2.22 Generalized 2.0
Generalized reduced reacƟon rate

reaction rate master plot for Isothermal 400C
Kel-F 800 2.5 C/min
1.6 5 C/min
20 C/min
1.2 10 C/min
0.8
0.4
0.0
0.0 0.2 0.4 0.6 0.8 1.0
FracƟon reacted
Fig. 2.23 Various profile 1.4

shapes available via the ePT

nucleation-growth model. 1.2
Note that these curves do not
address profile width, which 1.0
is equally important
0.8
0.6 m=1, n=1, q=0.99

m=1, n=1, q=0.999
0.4 m=0.5, n=1, q=0.99
m=0.5, n=1, q=0.999
0.2 m=0.25, n=1, q=0.99
m=0.25, n=1, q=0.999
0.0
0.0 0.2 0.4 0.6 0.8 1.0
1.4
1.2
1.0
0.8
0.6
m=1, n=0.5, q=0.99
0.4 m=1, n=0.5, q=0.999
m=1, n=2, q=0.99
0.2
m=1, n=2, q=0.999
0.0 m=0.5, n=0.5, q=0.999
0.0 0.2 0.4 0.6 0.8 1.0
FracƟon reacted
Although Burnham et al. [79] had already made the qualitative connection,
Sanchez-Jimenez et al. provide a more quantitative comparison. For the
chain-length parameter L favored in that paper, the random scission model reduces
to f(α) = 2(α1/2 − α). They report n = 1.12 and m = 0.40 as the equivalent ePT
parameters from their linear regression method. Similarly, Kinetics2015 obtains
n = 1.13 and m = 0.48. A point rarely mentioned in the literature is that sigmoidal
reactions need to start with a finite amount of reaction or the rate will be identically
zero for all time. In this case, α = 0.001 was assumed at time zero to create the
simulated data, which corresponds to q = 0.999 in the non-linear regression.
The application of generalized reduced reaction rate plots is also instructive for
materials with a reactivity distribution. Since, by definition, it is likely that the
activation energy will increase with conversion, it is not immediately obvious
whether one should use the mean activation energy or the isoconversional activa-
tion energy at the indicated level of conversion. However, numerical studies show
unambiguously that the mean activation energy is to be used throughout.
A reactivity distribution model is clearly more deceleratory than a first-order
reaction under isothermal conditions, but using the activation energy at the ith-level
of conversion gives an acceleratory phase, while using the mean energy gives
superimposable reaction rate plots for isothermal and constant-heating-rate
conditions.
A comparison of this method for model Gaussian reactions, a high-volatile coal,
and a petroleum source rock are shown in Fig. 2.24. The Gaussian model used a mean
activation energy of 30,070 K, a frequency factor of 2 × 1016 s−1, and an energy
distribution of 4.8 % of the mean. The petroleum source rock is fitted well by a
Gaussian energy distribution model having an energy distribution of 1.7 % of the
mean. The Pittsburgh #8 high volatile bituminous coal is more problematic, in that its
shape is not well represented by a Gaussian distribution, so depending on the fitting
method, one gets distributions of 3.5–7 % of the mean energy. That is qualitatively
consistent with its similarity to the 4.8 % Gaussian generalized reaction rate plot.
Fig. 2.24 Generalized 10

reaction rate master plot for a 4.8% Gaussian, 1 °C/min
Gaussian energy distribution
Generalized reaction rate
8 4.8% Gaussian, isothermal

model and data for a coal and
petroleum source rock Petroleum source rock 6.8 °C/min
6 Pittsburgh #8 coal 10 °C/min
0
0.0 0.2 0.4 0.6 0.8 1.0
Fraction reacted
Once a model has been selected, one can proceed to nonlinear regression to
optimize the model parameters. Good initial parameter estimates for refinement
must be generated by some combination of the methods outlined above, as poor
estimates may lead to a false minimum or complete failure of convergence. Our
experience is that no more than 4–5 parameters should be refined at once, so
holding some constant and working manually toward a global minimum is war-
ranted for complex models, particularly those involving overlapping reaction pro-
files. Although it might be nice to have an explicit recipe, each situation is slightly
different, and one must learn from experience and comparison to the approaches of
others.
Another consideration is what objective function to minimize. Ordinarily, one
would fit rate data to a rate equation by minimizing those residuals and fit fraction
reacted data to an integrated equation by minimizing those residuals.1 It is
important always to collect the complete thermal history and integrate the rate laws
over that thermal history rather than assume an isothermal or constant heating rate
analytical expression. Such approximations are not needed with today’s computa-
tional power and can lead to substantial errors in kinetic parameters, particularly for
nominally isothermal experiments.
However, data can be integrated or differentiated to obtain the other form to
regress, and there may be reasons to do so. If so, one should be sure to use
mathematical algorithms for differentiating and integrating that do not introduce
numerical errors. It may also be valuable to minimize the residuals simultaneously
on both rates and integrals, because they are sensitive to different aspects of the
reaction data. A final caution is that it is ordinarily best to weight each experiment
equally rather than each data point equally, because slower experiments typically
have more data points but do not contain significantly more information to drive the
optimization.
Exercises
1. What can and cannot be learned about global chemical kinetic parameters from
transition-state theory?
2. Explain one example for which the approximation dα/dt = k(T)f(α) is not valid.
3. Show that the formulas for a pseudo nth-order reaction in Table 2.1 are
equivalent to Eq. (2.11) in Burnham and Braun, Energy & Fuels 13, 1–22
(1999).
4. Kinetic Tmax values of a first-order reaction are 434.5, 460.6, and 488.6 °C at
constant heating rates of 2, 8 and 32 °C/min. Write a simple spreadsheet or
other program for Kissinger’s method and calculate A and E.
5. If there is a measurement error for Tmax of +1 °C at 32 °C/min and −1 °C at
2 °C/min, what are the resulting A and E? Now repeat the calculation with the
1
The presumption here in both cases is that the fit will be to the data as collected and not a
logarithmic form or a linearized transformation. Such manipulation tends to magnify the impor-
tance of data with large error bars, leading to poorer agreement on the most important portions.
signs of the temperature errors reversed. Plot the ln(A) versus E for these two
pairs along with the pair from problem 4. What does it mean?
6. Write a spreadsheet or other program to calculate the reaction rate and fraction
reacted of a first-order reaction for (at least) a constant heating rate. Justify your
choice of integration method. Generate simulated data for three parallel reac-
tions with A = 1×1013 s−1, E = 49, 51, and 53 kcal/mol, and weighting factors
of 0.2, 0.5 and 0.3, respectively, at 1, 5 and 25 °C/min. What are the tem-
peratures for 50 % conversion at the three heating rates?
7. For the simulated data from problem 5, what is the full width at half height of
the reaction rate profile at 5 °C/min? What is the full width at half height of the
reaction rate profile for the principal reaction channel (51 kcal/mol) at the same
heating rate? From the ratio of the two profile widths and the simpler equation
in Table 2.3, what is an estimate for the Gaussian E-distribution parameter, σ.
8. For the simulated data from problem 5, use Friedman’s method to calculate the
instantaneous first-order A and E values at 10, 30, 50, 70 and 90 % conversion.
9. What is the purpose of the parameter q in the extended Prout-Tompkins model
(Eq. 2.38) and what is the approximation associated with it having a constant
value.
10. Write a spreadsheet or other program to calculate the reaction rate and fraction
reacted for the extended Prout-Tompkins model (Eq. 2.38) for an arbitrary
thermal history. Plot the reaction rate curves at 480 °C and 10 °C/min for
A = 1×1014 s−1, E = 54, q = 0.99, and m = 0.0, 0.5, and 1.0. Plot the isothermal
times between 10 and 50 % conversion versus the profile full widths at half
height for m = 0.0, 0.5 and 1.0. Why does the curve have a negative slope?
References
1. Originally Published by S. Arrhenius in Z. Phys. Chem. 4, p. 226, It is covered in any

undergraduate physical chemistry text (1889)
2. Boltzmann Distribution (2016), https://en.wikipedia.org/wiki/Boltzmann_distribution.
Accessed 7 Aug 2016
3. Maxwell-Boltzmann Distribution (2016), https://en.wikipedia.org/wiki/Maxwell-
Boltzmann_distribution. Accessed 7 Aug 2016
4. A.K. Burnham, R.L. Braun, Global kinetic analysis of complex materials. Energy Fuels 13,
1–22 (1999)
5. Transition-state theory was first published by, H. Eyring in 1935, and its history is covered in
a special journal issue in his memory by K.J. Laidler, M.C. King, The development of
transition-state theory. J. Phys. Chem. 87, 2657–2664 (1983)
6. S. Vyazovkin, A.K. Burnham, J.M. Criado, L.A. Perez-Maqueda, C. Popescu, N.
Sbirrazzuoli, ICTAC recommendations for performing kinetic computations on thermal
analysis data. Thermochim. Acta 520, 1–19 (2011)
7. J.R. Frade, M. Cable, Theoretical solutions for mixed control of solid state reactions. J. Mat.
Sci. 32, 2727–2733 (1997)
8. Z. Du, A.F. Sarofim, J.P. Longwell, C.A. Mims, Kinetic measurement and modeling of
carbon oxidation. Energy Fuels 5, 214–221 (1991)
9. K.J. Jackson, A.K. Burnham, R.L. Braun, K.G. Knauss, Temperature and pressure
dependence of n-hexadecane cracking. Org. Geochem. 23, 941–953 (1995)
10. A.K. Burnham, R.K. Weese, A.P. Wemhoff, J.L. Maienschein, A historical and current
perspective on predicting thermal cookoff behavior. J. Therm. Anal. Cal. 89, 407–415 (2007)
11. E.A. Glascoe, J.M. Zaug, A.K. Burnham, Pressure-dependent decomposition kinetics of the
energetic material HMX up to 3.6 GPa. J. Phys. Chem. A 113, 13548–13555 (2009)
12. H. Freund, J.A. Couse, G.A. Otten, Effects of pressure on the kinetics of kerogen pyrolysis.
Energy Fuels 7, 1088–1094 (1993)
13. D.B. Anthony, J.B. Howard, Coal devolatilization and hydrogasification. AIChE J. 22, 625–
656 (1976)
14. E.A. Dorko, W. Bryant, T.L. Regulinski, Solid-state reaction kinetics. IV. Analysis of
chemical reactions by means of the Weibull function. Anal. Calorim. 3, 505–509 (1974)
15. R. Aris, Reactions in continuous mixtures. AIChE J. 35, 539–548 (1989)
16. J.H. Flynn, The temperature integral—its use and abuse. Thermochim. Acta 300, 83–92
(1997)
17. K.H. Van Heek, H. Juntgen, W. Peters, Kinetics of nonisothermal reactions, using thermal
decomposition reactions as an example. Ber. Bunsenges. Physik. Chem. 71, 113–121 (1967)
18. J.H. Campbell, G.J. Koskinas, N.D. Stout, Kinetics of oil generation from Colorado oil
shale. Fuel 57, 372–376 (1978)
19. A.W. Coats, J.P. Redfern, Kinetic parameters from thermogravimetric data. Nature 201, 68–
69 (1964)
20. C. Popescu, E. Segal, On the temperature integral in non-isothermal kinetics with linear
heating rate. Thermochim. Acta 75, 253–257 (1984)
21. G.I. Senum, R.T. Yang, Rational approximations of the integral of the Arrhenius function.
J. Therm. Anal. 11, 446 (1977)
22. L.A. Perez-Maqueda, J.M. Criado, The accuracy of Senum and Yang’s approximations to
the Arrhenius integral. J. Thermal Analysis 60, 909–915 (2000)
23. R.L. Braun, A.K. Burnham, Analysis of chemical reaction kinetics using a distribution of
activation energies and simpler models. Energy Fuels 1, 153–161 (1987)
24. W. Gautschi, W.F. Cahill, in Handbook of Mathematical Functions; ed. by M. Abramowitz,
I.A. Stegun. (National Bureau of Standards, AMS 55, Washington, D. C, 1964), pp. 228–
231
25. A.C. Hindmarsh, LSODE and LSODI two new. Initial value ordinary Differential Equation
Solvers. ACM SIGNUM Newsl. 15(4), 10–11 (1980)
26. S.V. Gulikeri, D. Luss, Analysis of activation energy of grouped parallel reactions.
AIChE J. 18, 277–282 (1972)
27. A.K. Burnham, L.N. Dinh, A comparison of isoconversional and model-fitting approaches to
kinetic parameter estimation and application predictions. J. Therm. Anal. Cal. 89, 479–490
(2007)
28. S. Vyazovkin, Isoconversional Kinetics of Thermally Stimulated Processes (Springer, 2015)
29. H.L. Friedman, Kinetics of thermal degradation of char-forming plastics from thermo-
gravimetry: application to a Phenolic Plastic. J. Polym. Sci., Part C. 6, 183–195 (1964)
30. T. Ozawa, A new method for analyzing thermogravimetric data. Bull. Chem. Soc. Jpn. 38,
1881 (1965)
31. J.H. Flynn, L.A. Wall, A quick, direct method for the determination of activation energy
from thermogravimetric data. Polymer Lett. 4, 323–328 (1966)
32. J.H. Flynn, The isoconversional method for determination of energy of activation at constant
heating rates: corrections for the Doyle approximation. J. Therm. Anal. 27, 95–102 (1983)
33. J.H. Flynn, Early papers by Takeo Ozawa and their continuing relevance. Thermochim. Acta
282(283), 35–42 (1996)
34. C.D. Doyle, Estimating isothermal life from thermogravimetric data. J. Appl. Polym. Sci. 6,
639–642 (1962)
References 71
35. T. Ozawa, Estimation of activation energy by isoconversional methods. Thermochim. Acta

203, 159–165 (1992)
36. M.J. Starink, Activation energy determination for linear heating experiments: a comparison
of the accuracy of isoconversional methods. Thermochim. Acta 404, 163–176 (2003)
37. S. Vyazovkin, Modification of the integral isoconversional procedure to account for
variation in the activation energy. J. Comput. Chem. 22, 178–183 (2001)
38. M.E. Brown, M. Maciejewski, S. Vyazovkin, R. Nomen, J. Sempere, A. Burnham,
J. Opfermann, R. Strey, H.L. Anderson, A. Kemmler, R. Drueleers, J. Janssens, H.O.
Desseyn, C.-R. Li, T.B. Tang, B. Roduit, J. Malek, T. Mitsuhashi, Computational aspects of
kinetic analysis. Part A: the ICTAC Kinetics Project—data, methods and results.
Thermochim. Acta 355, 125–143 (2000)
39. K. Miura, A new and simple method to estimate f(E) and k0(E) in the distributed activation
energy model from three sets of experimental data. Energy Fuels 9, 302–307 (1995)
40. K. Miura, T. Maki, A simple method for estimating f(E) and k0(E) in the distributed
activation energy model. Energy Fuels 12, 864–869 (1998)
41. V. Vand, A theory of the irreversible electric resistance changes of metallic films evaporated
in vacuum, Proc. Phys. Soc. (London, 1943), pp. 222–246
42. http://www.geoisochem.com/software/kinetics2015/index.html. Accessed 7 Aug 2016
43. http://www.akts.com/akts-thermokinetics-tga-dsc-dta-tma-ftir-ms/download-tga-dsc-dta-
tma-ms-ftir-akts-thermokinetics-software.html. Accessed 7 Aug 2016
44. H.E. Kissinger, Reaction kinetics in differential thermal analysis. Anal. Chem. 29, 1702–
1706 (1957)
45. H.E. Kissinger, Variation of peak temperature with heating rate in differential thermal
analysis. J. Res. Nat. Bur. Stand. 57, 217–221 (1956)
46. D. Chen, X. Gao, D. Dollimore, A generalized form of the Kissinger equation. Thermochim.
Acta 215, 109–117 (1993)
47. P.A. Redhead, Thermal desorption of gases. Vacuum 12, 203–211 (1962)
48. V. Dieckmann, Modelling petroleum formation from heterogeneous source rocks: the
influence of frequency factors on activation energy distribution and geological prediction.
Mar. Petrol. Geol. 22, 375–390 (2005)
49. Simple Linear Regression, http://en.wikipedia.org/wiki/Simple_linear_regression. Accessed
6 Mar 2014
50. A.K. Burnham, Obtaining reliable phenomenological chemical kinetic models for real-world
applications. Thermochim. Acta 597, 35–40 (2014)
51. A. Yelon, E. Sacher, W. Linert, Comment on the mathematical origins of the kinetic
compensation effect parts 1 and 2 by P.J. Barrie, Phys. Chem. Chem. Phys. 14, 318 and 327
(2012). Phys. Chem. Chem. Phys. 14, 8232–8234 (2012)
52. P.J. Barrie, Reply to comment on the mathematical origins of the kinetic compensation
effect. parts 1 and 2 by A. Yelon, E. Sacher, W. Linert, Phys. Chem. Chem. Phys. 14,
(2012). Phys. Chem. Chem. Phys. 14, 8235–8236 (2012)
53. Non-linear Least Squares, http://en.wikipedia.org/wiki/Non-linear_least_squares. Accessed
7 Aug 2016
54. Levenberg-Marquardt Algorithm, http://en.wikipedia.org/wiki/Levenberg%E2%80%
93Marquardt_algorithm. Accessed 7 Aug 2016
55. C.C. Lakshmanan, M.L. Bennett, N. White, Implications of multiplicity in kinetic
parameters to petroleum exploration: distributed activation energy models. Energy Fuels
5, 110–117 (1991)
56. J. Hillier, T. Bezzant, T.H. Fletcher, Improved method for the determination of kinetic
parameters from non-isothermal thermogravimetric analysis (TGA) data. Energy Fuels 24,
2841–2847 (2010)
57. https://www.netzsch-thermal-analysis.com/us/products-solutions/software/netzsch-
advanced-software/thermokinetics/. Accessed 7 Aug 2016
58. D.A. Ratkowsky, Handbook of Nonlinear Regression Models (Marcel Dekker, 1990),
pp. 36–48
59. A.K. Burnham, An nth-order Gaussian energy distribution model for sintering. Chem. Eng.
J. 108, 47–50 (2005)
60. B. Roduit, M. Hartmann, P. Folly, A. Sarbach, R. Baltensperger, Prediction of thermal
stability of materials by modified kinetic and model selection approaches based on limited
amount of experimental points. Thermochim. Acta 579, 31–39 (2014)
61. W.E. Brown, D. Dollimore, A.K. Galway, in Comprehensive Chemical Kinetics, ed. by C.H.
Bamfor, C.F. Tipper. Reactions in the solid state, vol. 22 (Elsevier, 1980), pp 49–65
62. M. Avrami, Kinetics of phase change, I. General theory, J. Chem. Phys. 7, pp. 1103–1112
(1939); II. Transformation-time relation for random distribution of nuclei, J. Chem. Phys. 8,
pp. 212–224 (1940); III. Granulation, phase change, and microstructure, J. Chem. Phys. 8,
pp. 177–184
63. W. Johnson, R. Mehl, Reaction kinetics in processes of nucleation and growth. Trans. AIME
135, 416–442 (1939)
64. A. Kolmogorov, A statistical theory for the recrystallization of metals, Akad. Nauk. SSSR,
Izv. Ser. Matem. 1, pp. 355–359 (1937)
65. B.V. Erofe’ev, A generalized equation of chemical kinetics and its application in reactions
involving solids, C.R. Dokl. Akad. Sci. USSR 52, pp. 511–514 (1946)
66. E.G. Prout, F.C. Tompkins, The thermal decomposition of potassium permanganate. Trans.
Faraday Soc. 40, 488–498 (1944)
67. U. Retter, D. Vollhardt, Chain nucleation: a critical consideration of Prout and Tompkins’
Law. Langmuir 8, 1693–1694 (1992)
68. J. Šesták, G. Berggren, Study of the kinetics of the mechanism of solid-state reactions at
increasing temperatures. Thermochim. Acta 3, 1–12 (1971)
69. J.B. Austin, R.L. Rickett, Kinetics of the decomposition of austenite at constant temperature,
AIME Technical publication. 964, p. 20 (1938). Trans. AIME 135, pp. 396–415
70. E.G. Prout, F.C. Tompkins, The thermal decomposition of silver permanganate. Trans. Far.
Soc. 42, 468–472 (1946)
71. B.V. Erofe’ev, in Reactivity of Solids, Proceedings of the 4th Intl. Symp., (Elsevier,
Amsterdam, 1961) pp. 273–282
72. M.A. Arshad, A. Maaroufi, Relationship between Johnson-Mehl-Avrami and
Šesták-Berggren models in the kinetics of crystallization in amorphous materials.
J. Non-Crystalline Solids 413, 53–58 (2015)
73. N.S. Akulov, Basics of Chemical Dynamics (in Russian) (Moscow State University, 1940)
74. M. Rios-Fachal, C. Garcia-Fernandez, J. Lopez-Beceiro, S. Gomez-Barreiro,
J. Tarrio-Saavedra, A. Ponton, R. Ariaga, Effect of nanotubes on the thermal stability of
polystyrene, J. Therm. Anal. Cal. 113, 481–487 (2013)
75. J. Tarrio-Saavedra, J. Lopez-Beceiro, S. Naya, M. Francisco-Fernandez, R. Artiaga,
Simulation study for generalized logistic function in thermal data modeling. J. Therm. Anal.
Cal. 118, 1253–1268 (2014)
76. A.K. Burnham, Use and misuse of logistic equations for modeling chemical reactions,
J. Therm. Anal. Cal. (2015)
77. Cheminform, St. Petersburg (CISP) Ltd, 197198, (14, Dobrolubov Ave, Saint Petersburg,
Russia)
78. M.E. Brown, The Prout-tompkins rate equation in solid-state kinetics. Thermochim. Acta
300, 93–106 (1997)
79. A.K. Burnham, R.L. Braun, T.T. Coburn, E.I. Sandvik, D.J. Curry, B.J. Schmidt, R.A.
Noble, An appropriate kinetic model for well-preserved algal kerogens. Energy Fuels 10,
49–59 (1996)
80. F.O. Rice, K.F. Herzfeld, The thermal decomposition of organic compounds from the
standpoint of free radicals. VI. The mechanism of some chain reactions, J. Amer. Chem. Soc.
56, 284–289 (1934)
81. J.H. Flynn, R.E. Florin, in Pyrolysis and GC in polymer analysis, ed. by S.A. Liebman, E.
J. Levy. Degradation and pyrolysis mechanisms, (Marcel Dekker, 1985) pp. 149–208
References 73
82. R.E. Lyon, M.T. Takemori, N. Safronava, S.I. Stoliarov, R.N. Walters, A molecular basis for
polymer flammability. Polymer 50, 2608–2617 (2009)
83. L.A. Wall, in Analytical Chemistry of Polymers, Part II, Chapter 5, ed. by G.M. Kline.
Pyrolysis,(Academic Press, 1962) pp. 181–248
84. L.A. Wall, S. Straus, Pyrolysis of polyolefins. J. Polym. Sci. 44, 313–323 (1960)
85. A.K. Burnham, R.L. Braun, Kinetics of Polymer Decomposition (Lawrence Livermore
National Laboratory Rept. UCID-21293, 1987) p. 28
86. A.K. Burnham, R.L. Braun, R.W. Taylor, T.T. Coburn, Comparison of isothermal and
nonisothermal pyrolysis data with various rate mechanisms: implications for kerogen
structure. Prepr. ACS Div. Petrol. Chem. 34(1), 36–42 (1989)
87. A.K. Burnham, Application of the Šesták-Berggren equation to organic and inorganic
materials of practical interest. J. Therm. Anal. Cal. 60, 895–908 (2000)
88. A.K. Burnham, R.K. Weese, Kinetics of thermal degradation of explosive binders Viton A,
Estane, Kel-F. Thermochim. Acta 426, 85–92 (2005)
89. J.D. Nam, J.C. Seferis, Generalized composite degradation kinetics for polymeric systems
under isothermal and nonisothermal conditions. J. Poly Sci. B Poly. Phys. 30, 455–463
(1992)
90. J.G. Reynolds, A.K. Burnham, Pyrolysis decomposition kinetics of cellulose-based materials
by constant heating rate micropyrolysis. Energy Fuels 11, 88–97 (1997)
91. R. Capart. L. Khezami, A.K. Burnham, Assessment of various kinetic models for the
pyrolysis of a microgranular cellulose. Thermochim. Acta. 417, 79–89 (2004)
92. D. Dollimore, B. Holt, Thermal degradation of cellulose in nitrogen. J. Poly. Sci. 11, 1703–
1711 (1973)
93. M. Grønli, M.J. Antal Jr., G. Várhegyi, A round-robin study of cellulose pyrolysis kinetics
by thermogravimetry. Ind. Eng. Chem. Res. 38, 2238–2244 (1999)
94. A.K. Burnham, X. Zhou, L.J. Broadbelt, Critical review of the global chemical kinetics of
cellulose thermal decomposition. Energy Fuels 29, 2906–2918 (2015)
95. S. Kim, Y. Eom, Estimation of kinetic triplet of cellulose pyrolysis reaction from isothermal
kinetic results. Kor. J. Chem. Eng. 23, 409–414 (2006)
96. H. Barud, C. Riberio, J. Capella, M.S. Crespi, S. Ribeiro, Y. Messadeq, Kinetic parameters
for thermal decomposition of microcrystalline, vegetal, and bacterial cellulose. J. Therm.
Anal. Cal. 105, 421–426 (2011)
97. P. Sanchez-Jimenez, L. Perez-Maqueda, A. Perejon, J.M. Criado, Generalized master plots
as a straightforward approach for determining the kinetic model: the case of cellulose
pyrolysis. Thermochim. Acta 552, 54–59 (2013)
98. F.H. Constable, The mechanism of catalytic decomposition. Proc. Royal. Soc. A 108, 355–
378 (1925)
99. G.J. Pitt, The kinetics of the evolution of volatile products from coal. Fuel 41, 267–274
(1962)
100. P. Hanbaba, Reaktionkinetische Untersuchungen sur Kohlenwasserstoffenbindung aus
Steinkohlen bie niedregen Aufheizgeschwindigkeiten. Dissertation, University of Aachen,
(1967)
101. P. Ungerer, R. Pelet, Extrapolation of the kinetics of oil and gas generation from laboratory
experiments to sedimentary basins. Nature 327, 52–54 (1987)
102. L. Kolar-Anić and S. Veljković, Weibull distribution and kinetics of heterogeneous
processes. J. Chem. Phys. 63, pp. 663–668 (1975); Statistical foundations of heterogeneous
kinetics, ibid., pp 669–673
103. T.C. Ho, R. Aris, On apparent second-order kinetics. AIChE J. 33, 1050–1051 (1987)
104. J. Cai, R. Liu, Weibull mixture model for modeling nonisothermal kinetics of thermally
stimulated solid-state reactions: application to simulated and real kinetic conversion data.
J. Phys. Chem. B. 111, 10681–10686 (2007)
105. E.M. Suuberg, Approximate solution technique for nonisothermal, gaussian distributed
activation energy models. Combust. Flame 50, 243–245 (1983)
106. A.K. Burnham, R.L. Braun, H.R. Gregg, A.M. Samoun, Comparison of methods for
measuring kerogen pyrolysis rates and fitting kinetic parameters. Energy Fuels 1, 452–458
(1987)
107. P. Sundararaman, P.H. Merz, R.G. Mann, Determination of kerogen activation energy
distribution. Energy Fuels 6, 793–803 (1992)
108. J. Cai, T. Li, R. Liu, A critical study of the Miura-Maki integral method for the estimation of
the kinetic parameters of the distributed activation energy model. Bioresour. Technol 102,
3894–3899 (2011)
109. S. Vyazovkin, K. Chrissafis, M.L. Di Lorenzo, N. Koga, M. Pijolet, B. Roduit, N.
Sbirrazzuoli, J.J. Sunol, ICTAC Kinetics Committee recommendations for collecting
experimental thermal analysis data for kinetic computations. Thermochim. Acta 590, 1–23
(2014)
110. R.E. Lyon, N. Safronova, J. Senese, S. Stoliarov, Thermokinetic model of sample response
in nonisothermal analysis. Thermochim. Acta 545, 82–89 (2012)
111. K.E. Peters, A.K. Burnham, C.C. Walters, Petroleum generation kinetics: single versus
multiple heating-ramp open-system pyrolysis. AAPG Bull. 99, 591–616 (2015)
112. F.J. Gotor, J.M. Criado, J. Malak, N. Koga, Kinetic analysis of solid-state reactions: the
universality of master plots for analyzing isothermal and nonisothermal experiments.
J. Phys. Chem. A 104, 10777–10782 (2000)
113. P.E. Sanchez-Jimenez, L.A. Perez-Maqueda, A. Perejon, J.M. Criado, A new model for the
kinetic analysis of thermal degradation of polymers driven by random scission. Poly.
Degrad. Stab. 95, 733–739 (2010)
114. R. Simha, L.A. Wall, Kinetics of chain depolymerization. J. Phys. Chem. 56, 707–715
(1952)
http://www.springer.com/978-3-319-49633-7

Introduction To Chemical Kinetics

Uploaded by

Copyright:

Available Formats

Introduction To Chemical Kinetics

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Introduction To Chemical Kinetics

Uploaded by

Copyright:

Available Formats

Chapter 2

Introduction to Chemical Kinetics

Chemical kinetics describe of the rate at which chemical structures change.

© Springer International Publishing AG 2017 25

2.1 The Activation Energy and Frequency Factor

where A is a frequency factor, as it has units of reciprocal time in addition to other

Fig. 2.1 Schematic

k is the reaction rate constant

While some people studying complex reactions in a phenomenological or global

2.2 Introduction to Conversion and Pressure Dependence

It is commonly assumed for practical purposes that the dependences on conversion,

dx=dt ¼ kðTÞf ð1 xÞhðPÞ ð2:4Þ

da=dt ¼ kðTÞf ðaÞhðPÞ: ð2:5Þ

Table 2.1 Selection of commonly used kinetic models

dxi =dt ¼ ki ðTÞf ð1 xi Þ: ð2:7Þ

A Gaussian distribution in activation energy, which is not the same as a

2.3 Integrating the Reaction Rate

it is important to achieve high conversions over the entire range of temperatures to

For ﬁrst and nth-order reactions, respectively, this equation evolves to

For isothermal conditions, k is constant, so one obtains the familiar equations

In the middle of the 20th century, it became fashionable to conduct kinetic

and y = E/RT, a1 = 2.334733, a2 = 0.250621, b1 = 3.330657, and b2 = 1.681534.

Here a provision is made for a variable frequency factor as could be derived

2.4 Methods of Kinetic Analysis

2.4.1 Isoconversional Methods

Fig. 2.3 Graphical

lnðbi Þ ¼ const 1:052Ea =RTa;i : ð2:20Þ

Slightly more accurate activation energies can be obtained by replacing the

Vyazovkin’s advanced integral method obtains activation energies practically

This must be solved by iteration until convergence by substituting previous

For a ﬁrst order reaction, df(x)/dx = − 1, so

isoconversional methods. If one wants to model two different measures of reaction

2.4.2 Model Fitting Methods

The example of Kissinger’s equation in Eq. (2.26) means plotting ln(β/T2max)

2.48e14 s-1 2.09e13 s-1

Fig. 2.4 Two realizations of a Kissinger plot including measurement noise

2.5 Sigmoidal Reaction Models

As mentioned in the previous section, the isothermal reaction rate of sigmoidal

which corresponds to the rate form

da=dt ¼ kð1 aÞa: ð2:32Þ

The Prout-Tompkins model is more closely related to autocatalytic reactions

da=dt ¼ kam ð1 aÞn ½lnð1 aÞs : ð2:33Þ

ln½a=ð1 aÞ ¼ klnt þ c; ð2:34Þ

for analyzing decomposition kinetics of austenite, an allotrope of iron or steel. The

1 a ¼ 1=ð1 þ sebðtcÞ Þ1=s ð2:35Þ

where b is a reaction rate parameter, t is either time or temperature, and c is the

da=dt ¼ bebðtcÞ =ð1 þ sebðtcÞ Þs þ 1

da=dt ¼ k1 ð1 aÞn1 þ k2 am ð1 aÞn2 : ð2:37Þ

da=dt ¼ k2 ð1 aÞn ðam þ zÞ  k2 ð1 aÞn ½1 qð1 aÞm ð2:38Þ

where z = k1/k2 ≈ 1 − q. The left-hand approximation is used in CISP software

a ¼ 1=ðekt ð1 a0 Þ=a0 þ 1Þ ð2:39Þ

Fig. 2.5 Comparison of extended Prout-Tompkins and Avrami-Erofe’ev approaches to modeling

0.2 m=0.34, n=1.67, q=0.999 0.2

Random iniƟaƟon End iniƟaƟon

RelaƟve reacƟon rate

small zip small zip

Fig. 2.8 Effect of chain

generating a volatile fragment for a given scission is proportional to the amount of

Fig. 2.9 Fit of the ePT model 1.0

da=dt ¼ k2 ð1 aÞn ðam þ zÞ k2 ð1 aÞn ½1 qð1 aÞm ð2:38Þ