You are on page 1of 10

Fuel 143 (2015) 438447

Contents lists available at ScienceDirect

Fuel
journal homepage: www.elsevier.com/locate/fuel

Model-free rate expression for thermal decomposition processes: The


case of microcrystalline cellulose pyrolysis
Lina N. Samuelsson a, Rosana Moriana b,, Matthaus U. Babler a, Monica Ek b, Klas Engvall a
a
b

Department of Chemical Engineering and Technology, KTH Royal Institute of Technology, SE-100 44 Stockholm, Sweden
Department of Fiber and Polymer Technology, KTH Royal Institute of Technology, SE-100 44 Stockholm, Sweden

h i g h l i g h t s
 We derive a model-free rate expression using isoconversional principles.
 Thermal and kinetic analysis revealed inuence of initial sample mass.
 Derived kinetics, based on dynamic data, can model isothermal cellulose pyrolysis.

a r t i c l e

i n f o

Article history:
Received 22 September 2014
Received in revised form 21 November 2014
Accepted 24 November 2014
Available online 4 December 2014
Keywords:
Model-free
Isoconversional
Kinetics
Thermogravimetric analysis
Pyrolysis
Cellulose

a b s t r a c t
We explore the possibility to derive a completely model-free rate expression using isoconversional
methods. The Friedman differential method (Friedman, 1964) and the incremental integral method by
Vyazovkin (2001) were both extended to allow for an estimation of not only the apparent activation
energy but also the effective kinetic prefactor, dened as the product of the pre-exponential factor and
the conversion function. Analyzing experimental thermogravimetric data for the pyrolytic decomposition
of microcrystalline cellulose, measured at six different heating rates and three different initial sample
masses (1.510 mg), revealed the presence of secondary char forming reactions and thermal lag, both
increasing with increased sample mass. Conditioning of the temperature function enables extraction of
more reliable prefactors and we found that the derived kinetic parameters show weak dependence on
initial sample mass. Finally, by successful modeling of quasi-isothermal experimental curves, we show
that the discrete rate expression estimated from linear heating rate experiments enables modeling of
the thermal decomposition rate without any assumptions regarding the chemical process present. These
ndings can facilitate the design and optimization of industrial isothermal biomass fed reactors.
2014 Elsevier Ltd. All rights reserved.

1. Introduction
The threat of global warming and decreasing fossil fuel supplies
have paved the way for increased utilization of biomass as a feedstock for chemical production and thermochemical conversion
processes. For environmental and economic reasons, such processes need to be energy efcient and selective, something that
can be achieved partly through modeling of the reaction kinetics.
As cellulose is one of the main components in agricultural and
forestry biomass its pyrolysis kinetics has been systematically
studied for more than half a century [1]. In a majority of these
studies, the kinetic triplet, i.e. the activation energy, the preexponential factor and the conversion function (also called kinetic
model) has been completed through so-called model-tting
Corresponding author. Tel.: +46 70 759 54 34.
E-mail address: rosana@kth.se (R. Moriana).
http://dx.doi.org/10.1016/j.fuel.2014.11.079
0016-2361/ 2014 Elsevier Ltd. All rights reserved.

methods (refer to e.g. [2] for a summary of model-tting methods).


The rst step in applying such methods is usually to establish a
reaction scheme, and for each step in this scheme select a conversion function describing the reaction rates conversion dependency. Thereafter regression is used to derive an activation
energy and pre-exponential factor for each step in the reaction
scheme. Model-tting methods benet from their ability to model
the product distribution, since each step in the reaction scheme
lead to different products. A drawback of these methods is that,
if the reaction scheme, or any of the conversion functions used,
is faulty, the derived kinetic triplet will be biased.
Many different conversion functions have been used to model
cellulose pyrolysis kinetics during the years. Nevertheless, there
is still no general conclusion of what conversion function or reaction scheme that best describes the process. A review of previous
work in this area [1,3,4], reveals that models such as the single
step, rst order reaction model [59], the BroidoShazadeh

L.N. Samuelsson et al. / Fuel 143 (2015) 438447

model with modications [10] and nucleation models [1114] are


the most commonly applied. Recently a new conversion function,
based on random scission of the polymer chain, was suggested
for cellulose pyrolysis [15]. Additionally, some studies suggest that
depending on the heating rate [16] and the temperature range
[4,17], either one or another of these models is valid. As a result
of the use of different reaction schemes, conversion functions, as
well as the inuence from differences in samples [9] and experimental setups, the derived kinetic parameters reported in literature are widely scattered [1].
In 2011 the Kinetics Committee of the International
Confederation for Thermal Analysis and Calorimetry (ICTAC) [2]
published recommendations on how to estimate thermal decomposition kinetics of solids. One of their recommendations is to
use isoconversional methods to derive the apparent activation
energy as a function of conversion. Such relations give valuable
information about changes in the reaction kinetics during the
pyrolysis process and this insight can thus facilitate the choice of
a suitable conversion function. In recent years an increasing number of publications report on the use of isoconversional methods to
study cellulose pyrolysis kinetics [11,1315,1820]. In these studies, only the apparent activation energy is obtained in a strictly
model-free manner. The kinetic triplet is then completed through
a combination of model-tting and isocoversional methods.
In this study we explore the possibility to apply the principles of
isoconversional methods to derive a discrete rate expression without any model-tting tools, thus avoiding the problem of selecting
a conversion function. For this purpose we extend the isoconversional methods proposed in the literature to also extracting
an effective kinetic prefactor, dened as the product of the
pre-exponential factor and the conversion function. These values
together with the apparent activation energy provide an effective
rate expression that can be used to describe the thermal decomposition process at an arbitrary temperature program.
The two isoconversional methods adopted in this work are the
differential method originally proposed by Friedman [21] and the
incremental integral method introduced by Vyazovkin [22]. The
differential method was chosen because of its ease of implementation and since it is built on the least number of assumptions [2].
The incremental integral method was chosen in preference over
two classical linear integral methods, i.e. OzawaFlynnWall and
KissingerAkahiraSunose, since those methods both assume constant activation energy from time zero to the conversion of interest
[23]. This is not the case for the incremental method.
Similar strictly model-free isoconversional approaches were
used in the works by Cai and Chen [24] and imon et al. [25]. Cai
and Chen [24] investigated the combustion kinetics of pyrolysis
char by an isoconversional iterative linear integral method. This
particular method was used since the incremental integral method
by Vyazovkin was criticized for being computationally slow [26]. In
the present study, extra computational time was not an issue.
imon et al. [25] used an incremental integral method based on a
least-square optimization to obtain the kinetic parameters of NiS
recrystallization. This method, although easier to apply than the
method of Vyazovkin, is based on integration in the temperature
domain and it may thus introduce errors due to the explicit use
of the heating rate in the kinetic evaluation. This issue can easily
be avoided by using experimental data that follow the desired
heating rate.
The selected isoconversional methods and their extensions are
applied to pyrolysis data obtained from thorough thermogravimetric experiments on commercial microcrystalline cellulose. In the
thermogravimetric experiments, six different heating rates and
three different sample masses are used, allowing us to assess
the effect of thermal lag and initial sample mass on the derived
kinetics. The kinetic parameters derived from both the differential

439

and integral methods are compared with previously published


results and subject to error propagation analysis. Furthermore,
we highlight the advantages of the presented methodology by
comparing our results with the information obtained from generalized master plots [27]. The extracted kinetic data is nally validated by reconstructing quasi-isothermal conversion curves
measured in separate experiments.
2. Materials and methods
Sigmacell cellulose type 20 (Sigma Aldrich, particle size 20 lm,
average molecular weight 3640 kDa and degree of polymerization
about 200 according to Sigma Aldrich technical service) was used
as received for all the pyrolysis experiments.
Pyrolysis tests were performed on a Mettler Toledo TGA/DSC 1
STAR System (balance resolution 0.1 lg, horizontal balance setup,
balance signal noise below 2 lg, temperature accuracy 2 K). The
temperature scale of the instrument was calibrated using the melting points of the standards In, Zn and Al. The sample thermocouple,
measuring the temperature used for the kinetic analysis, was
located directly under and in contact with the sample crucible.
The cellulose powder was placed in a cylindrical 70 lL alumina
crucible (depth 3.5 mm) without lid and pyrolyzed from room
temperature to approximately 1023 K. Cellulose samples of three
different initial masses were used: 1.6 (5%), 4.0 (3%) and 9.9
(2%) mg, where number in brackets indicate variations among
repetitions. For the lowest mass the sample thickness in the crucible was less than 0.5 mm while for the highest mass it was about
3 mm. For each sample mass, a minimum of ve different heating
rates were employed: 3, 5, 7, 10 and 15 K/min. The two higher
sample masses were also pyrolyzed using 2 K/min. For each heating rate and mass, two to four runs were performed. Finally, to further investigate the presence of thermal lag, four runs at 15 K/min
and with a mass of 2.7 mg were performed.
In addition to the dynamic experiments, two quasi-isothermal
runs were performed. The end temperatures were 553 K and
573 K. A heating rate of 10 K/min was used to reach the isotherms.
An initial sample mass of 4 mg was used for both of the isothermal
runs. Only one run per temperature was performed.
The measurements were all corrected with a blank run preceding the sample run. All of the pyrolysis experiments were run at
atmospheric pressure in nitrogen (purity 99.999%) with a volumetric ow rate of 70 mL/min.
For data acquisition, the highest possible sampling rate was
used, resulting in approximately 29  103 to 54  103 data points
per run, depending on the length of the run. A high sampling rate
was used since kinetic evaluation of analytical data showed that
increasing the number of data points decreased the uncertainty
in the derived kinetic parameters. To minimize the numerical
manipulation of the collected data, the sampled mass loss data
was used for kinetic analysis without previous smoothing.
The measured mass loss is used to characterize the pyrolysis
process by means of the extent of conversion at, dened as:

at

m1  mt
;
m1  m2

where m1 is the mass of the dry sample, mt is the sample mass at


time t, and m2 is the mass of the char, i.e. the solid residue left after
pyrolysis. In the present work, m1 and m2 are taken as the sample
mass at 423 K and 1023 K, respectively. Since isoconversional methods are sensitive to the denition of conversion, the above boundaries were chosen to ensure that no signicant mass loss
processes were present at those temperatures (validated by varying
the lower temperature that determines m1 , see Fig. S1 in the
Supplementary Material).

440

L.N. Samuelsson et al. / Fuel 143 (2015) 438447

3. Isoconversional analysis
3.1. Isoconversional methods
Isoconversional methods are based on the isoconversional principle stating that, at a certain extent of conversion, the rate of
decomposition is uniquely determined by the current sample
temperature.
The single-step approximation is the simplest model following
this principle:

da
kTtf a
dt

where kTt is the temperature function, T is the sample


temperature at time t, and f a is the conversion function. The latter
is typically, but not necessarily [28], assumed to follow Arrhenius
law, with an apparent activation energy Ea :



Ea
;
kT A exp 
RT

where A is the apparent pre-exponential factor and R is the gas


constant.
Traditionally, isoconversional methods are used for determining the apparent activation energy as function of conversion without making any assumptions on the conversion function f a. In
this work, we extend the methods and additionally determine an
effective prefactor, i.e. the pre-exponential factor formed by the
product of the apparent pre-exponential factor and the conversion
function:

c0a Af a

presents a non-linear optimization problem with the two


unknowns c0a and Ea . In the original paper by Friedman [21], this
problem is transformed into a linear problem by simply taking
the logarithm of Eq. (5) which then allows for determining c0a and
Ea from a plot of ln ri;a versus 1=T i;a t. Although widely applied, this
linearization of Eq. (5) will affect the experimental error in r i;a and
has a tendency to overestimate the uncertainty in experiments with
a low heating rate (for which r i;a is small). This suggests to remain
with the non-linear problem as given in Eq. (5) and use a leastsquare solver for nding c0a and Ea .
The integral method is based on the assumption that over a
small increment of conversion the kinetics are governed by a single
reaction and accordingly c0a and Ea can be treated as constants. The
single-step approximation, Eq. (2), can then be applied for each
increment of conversion [28]. Denoting such an increment by Da,
integration of Eq. (2) gives:

Da c0a Ji Ea ;
where

J i Ea


r i;a c0a exp 


Ea
;
RT i;a t

where T i;a t is the temperature at which a conversion a is reached


in the i-th experiment. For a given conversion, each experiment provides a value for r i;a and T i;a t. Fitting Eq. (5) to these values

Fig. 1. Schematic of the evolution of conversion at two heating rates b1 and b2 , and
the quantities used in the isoconversional methods (see text for details).

t i;a

t i;aDa



Ea
dt;
exp 
RT i t

and T i t is the temperature evolution in the i-th experiment; the


integration limits indicate the times at which the conversions
a  Da and a, respectively, are reached in experiment i (see
Fig. 1). For a given conversion, Eq. (6) holds for any experiment,
which implies that for a given conversion the Ji s should be equal.
The optimization problem to nd Ea is thus to minimize the variation of the J i s. In the method proposed by Vyazovkin, this is done by
minimizing the following function:

As with Ea , the prefactor ca is extracted as a discrete value for a


given conversion. The set of these parameters provide a complete
kinetic description.
In this work, the differential method of Friedman and the incremental integral method of Vyazovkin are used. The differential
method is based on approximating the rate of conversion, i.e. the
derivate on the left hand side of Eq. (2), from the experimental
data. For a given conversion a, let us denote this approximated rate
by ri;a , where i denotes the index of the experiment (how to
numerically approximate the derivate da=dt is described below),
as illustrated in Fig. 1. Applying this together with Eqs. (4) and
(3), Eq. (2) reads as:

UEa

n X
n
X
J Ea
i

i1 ji

Jj Ea

where n is the number of experiments. Having found Ea from minimizing Eq. (8), we can recompute the J i s, which allows for extracting the prefactor c0a using:

E a i;
c0a Da=hJ i b

where hJi b
E a i is the average of Ji b
E a evaluated at the optimized
activation energy b
E a . Details on the implementation of both methods are given below.
3.2. Error estimation and conditioning
Estimation of pyrolysis kinetics is typically based on a large set
of experimental data where the extracted parameters represent
the best t of the underlying model to the data. To get an idea
of the quality of this best t and the adequacy of the proposed
model, next to visual inspection of the tted data, analysis of the
uncertainty of the estimated parameters is crucial. In this section,
we rst discuss how to estimate the uncertainty in the integral
method of Vyazovkin and thereafter we show how to condition
the problem in order to reduce the uncertainty. The latter strategy
is also applicable to the differential method, discussed at the end of
this section.
The uncertainty estimation for the apparent activation energy
derived with the integral method was addressed in a paper by
Vyazovkin and Wight [29]. Although the method applied in this
work is different from that in [29], both methods use the same
starting criteria. This starting criteria, i.e. the isoconversional principle, states that for a given conversion interval, independent on
heating rate, the time integral J i should be equal for all experimental curves. The variation in the J i s, evaluated with the optimized
E a , can therefore be seen as the error (or
activation energy, b

L.N. Samuelsson et al. / Fuel 143 (2015) 438447

residual) of the minimization problem. We dene this error as (in


E a ):
the following J i  J i b

dJ i J i  hJ i i:

10

The error dJ i is related to the error in the activation energy, dEa ,


through [30]:

dJ i X i dEa ;

11

where X i @Ji =@Ea b is the derivate of Eq. (7) with respect to Ea ,


Ea
evaluated at the optimized activation energy. From Eq. (11), the
error in the activation energy follows as [31]:

dEa

q
1
dJ 2 X T X

12

where dJ2 is the variance of the dJi s, and X is the vector containing
X i . From the error in the activation energy given in Eq. (12) we can
proceed and deduce the error in the effective prefactor. The error in
b
E a propagates to the estimation of the prefactor c0a , whose error is:

dc0a

Da
hJ i i2

dJ:

13

This expression deserves a closer look. Indeed, Eq. (13) gives an


error dc0a that is of similar magnitude as the estimated parameter c0a
(cf. Eq. (9)). This is readily seen from an order of magnitude estimate of the quantities entering Eq. (13): Writing Eq. (11) as
dJ  jX jdEa with jX j  J=RT a , where T a is the characteristic temperature at which conversion a is reached, we can use Eqs. (9) and (13)
to obtain:

dc0a dEa

;
c0a
RT a

14

which for typical values of dEa and T a is O1.


This large error in the estimation of the prefactor c0a presents a
severe limitation of our methodology, and Eq. (9) actually may lead
to meaningless results. To overcome this difculty, we introduce a
preconditioning of the kinetic model by substituting the original
Arrhenius expression (Eq. (3)) with a conditioned form,




Ea 1 1
;
kT kT 0 exp 

R T T0

15

where kT 0 is the rate constant at a reference temperature T 0 chosen within the range of interest. In the use of kinetic data, kT 0
takes on a similar role as the pre-exponential factor A while the
treatment of its uncertainty is substantially simplied. To illustrate
this, let us repeat the derivation from above by using Eq. (15)
instead of Eq. (3):
Substituting Eq. (15) into Eq. (2) and repeating some of the previous steps will lead to slightly different expressions. In particular,
instead of Eq. (6), the starting equation for the integral method
becomes:

!
b
Ea
Da ca exp
J;
RT 0 i

16

where J i is given in Eq. (7), and

ca kT 0 f a;

17

is the effective prefactor of the conditioned problem. The latter is


estimated in a similar way as c0a :

!
b
Ea
Da
:
ca
exp 
hJ i i
RT 0

18

441

where dEa is the error in the activation energy given by Eq. (12).
Eq. (19) is the actual relation hereafter used for estimating the error
in ca in the integral method. Applying the same order of magnitude
approximation that leads to Eq. (14), the above simplies to:



dca dEa 1
1
:


ca
R Ta T0

20

Hence, the error in the effective prefactor ca is reduced by dEa =RT 0


with respect to the original prefactor c0a (see Eq. (14)). This difference makes ca more suitable to study and its determination more
robust, while the kinetic information it carries is essentially the
same as the one in the original prefactor.
Correspondingly for the differential method, instead of Eq. (5),
the governing equation becomes




Ea 1
1
ri;a ca exp 

;
R T i;a T 0

21

which is equivalently solved using a least-square solver, which provides estimated values for the two unknowns ca and Ea . The errors
in these two parameters are obtained from the square root of the
diagonal elements of the covariance matrix [31].
3.3. Reconstruction of at
The extracted kinetic data is validated by reconstructing the
extent of conversion from the model equations. Reconstruction of
at is readily done by integrating Eq. (2) with the constituting
functions ca and Ea taken as piecewise constant functions, as determined by kinetic analysis. The initial condition to Eq. (2) is taken as
the rst value of a for which the kinetic analysis was performed
minus the increment Da used in the integral method, and the time
at which this conversion is achieved. For reconstructing runs with
a linear heating rate b, the temperature in Eq. (2) is simply
Tt T t0 bt, while for isothermal runs Tt T iso .
3.4. Generalized master plots
Master plots are applied in this study to enable an independent
comparison of the methodology presented in this work. Generalized master plots are chosen over the traditional master plots
since, in the case of non-isothermal data, the former enable us to
select a suitable conversion function from one single experimental
curve [27].
The criterion to use the generalized master plots is the same as
for traditional master plots, i.e. we assume that one single conversion function can describe the whole process [2]. This means that
we have to select one apparent activation energy, E0 , to represent
the whole conversion process. If the derived apparent kinetic
parameters show a signicant dependence on conversion then
master plots are not applicable [2].
Two different types of normalized generalized master plots are
used in this work, i.e. the differential form and the integral form
[27]. These two plots complement each other by creating distinguishable curves in the range a < 0:5 and a > 0:5, respectively.
The experimental master plots are compared against three theoretical conversion functions: the rst order model (F1) [27], a random
nucleation and growth model (Am, m = 2) [27], as well as the random scission model (L2), published recently [15]. All the applied
equations related to master plots are found in the Supplementary
Material.

The error in ca is obtained from the propagation of the error in the


activation energy:

3.5. Implementation



dca
hX i i
1
;
dEa 

hJ i i RT 0
ca

The numerical routines were implemented in Python, using the


modules Scipy and Numpy [32]. In the differential method, the

19

L.N. Samuelsson et al. / Fuel 143 (2015) 438447

4.1. Thermal analysis


Let us rst discuss the mass loss curves obtained from thermogravimetric experiments at a linear heating rate. Fig. 2 shows arbitrarily selected mass loss curves at two heating rates for each
initial sample mass tested. From this gure it is obvious that the
decomposition takes place in a relatively narrow temperature
range: For a heating rate of 3 K/min the decomposition takes place
between 520 and 620 K, while for higher heating rates the temperature range is shifted to higher temperatures. This shift in temperature with increasing heating rate is a direct consequence of the
reaction kinetics, which is readily seen by integrating Eq. (2) with
T T 0 bt and an arbitrary but nite activation energy.
Looking at the inuence of sample mass at a xed heating rate, a
slight increase of the decomposition temperature with increasing
sample mass can be observed. To investigate this further, in
Fig. 3 we plot the peak temperature T peak as a function of heating
rate b. The peak temperature is dened as the temperature at
which the mass loss rate is at a maximum (the inset of Fig. 3 shows
the mass loss rate, dened as the derivative, with respect to temperature, of the mass loss curves; the derivative shows one peak).
The peak temperature generally increases with increasing heating

1.0
= 15 Kmin1

m(T )/m1

0.8
0.6
= 3 Kmin1

0.2
0.0

500

550

615
610
605
600
595
590
585
580

3 Kmin1
15 Kmin1

T /K
2

10

12

14

16

/Kmin1
Fig. 3. Mean peak temperatures plotted as a function of heating rate for sample
mass 1.6 mg (squares), 4.0 mg (circles) and 9.9 mg (triangles). Bars indicate
standard deviation in peak temperatures. Inset: Mass loss rate as a function of
sample temperature. Rate curves are chosen arbitrary among the repeated runs. A
SavitskyGolay lter was used to smooth the rate curves to increase visibility.

4. Results and discussion

0.4

620

dm
/mgK1
dT

derivate da=dt was approximated by central differences, and


scipy.optimize.curve_t was used for solving the least square
problem. In both the differential and integral method,
scipy.interpolate.interp1d was used for nding t i;a and t i;aDa . In
the integral method, the integration of the J i s was performed
numerically using trapezoidal method. The minimization of the
U-function was done using scipy.brent. An increment of
Da 0:04 was used; using an increment of 0.02 or 0.08 did not signicantly change the results. The analysis was performed for 21
values of conversion, distributed uniformly in between 0.10 and
0.94. The reference temperature was chosen as T 0 600 K.
Master plots were created by taking the time derivative of the
experimental conversion data. The resulting curves were smoothed
with a SavitskyGolay lter before generating the master plots.
The program written to perform the numerical operations was
veried by simulated data with known activation energies, preexponential factors and conversion functions.

Tpeak /K

442

600

650

700

750

800

T /K
Fig. 2. Normalized sample mass as a function of sample temperature for two of the
used heating rates. Three different initial sample masses, m1 , (from bottom to top):
1.6 (blue), 4.0 (red) and 9.9 mg (green) are plotted. The plotted curves were chosen
arbitrary among the repeated runs. (For interpretation of the references to color in
this gure legend, the reader is referred to the web version of this article.)

rate and sample mass, apart from some variations at lower heating
rates: For a heating rate of 3 K/min, the difference in peak temperature between the highest and lowest mass is ca. 2 K, while at
15 K/min, the difference is ca. 7 K. This increase reects the thermal
lag, i.e. the delay in temperature due to nite heat capacity and
heat (and mass) transfer limitations in the sample [9]. A common
strategy to avoid such effects is to reduce the sample mass until
the peak temperature does not vary with initial sample mass.
Reducing the sample mass is meaningful only as long as the noise
in the recorded signal is low enough to allow the reduced sample
mass to be measured accurately. In this study, a balance noise of
2 lg hindered us from decreasing the initial sample mass below
1.6 mg. For this initial sample mass the nal mass was 1664 lg
(14% in Fig. 4a), resulting in a maximum noise in the signal of
10%. This must be considered an upper limit of what we can
accept.
The presence of thermal lag was investigated through a series of
additional tests at 15 K/min, including a fourth sample mass of
2.7 mg, see Fig. S2 in the Supplementary Material. These tests support the trend seen in Fig. 3, namely the peak temperatures
increase with increased initial sample mass for all investigated
masses. This indicates that thermal lag cannot be ruled out completely for the lowest sample mass. Note also that, due to the
low temperature accuracy of the instrument, the small differences
in thermal lag resulting from similar masses are difcult to measure accurately.
Returning our attention to Fig. 2, we notice that the nal masses
increase with increased initial sample mass. This behavior, also
reported in earlier studies [9,12,13,33], may hint to a secondary
solidvapor reaction, resulting in increased char yield [4,12].
Secondary solidvapor reactions are favored by increased sample
thickness, which increases the residence time of the vapors in
the solid matrix, giving molecules increased possibility to react.
Fig. 4 shows the char yields in each of our experiments, plotted
versus the initial sample mass, and versus heating rate. A clear correlation between end mass and sample mass is revealed, Fig. 4a,
while the data shows no consistent trend between end mass and
heating rate among the whole dataset, Fig. 4b: the discrepancy in
char yield for a given heating rate is in some cases larger than
the difference observed between different heating rates. These
observations are in agreement with the results of Vrhegyi et al.
[34], justifying the use of isoconversional analysis at these process
conditions, since kinetics is not a function of the temperature

L.N. Samuelsson et al. / Fuel 143 (2015) 438447

A comparison of our char yields with those reported in Table 1


reveals good agreement with the values reported by Cabrales and
Abidi [19]. Signicantly higher values were reported by Miranda
et al. [36] and Capart et al. [13]. The value reported by the latter
study is surprisingly high, considering the high volumetric gas ow
rate used. A high gas ow rate should increase the escape rate of
volatiles evolved from the sample and thus decrease the char yield.
The discrepancies in reported char yields may result from differences in sample volume (vapor residence time), ash amounts or
sample morphology.

m2 /m1 100

6
5
4
2 K/min
3 K/min
5 K/min
7 K/min
10 K/min
15 K/min

3
2

4.2. Kinetic analysis

1
0

10

m1 /mg

(a)
8
7

m2 /m1 100

6
5
4
3
2
1
0

10

12

443

14

16

/Kmin1

(b)
Fig. 4. Char yield, as a function of initial dry mass m1 (a) and heating rate b (b). In
(a) different heating rates are indicated by different markers. In (b) initial sample
mass 1.6 mg (squares), 4.0 mg (circles) and 9.9 mg (triangles).

program [35]. From Fig. 4a, we also notice that the experiments
with the lowest sample mass are subject to stronger variations in
char yield compared to the experiments performed with higher
masses. This behavior may be a result of the balance resolution
or inhomogeneities in the cellulose sample, both of probable relevance at such low sample masses.
Table 1 compares the peak temperatures shown in Fig. 3 with
peak temperatures reported in the literature, including also cellulose type, char yields m2 =m1 , and some relevant experimental
conditions (initial sample mass m1 , heating rate b, end temperature
T 2 and volumetric ow rate m_ ). References included in Table 1 were
selected based on two criteria: (i) similarity to our study and (ii)
availability of data. The peak temperature reported by Grnli
et al. [6] is very similar to those found in this study. In the four
other studies the peak temperatures are signicantly higher,
despite the similar sample mass and heating rates. This discrepancy could be due to thermal gradients (internal or external) or
more thermally stable samples due to e.g. larger particle size
(which would decrease the reactivity in a nucleation type process
[14]), higher crystallinity or higher degree of polymerization [37].
Note also that the peak temperature and char yield, reported by
Grnli et al. [6], are mean values derived from a round-robin study
including eight laboratories.

4.2.1. Apparent activation energy


Analysis of the mass loss curves, presented above, revealed the
presence of thermal effects, becoming important with increased
sample mass and heating rate. This suggests limiting the kinetic
analysis to the lowest sample mass. However, as presented below,
the data for the lowest sample mass shows some effects that
makes us question the robustness of the data. We therefore performed the kinetic analysis for all three sample masses, which
allows assessing the inuence of experimental noise, initial sample
mass and thermal lag on the kinetic analysis.
Fig. 5 shows the apparent activation energy as a function of conversion derived from the lowest (square symbols), the intermediate (circle symbols) and the highest (triangle symbols) sample
mass. For all sample masses, the differential method (full symbols)
leads to similar values as the integral method (open symbols).
Interestingly though, the apparent activation energy from the
low sample mass is ca. 30 kJ/mol higher than that for the two
higher sample masses: For the lowest sample mass, Ea varies
between ca. 210 and 240 kJ/mol, while for the intermediate and
high sample mass, Ea varies between ca. 170 and 200 kJ/mol in
the conversion range 0.10.8. Moreover, the errors in Ea derived
from the low sample mass are considerably larger than those of
the intermediate and high sample mass. Also, the errors derived
for the differential method are signicantly larger than for the integral method (whose error bars in Fig. 5 are often covered by the
markers). The latter observation might be partly due to the differences in the two methods used for estimating the error; in the latter method the error is based on the variation of the embedded
integral terms (see Eq. (12)).
The difference in curve magnitude with varying sample mass,
seen in Fig. 5, requires an explanation. It is unlikely that thermal
lag is the cause: If the difference in magnitude was caused by thermal lag one would expect a trend of decreasing activation energies
with increasing sample mass. However, as seen in Fig. 5 the two
higher sample masses result in similar activation energies, which
makes this possibility improbable for explaining the observed
behavior. Another reason might be the quality of the underlaying
data, as suggested by the large error bars for the low sample mass.
To test the reliability of all datasets we performed the kinetic analysis on (i) half of the experimental data i.e. excluding half of the
repetitions for each heating rate, (ii) all data excluding the heating
rate 15 K/min and (iii) all data excluding the heating rates 10 and
15 K/min. For the intermediate and high sample mass the three
reduced datasets do not signicantly change the derived activation
energies, i.e. they stay in between ca. 170 and 200 kJ/mol, except at
the highest conversion. On the contrary, removing the two highest
heating rates from the lowest sample mass led to an average
increase of 10 kJ/mol, while removing only 15 K/min decreased
the activation energy by ca. 20 kJ/mol. Analyzing only half of the
dataset had no signicant inuence. This implies that the kinetics
derived from the low sample mass data is sensitive to the high
heating rate data, probably due to discrepancies in nal char yield.
Consequently, the results derived from the low sample mass data

444

L.N. Samuelsson et al. / Fuel 143 (2015) 438447

Table 1
Mean peak temperatures and char yields, both with standard deviations, for pyrolysis of microcrystalline cellulose. Values in parenthesis correspond to heating rates in
parenthesis.
Ref.

Material

m1 /mg

m2
m1

T peak /K

 100

T 2 /K

m_ /mL min1

b/K min1
2 (2.1)

3 (4)

7 (8)

10

15 (16)

This

Sigmacell

1.6 5 %
4.0 3 %
9.9 2 %

585 0.4
586 0.9

591 0.3
590 0.1
593 0.6

597 0.1
598 0.2
601 0.4

601 0.1
603 0.3
606 0.0

607 0.5
610 0.9
611 0.3

612 0.1
616 0.7
619 0.5

2.6 0.8
4.2 0.6
6.8 0.3

1023
1023
1023

70

[19]
[6]
[36]
[13]
[7]

Avicel PH 105
Avicel PH 105
Quimsul cellulose
Whatman CC31
Acros cellulose

23
4.1 1.3
5
5.84
69

(600)

(603 0.3)

607 0.9
600 5
622

(617 0.2)

621 0.2

(629 0.2)

624

3.024.05
7.2 2.4
11.8
14.21
5.1

873

873
973
1073

20

100
1500
100

280
260

E /kJmol1

240
220
200
180
160
140
120
100

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

0.9

Fig. 5. Apparent activation energy as a function of conversion derived from


experiments with sample mass 1.6 mg (squares), 4.0 mg (circles) and 9.9 mg
(triangles). Closed symbols: differential method, open symbols: integral method.
Error bars are partly covered by the markers.

have to be interpreted with care, i.e. we have to assume that the


error is at least twice as large as indicated by the error bars in
Fig. 5. Additionally, it shows, for the two kinetic analysis methods
used here, results cannot be considered reliable only based on the
criteria that both methods give rise to similar results.
Regarding the individual datasets, shown in Fig. 5, for the lowest sample mass the apparent activation energy initially decreases
and reaches a plateau, after which it increases at high conversion.
For the intermediate sample mass the trend is different: At low
conversion a plateau around 180 kJ/mol is visible, while for
a > 0:4 the kinetic values are slowly increasing. A slightly different
behavior is observed for data derived from the highest sample
mass. Here we see a weak initial decrease, followed by a minimum
at a  0:5. At high conversion, a > 0:8, the values rapidly decrease.
Table 2
Apparent activation energies for cellulose pyrolysis at intermediate conversion. The
lower uncertainties in values derived with the incremental integral method stem
from the different method used for its estimation.

Ea0:5 /kJ mol1

m1 /mg

Kinetic analysis method

1.6 5%

Non-linear differential method


Incremental integral method

223 12a
216 0.25a

4.0 3%

Non-linear differential method


Incremental integral method

178 3.0
183 0.12

9.9 2%

Non-linear differential method


Incremental integral method

172.5 1.8
178.4 0.12

Errors might be substantially larger (see text).

The difference to the intermediate sample mass kinetics are


marginal, at least for conversions a < 0:8. At this moment it is
not possible to assign what chemical or physical processes that
cause these curve characteristics and the different magnitudes of
the derived apparent activation energies. However, it is likely that
the differences stem from differences in the pyrolytic processes,
such as the secondary solidvapor reactions which, as shown
above, depend on the sample mass.
In Table 2 we present apparent activation energies derived at
a 0:5 for each sample mass used. A comparison of these values
with similar experimental studies, applying isoconversional and
model-tting methods based on single-step models, reveals that
our values are in reasonable agreement with values found in the literature. Literature reports activation energies in the range between
180 and 250 kJ/mol [710,13,15,19,20,36], which well covers the
values found in this work. Note that, when possible, this comparison, between our values and those reported in literature, was based
on activation energies derived at intermediate conversion.
4.2.2. Effective prefactor
Next, we turn to the effective prefactor. At rst, we considered
the unconditioned problem, based on the common form of the
Arrhenius law (Eq. (3)). This lead to the prefactor c0a , given by
Eq. (9), shown in Figs. S3 and S4 in the Supplementary Material.
From these gures it is apparent that c0a is subject to strong uctuations. Adding error bars to the gures (Figs. S3b and S4b) reveals
the large uncertainty of the derived data, which signicantly renders its robustness and no conclusion can be drawn from the data.
This motivated us to reformulate the problem and use the condition form of kT given by Eq. (15) which, as shown next, enabled
us to derive more robust data with reasonably small uncertainty.
Fig. 6 shows the conditioned prefactor ca as a function of conversion. Also here, we show data derived from all three sample
masses using both the differential and integral method. As for
the discussion of the apparent activation energy trends in Fig. 5,
the different curve shapes in Fig. 6 could indicate the presence of
different pyrolytic processes. Being aware of the fact that the prefactor is not derived independently of the apparent activation
energy (cf. Eqs. (18) and (21)), the data for the lowest sample mass
in Fig. 6 is subject to the same concerns as those stated for the
apparent activation energy, i.e. the errors might be larger than
those shown in Fig. 6. For the lowest sample mass, the effective
prefactor for both methods exhibit a monotonous decay over the
whole range of analyzed conversions. After a sharp decrease at
low conversions, from a  0:2 onwards the decay evolves linearly
up to the highest conversion analyzed.
As for the apparent activation energy, the effective prefactors
derived from the intermediate and high sample masses show a different behavior. While at a > 0:4 a linear decay is seen also in these
cases, for low conversion, ca describes a plateau at signicantly
lower values compared to the prefactor derived from the lowest

445

0.5

2.5

0.4

2.0

f ()/f (0.5)

c /min1

L.N. Samuelsson et al. / Fuel 143 (2015) 438447

0.3
0.2
0.1
0.0

1.5
1.0
0.5

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

0.0

0.9

0.2

0.8

(a)

g()/g(0.5)

4.2.3. Master plots


In Fig. 6 we see a pronounced difference between the effective
prefactors derived from the low sample mass, compared to those
derived from the two higher sample masses. The effective prefactor
ca is dened as the product of the conversion function and the
exponential of the activation energy, (cf. Eq. (17)). Noticing the
small variations of the apparent activation energy within
0.2 < a < 0.8 we argue that the differences seen in Fig. 6 are mainly
caused by a difference in the conversion function. To explore this
possibility, we analyzed our data with an independent method,
namely generalized master plots. Master plots were constructed
for each initial sample mass and a heating rate of 3 K/min, following the procedure of Gotor et al. [27]. The resulting differential and
integral master plots are shown in Fig. 7a and b, respectively.
To create the experimental master curves (solid lines) one
apparent activation energy per sample mass was chosen, based
on the data in Table 2. From Fig. 7a, we see that the experimental
curve derived from the low sample mass data is close to a rst
order model (dashed line), while the two curves derived from
higher sample masses are close to a random scission model (dotted
line) (the shape of the experimental master plots differed slightly
as a function of the chosen apparent activation energy, see
Fig. S5 in the Supplementary Material). On the contrary, from the
integral master plot, Fig. 7b, we see that at a > 0:5 all of the experimental curves agree best with the F1 model. These results imply
that at low conversion different reaction mechanisms govern the
process, depending on the initial sample mass, while at high conversion the pyrolysis can be approximated as a single, rst order
reaction. Such dependence of the kinetics on the initial sample
mass suggests a more complicated reaction mechanism than proposed by single conversion function models.
If the conversion function is of interest there is a means to compare generalized master plots with plots of the effective prefactor.
This comparison is straight forward for processes with a constant
apparent activation energy. However, for processes with varying
apparent activation energies, generalized master plots with varying E0 should be derived to check the reliability of the resulting
proles (see Fig. S5 in the Supplementary Material).
Finally, we would like to stress that by using the methodology
presented in this work, there is no need to determine the

0.6

Fig. 6. Effective prefactor as a function of conversion derived from experiments


with sample mass 1.6 mg (squares), 4.0 mg (circles) and 9.9 mg (triangles). Closed
symbols: differential method, open symbols: integral method. Reference temperature, T 0 600 K. Error bars are partly covered by the markers.

sample mass. This plateau is more pronounced for the intermediate sample mass than for the highest mass, where the trend is
smoother.

0.4

0
0.0

0.2

0.4

0.6

0.8

1.0

(b)
Fig. 7. Normalized experimental and theoretical master plots based on (a) the
differential form of the generalized kinetic equation and (b) the integral form of the
generalized kinetic equation. Solid lines: experimental curves at 3 K/min and
1.6 mg (blue), 4.0 mg (red) and 9.9 mg (green). Theoretical curves are F1 (dashed),
L2 (dotted) and A2 (dash-dotted). E0 : 219 kJ/mol (1.6 mg), 180 kJ/mol (4.0 mg) and
175 kJ/mol (9.9 mg). (For interpretation of the references to color in this gure
legend, the reader is referred to the web version of this article.)

conversion function (or functions) as the kinetic information it carries is contained within the effective prefactor, obtained in discrete
form. This is illustrated in the next section.
4.3. Reconstructed and modeled conversion under dynamic and quasiisothermal conditions
To validate the derived kinetic parameters, the evolution of at
under both dynamic and isothermal conditions was reconstructed
and compared to experimental data. Fig. 8 shows reconstructed
dynamic experiments for three heating rates. The reconstructed
curves are based on kinetic data obtained from the intermediate
sample mass and compared to experiments with the same initial
sample mass. The agreement is generally good and both of the
kinetic methods applied in this work lead to satisfactory results.
Similar results were found when reconstructing data for the low
and high sample masses; the low mass gave slightly worse t
while the agreement was somewhat better for the highest sample
mass (data not shown).
Fig. 9 shows the quasi-isothermal extent of conversion (lines)
together with the modeled conversion (symbols) using the kinetic

446

L.N. Samuelsson et al. / Fuel 143 (2015) 438447

data obtained from dynamic experiments. The experimental conversion is obtained by using the nal mass from the mass loss
curves shown in the inset of Fig. 9. From this inset, it is visible that
the end mass for the high temperature isotherm levels out at 9%. For
the low temperature isotherm the sample mass exhibits a relatively
slow decay, not reaching a plateau within the duration of the experiment. This leaves us with some uncertainty regarding the end mass
that would be achieved for the given isotherm and sufciently long
time. By visual inspection we estimate the end mass for the low isotherm to be 15%. Additionally, the well-known issue with isothermal experiments arose, i.e. non-negligible conversion occurred
during the dynamic segment preceding the isothermal stage. For
the 553-K isotherm, a conversion of 5% was achieved before the isotherm was reached, while for the 573-K isotherm 18% conversion
was reached. In the modeling this non-isothermal step was ignored.
Despite this, in general, the t is satisfactory.
For both of the isotherms, the modeled curves based on intermediate (circles) and the high (triangles) sample mass data, show
a clearly better t compared to kinetic data obtained from the

lowest sample mass (squares). Considering the large uncertainty


in the kinetic data derived from the low sample mass, this result
is not surprising.
To further strengthen these results, the modeled isothermal
curves in Fig. 9 were differentiated and the resulting da=dt curves
were plotted as a function of conversion. The result, plotted together
with the quasi-isothermal experimental curves, is shown in Fig. S6 in
the Supplementary Material. This gure conrms the good agreement, seen in Fig. 9, between the modeled and experimental data.
Additionally, the new plot reveals that the curve derived from the
4.0 mg data shows the best t to the isothermal curves, as expected.
The agreement of the modeled conversion, derived from dynamic
heating rate experiments, with the quasi-isothermal experiments is
a clear indication of the feasibility of this methodology to derive an
effective rate expression using isoconversional methods. Indeed,
both the differential and integral method provide model-free kinetic
parameters that can be used to describe the pyrolysis process of the
microcrystalline cellulose studied in this work.
5. Conclusions

1.0

0.8

0.6

0.4

0.2

0.0

520

540

560

580

600

620

640

660

680

T /K
Fig. 8. Reconstructed conversion for dynamic experiments at heating rates (from
left to right): 3, 7 and 15 K/min. Solid lines: experimental 4.0 mg sample mass;
symbols: reconstructed data, differential (closed symbols) and integral method
(open symbols). Experimental curves were chosen arbitrary among the repeated
runs.

573 K
553 K

1.0
0.8

1.0
m(t)/m1

0.6
0.4
0.2

0.8
0.6

553 K

0.4

573 K

0.2
0.0

50

100 150 200


t/min

0.0
0

100

200

300

400

500

 The differential method by Friedman, used in its non-linear


form, and the incremental integral method by Vyazovkin, integrated in time domain, were extended to derive an effective,
model-free rate expression for the pyrolytic decomposition of
microcrystalline cellulose, measured in thermogravimetric
experiments at atmospheric pressure. For this system, the integral and the differential method lead to similar kinetic results.
 The proposed model-free isoconversional methodology was
used to derive a rate expression from dynamic data that can
model conversion proles of quasi-isothermal experiments.
These ndings can facilitate the development of improved modeling tools, useful in the design and optimization work of industrial isothermal biomass-fed reactors.
 For a conversion range of 0.1 < a < 0.8, experiments with an initial sample mass of 4.0 mg and 9.9 mg gave apparent activation
energies of 170200 kJ/mol. For a sample mass of 1.6 mg the
derived apparent activation energy varied signicantly as a
function of the input data and no reliable values could be
extracted.
 Both methods were sensitive to discrepancies in the experimental data, especially deviations in nal conversion. We therefore
recommend performing kinetic analysis on reduced amount of
data as part of the kinetic evaluation work, since it will help elucidate the reliability of the derived parameters.
 By using a conditioned temperature function, within which the
prefactor is replaced by the rate constant at a reference temperature, we could estimate an effective prefactor with signicantly increased reliability.
 Thermal lag was present, as shown from dynamic thermogravimetric experiments at various sample masses.
 The normalized nal mass was found to increase with increasing sample mass, independent of the heating rate. This could
indicate secondary vaporsolid reactions that become relevant
as the sample layer thickness in the crucible increases. Due to
this inuence of sample dimension, we suggest that future
studies of cellulose pyrolysis kinetics report both the sample
mass and the sample thickness used.

t/min
Fig. 9. Conversion for quasi-isothermal curves as a function of time. Solid thick
lines: experiments. Symbols: modeled data with kinetic parameters derived from
sample mass 1.6 mg (squares), 4.0 mg (circles) and 9.9 mg (triangles). Closed
symbols: differential method, open symbols: integral method. Inset: experimental
mass loss curves for the two isotherms. In both panels, time zero is at 10%
conversion.

Acknowledgements
This work was carried out within the Swedish Gasication
Centre consortium. Funding from the Swedish Energy Agency and
the academic and industrial partners is gratefully acknowledged.

L.N. Samuelsson et al. / Fuel 143 (2015) 438447

R.M. acknowledges nancial support by the Wallenberg and


Lars-Erik Thunholm Foundation.
L.N.S. would like to thank Rasmus Karlsson and Tommy Zavalis
for their input on the numerics, Peter Gille for his help regarding
time optimization of the code used and nally Tibor Dubaj and Sergey Vyazovkin for their generous email correspondence regarding
isoconversional analysis.
Finally, the authors would like to thank the reviewers selected
by Fuel for their constructive critique.

Appendix A. Supplementary material


Supplementary data associated with this article can be found, in
the online version, at http://dx.doi.org/10.1016/j.fuel.2014.11.079.

References
[1] Serbanescu C. Kinetic analysis of cellulose pyrolysis: a short review. Chem Pap
2014;68:84760.
[2] Vyazovkin S, Burnham AK, Criado JM, Prez-Maqueda LA, Popescu C,
Sbirrazzuoli N. ICTAC kinetics committee recommendations for performing
kinetic computations on thermal analysis data. Thermochim Acta 2011;520:
119.
[3] Ld J. Cellulose pyrolysis kinetics: an historical review on the existence and
role of intermediate active cellulose. J Anal Appl Pyrol 2012;94:1732.
[4] Blasi CD. Modeling chemical and physical processes of wood and biomass
pyrolysis. Prog Energy Combust Sci 2008;34:4790.
[5] Fisher T, Hajaligol M, Waymack B, Kellogg D. Pyrolysis behavior and kinetics of
biomass derived materials. J Anal Appl Pyrol 2002;62:33149.
[6] Grnli M, Antal MJ, Vrhegyi G. A round-robin study of cellulose pyrolysis
kinetics by thermogravimetry. Ind Eng Chem Res 1999;38:223844.
[7] Lin YC, Cho J, Tompsett GA, Westmoreland PR, Huber GW. Kinetics and
mechanism of cellulose pyrolysis. J Phys Chem C 2009;113:20097107.
[8] Shaik SM, Koh C, Sharratt PN, Tan RB. Inuence of acids and alkalis on
transglycosylation and beta-elimination pathway kinetics during cellulose
pyrolysis. Thermochim Acta 2013;566:19.
[9] Antal MJ, Vrhegyi G, Jakab E. Cellulose pyrolysis kinetics: revisited. Ind Eng
Chem Res 1998;37:126775.
[10] Conesa JA, Caballero J, Marcilla A, Font R. Analysis of different kinetic models in
the dynamic pyrolysis of cellulose. Thermochim Acta 1995;254:17592.
[11] Reynolds JG, Burnham AK. Pyrolysis decomposition kinetics of cellulose-based
materials by constant heating rate micropyrolysis. Energy Fuels 1997;11:
8897.
[12] Vlker S, Rieckmann T. Thermokinetic investigation of cellulose pyrolysisimpact of initial and nal mass on kinetic results. J Anal Appl Pyrol 2002;62:
16577.
[13] Capart R, Khezami L, Burnham AK. Assessment of various kinetic models for
the pyrolysis of a microgranular cellulose. Thermochim Acta 2004;417:7989.
[14] Mamleev V, Bourbigot S, Yvon J. Kinetic analysis of the thermal decomposition
of cellulose: the main step of mass loss. J Anal Appl Pyrol 2007;80:15165.

447

[15] Snchez-Jimnez PE, Prez-Maqueda LA, Perejn A, Criado JM. Generalized


master plots as a straightforward approach for determining the kinetic model:
the case of cellulose pyrolysis. Thermochim Acta 2013;552:549.
[16] Banyasz JL, Li S, Lyons-Hart JL, Shafer KH. Cellulose pyrolysis: the kinetics of
hydroxyacetaldehyde evolution. J Anal Appl Pyrol 2001;57:22348.
[17] Milosavljevic I, Suuberg EM. Cellulose thermal decomposition kinetics: global
mass loss kinetics. Ind Eng Chem Res 1995;34:108191.
[18] Mamleev V, Bourbigot S, Bras ML, Yvon J, Lefebvre J. Model-free method for
evaluation of activation energies in modulated thermogravimetry and analysis
of cellulose decomposition. Chem Eng Sci 2006;61:127692.
[19] Cabrales L, Abidi N. On the thermal degradation of cellulose in cotton bers. J
Therm Anal Calorim 2010;102:48591.
[20] Moriana R, Vilaplana F, Karlsson S, Ribes A. Correlation of chemical, structural
and thermal properties of natural bres for their sustainable exploitation.
Carbohydr Polym 2014;112:42231.
[21] Friedman HL. Kinetics of thermal degradation of char-forming plastics from
thermogravimetry. Application to a phenolic plastic. J Polym Sci Pol Sym
1964;6:18395.
[22] Vyazovkin S. Modication of the integral isoconversional method to account
for variation in the activation energy. J Comput Chem 2001;22:17883.
[23] Burnham AK, Dinh LN. A comparison of isoconversional and model-tting
approaches to kinetic parameter estimation and application predictions. J
Therm Anal Calorim 2007;89:47990.
[24] Cai J, Chen Y. Iterative linear integral isoconversional method: theory and
application. Bioresour Technol 2012;103:30912.
[25] imon P, Thomas PS, Okuliar J, Ray AS. An incremental integral isoconversional
method. J Therm Anal Calorim 2003;72:86774.
[26] Budrugeac P. An iterative model-free method to determine the activation
energy of non-isothermal heterogeneous processes. Thermochim Acta
2010;511:816.
[27] Gotor FJ, Criado JM, Malek J, Koga N. Kinetic analysis of solid-state reactions:
the universality of master plots for analyzing isothermal and nonisothermal
experiments. J Phys Chem A 2000;104:1077782.
[28] imon P. Considerations on the single-step kinetics approximation. J Therm
Anal Calorim 2005;82:6517.
[29] Vyazovkin S, Wight CA. Estimating realistic condence intervals for the
activation energy determined from thermoanalytical measurements. Anal
Chem 2000;72:31715.
[30] Taylor JR. An introduction to error analysis. 2 ed. Sausalito
California: University Science Books; 1997.
[31] Montgomery D, Runger G. Applied statistics and probability for engineers. 5
ed. John Wiley & Sons; 2010.
[32] Oliphant TE. Python for scientic computing. Comput Sci Eng 2007;9:1020.
[33] Suuberg EM, Milosavljevic I, Oja V. Two-regime global kinetics of cellulose
pyrolysis: the role of tar evaporation. Symp (Int) Combust 1996;26:151521.
[34] Vrhegyi G, Jakab E, Antal Jr MJ. Is the BroidoShazadeh model for cellulose
pyrolysis true? Energy Fuels 1994;8:134552.
[35] Opfermann J, Kaisersberger E, Flammersheim H. Model-free analysis of
thermoanalytical data-advantages and limitations. Thermochim Acta
2002;391:11927.
[36] Miranda M, Bica C, Nachtigall S, Rehman N, Rosa S. Kinetical thermal
degradation study of maize straw and soybean hull celluloses by
simultaneous
DSc-tga
and
mdsc
techniques.
Thermochim
Acta
2013;565:6571.
[37] Matsuoka S, Kawamoto H, Saka S. What is active cellulose in pyrolysis? An
approach based on reactivity of cellulose reducing end. J Anal Appl Pyrol
2014;106:13846.

You might also like