Computers and Chemical Engineering 31 (2007) 1020–1034

Molecular reconstruction of naphtha steam cracking feedstocks

based on commercial indices
Kevin M. Van Geem a , Damien Hudebine b , Marie Françoise Reyniers a ,
François Wahl b , Jan J. Verstraete b , Guy B. Marin a,∗
a Laboratorium voor Petrochemische Techniek, UGent, Krijgslaan 281 (S5) B-9000 Gent, Belgium
b IFP-Lyon, Process Development and Engineering Division, BP 3, 69390 Vernaison, France

Received 20 January 2006; received in revised form 29 August 2006; accepted 1 September 2006
Available online 6 December 2006

Abstract
A new method for feedstock reconstruction of industrially important hydrocarbon fractions is developed using as input the analytically determined
commercial indices such as the average molecular weight, the specific density, the H/C-ratio, the paraffin, isoparaffin, olefins, naphthenes, aromatics
(PIONA) weight fractions and a set of ASTM boiling points. The method is based on Shannon’s entropy criterion and allows generating a molecular
composition of a naphtha fraction that meets all the boundary conditions set by the industrially available commercial indices. The most important
components present in a naphtha fraction are selected as possible feedstock components. This approach is very fast because the optimization
function can be transformed from a nonlinear equation in the N mole fractions (N ∼ 102 ) into a nonlinear equation in maximally 14 parameters,
making it very attractive for the petrochemical industry. A good correspondence exists between the simulated weight fractions and the actual
analytically determined ones provided that the specified commercial indices are sufficiently accurate. The feedstock module is implemented in a
simulation package for steam cracking but can be used as a stand-alone tool too. The combination of these two simulation tools makes it possible to
obtain a good agreement between the simulated and experimentally determined product yields for a set of 50 pilot plant experiments with different
naphtha feedstocks under widely varying operating conditions.
© 2006 Elsevier Ltd. All rights reserved.

Keywords: Steam cracking; Molecular reconstruction; Simulation; Naphtha; Optimization

1. Introduction and obtaining this kind of information is not straightforward.

The petrochemical industry is continually striving to improve
the performance of their installations. To this end, accurate
mathematical simulation models are an indispensable tool. For
reactor modeling, only fundamental kinetic models are able to
simulate the chemical kinetics over a wide range of process
conditions and for a wide range of feedstocks (Froment, 1992).
These reactor models account for both the chemical reactions
and the physical transport phenomena. Simulation models have
the advantage that once the model is developed, results can be
easily gathered and computer simulations take only a limited
time (Dente & Ranzi, 1979). One of the major problems of
these models is that a detailed feedstock composition is needed,
∗ Corresponding author. Tel.: +32 9 264 45 16; fax: +32 9 264 49 99.
E-mail address: kevin.vangeem@ugent.be (K.M. Van Geem).
During the past decade, several analytical techniques to obtain
a detailed molecular composition, e.g. gas chromatography
(GC), gas chromatography–mass spectrometry (GC–MS)
and high performance liquid chromatography (HPLC) have
been developed and improved. However, all these techniques
suffer from similar shortcomings; they are error-prone and
time-consuming. Nowadays there is an increasing trend to
replace these time-consuming analytical techniques by software
modules that can be easily implemented in simulation packages.
All these modules operate along the same fundamental princi-
ple: the reconstruction of the composition of a mixture based
on some easily obtainable average properties of the mixture,
i.e. the so-called commercial indices of the mixture. These are
for example the average molecular weight, the specific density,
the H/C-ratio, the paraffin, isoparaffin, olefins, naphthenes,
aromatics (PIONA) weight fractions and a set of ASTM boiling
points.

0098-1354/$ – see front matter © 2006 Elsevier Ltd. All rights reserved.
doi:10.1016/j.compchemeng.2006.09.001
 K.M. Van Geem et al. / Computers and Chemical Engineering 31 (2007) 1020–1034 1021

(1994) and Hudebine and Verstraete (2004). Although methods

Nomenclature based on reconstruction via lumped components are signifi-
List of symbols cantly faster, they are rather limited in their application range
cj concentration of species j (mol m−3 ) and not easily extendable. Moreover, methods using lumped
CIP coil inlet pressure (MPa) components have traditionally a rather limited flexibility as
CIT coil inlet temperature (K) they are based on a limited number of well-defined commercial
Cn carbon number (–) indices. Hence, the method can only be applied if all the
COP coil outlet pressure (MPa) required indices are available. Moreover, additional analytical
COT coil outlet temperature (K) information about the mixture cannot improve the predicted
d specific density (g/m3 ) composition because this extra information cannot be taken
dreac reactor diameter (m) into account. The stochastic methods such as those developed
fi ,j stoichiometric coefficient of molecule i for con- by Neurock et al. (1994) and Hudebine (2003) do not show
straint j variable these disadvantages but are computationally very demanding
fj value of constraint j variable (Hudebine, Vera, Wahl, & Verstraete, 2002). The main reason
F flow rate (g s−1 ) of the time-consuming character of the stochastic methods
%Fk weight percent of PIONA class k (–) is that a new library of possible molecules is generated for
%Gk volume percent of the boiling point k (–) each new simulation of the molecular composition. Therefore,
M molar mass (g mol−1 ) stochastic methods are less attractive for implementation in fast
N number of molecules in the library (–) commercial packages. The method developed by Liguras and
P pressure (MPa) Allen (1989), Allen and Liguras (1991) is significantly faster
S Shannon’s entropy (–) than the stochastic reconstruction methods of Neurock et al.
Tb boiling point (K) (1994) or Hudebine et al. (2002) because these authors worked
Tb,i% boiling point of i vol.% (K) with a predefined molecular library. To determine the mole
v additive atom or group contribution of the molar fractions of the molecules considered in the library, Liguras and
volume (m3 mol−1 ) Allen (1989a,b), Allen and Liguras (1991) used information
V molar volume (m3 mol−1 ) from both NMR (nuclear magnetic resonance) spectrometry
xi mole fraction of component i (–) and GC–MS to specify over 190 constraints for the system
of mole fractions. As the number of components exceeds the
Greek symbols number of boundary conditions, the authors opted to solve this
εk variable concerning constraint k (–) system by minimizing of a weighted objective function P:
λ Lagrange multiplier (–)
μ Lagrange multiplier (–) 
N
P= ωi xi (1)
␯ Lagrange multiplier (–)
i=1
π probability (–)
σk error linked to constraint k variable with ωi the weight for component i. Allen and Liguras sug-
σm standard deviation of the distribution linked to gested, for example, to minimize the objective function P
constraint m variable using the enthalpies of formation as weights for the different
ξ objective function (–) molecules. A disadvantage of the approach is that a huge
amount of analytical information is needed to generate all these
Subscripts constraints. Still, the method is potentially very useful provided
i molecule i/pseudo-component i that its flexibility is drastically increased.
j exact linear constraint j Clearly for each type of reconstruction method, several
k constraint with uncertainty k improvements are possible. The main objective in developing
m constraint with imposed distribution (m) a new method for feedstock reconstruction is to retain the
strong points and reduce/eliminate the weaknesses as much
Superscripts as possible. The main point in favor of the stochastic methods
calc calculated and those approaches used in SPYRO (Dente & Ranzi, 1979;
exp experimental Dente et al., 1988) and CRACKER (Joo et al., 2001) is that
the reconstruction of a detailed molecular composition is
based on readily available commercial indices of the mixture.
In literature, a distinction is made between two methods for Therefore, the novel method developed at IFP (Hudebine et
feedstock reconstruction: methods using a pseudo-component al., 2002; Hudebine, 2003; Wahl, Hudebine, & Verstraete,
representation (lumped components), e.g. SPYRO (Dente & submitted for publication) and used in this work also starts from
Ranzi, 1979; Dente et al., 1988) and CRACKER (Joo, Park, easily obtainable commercial indices as input for obtaining
& Lee, 2001), and methods that try to obtain a more detailed the detailed feedstock composition. Moreover, similar to the
molecular composition, e.g. Liguras and Allen (1989a,b), approach followed by Liguras and Allen (1989a,b), this method
Quann and Jaffe (1996), Neurock, Nigam, Trauth, and Klein determines the weight fractions of the molecules in the library
1022 K.M. Van Geem et al. / Computers and Chemical Engineering 31 (2007) 1020–1034
that minimize an objective function and meet all the boundary
conditions imposed by the commercial indices. However, this
specific method is based on the optimization of Shannon’s
entropy criterion (1948), originally applied in the information
theory, and results in some important advantages (Hudebine,
2003; Wahl et al., submitted for publication). In what follows
the principles of this method for feedstock reconstruction and its
build-up are discussed and explained. The approach is validated
via comparison with analytical results obtained for a number of
naphtha feedstocks and by comparison between simulated and
experimental results obtained from a set of pilot plant experi-
ments. Suggestions are made to extend the developed approach
for other fractions such as gasoils or even heavier fractions.
2. Feedstock characterization by maximization of the
entropy
2.1. Overview of the feedstock characterization method
Shannon’s entropy theory is widely applied in all sorts of
engineering fields, ranging from quantum chemistry over civil
engineering to hydrodynamics. Shannon’s entropy is defined as:

N 
N
S(πi ) = − πi ln πi with πi = 1
i=1 i=1
in which S represents Shannon’s entropy and πi is the probability
of a certain state. Shannon’s entropy is a measure of the homo-
geneity of a probability distribution. The principle of maximum
Shannon entropy states that if only partial information concern-
ing the possible outcomes is available, the probabilities are to be
chosen so as to maximize the uncertainty on the missing infor-
mation (Shannon, 1948). A larger entropy corresponds to a more
uniform distribution and hence to a larger uncertainty. Although
information provides reasons for preferring some possibilities
over others, the best solution to avoid unwarranted conclusions is
to assign a probability distribution which is as uniform as possi-
ble. This implies that the entropy has to be maximized subject to
constraints representing the available information. By applying
Fig. 1. General scheme of the method of maximization of the entropy.
the principle of maximum entropy, the most random distribution
subject to the given constraints can be obtained. Applying this
theory to the composition of petroleum fractions implies that the
probabilities in Eq. (2) are replaced by the mole fractions xi of
the feedstock components i. The first hypothesis of the method
of maximization of the entropy is that in the absence of any
information, it is impossible to favor one molecule over another,
and thus that the distribution of mole fractions is uniform.
In Fig. 1, an overview of the method developed by IFP for
feedstock reconstruction based on maximization of the entropy
is given. First, a molecular library is selected. Next, the mole
fractions of the molecules contained in the library are adjusted in
order to obtain a mixture with the desired characteristics, i.e. the
characteristics imposed by the commercial indices of the feed-
stock, e.g. the average molecular weight or the specific density.
Evidently, the mole fractions should also meet the maximization
criterion defined in Eq. (2). The computer program maximizes
the objective function and delivers a molecular composition that
meets the specified boundary conditions. The results obtained
for a set of test problems can be tested to refine the molecular
library and function as a feedback control loop. In the following
paragraphs, the method’s build-up is discussed in more detail.
2.2. Library selection
(2)
Although the selection of a molecular library might seem triv-
ial, it is of crucial importance for the success of the feedstock
characterization method. Evidently, a library containing compo-
nents that are not representative for a feedstock can never result
in an accurate characterization. One of the main reasons that
a judicial selection of the library is of the utmost importance
for the method based on the maximization of the information
entropy is that this is a statistical method where no correlations
are implemented. This method selects a single composition with
maximum Shannon entropy out of a number of compositions
that meet the boundary conditions. The molecular library deter-
mines the number of possible compositions that satisfy all the
boundary conditions, and thus affects the outcome of the maxi-
mization method. If an important component is not included in
 K.M. Van Geem et al. / Computers and Chemical Engineering 31 (2007) 1020–1034 1023

Table 1
Components included in the library of molecules for naphtha feedstock representation
C3 Propyne, 1,2-propadiene, propene, n-propane C4 Iso butane, n-butane
C5 2-methyl butane, n-pentane, cyclo pentane C6 2,3-Di-methyl butane, 2-methyl pentane, 3-methyl pentane, n-
hexane, methyl cyclo pentane, cyclo hexane, benzene
C7 2,3-Di-methyl pentane, 2-methyl hexane, 3-methyl hexane, C8 2-Methyl heptane, 3-methyl heptane, n-octane, di-methyl cyclo hex-
n-heptane, methyl cyclo hexane, di-methyl cyclo pentane, toluene ane, tri-methyl cyclo pentane, ethylbenzene, xylene
C9 2-Methyl octane, 3-methyl octane, n-nonane, tri-methyl C10 n-Decane
cyclohexane, tri-methyl benzene
C11 n-Undecane C12

the library, the composition obtained via the method can never be components. In supplementary material Annex 1 the extended
representative for the mixture. There are several possibilities for library of 173 components is given. This extension involves
constructing a molecular library (Hudebine, 2003) ranging from mainly isomers of one of the 37 components in Table 1. As will
experimental methods (Wahl et al., submitted for publication), be illustrated in the next paragraphs, including more detail in the
over group contribution methods (Hudebine et al., 2002) and feedstock library does not necessarily lead to better results. The
even stochastic methods (Hudebine & Verstraete, 2004). As the library in Table 1 is sufficiently detailed to allow an accurate sim-
method in this work is specifically developed to reconstruct ulation of the naphtha steam cracking process. Logically, other
the compositions of naphthas, we opted for an experimental types of oil fractions require the generation of a new molecular
method because the detailed composition of naphtha fractions library. For example, for gas oils a similar approach can be used.
can still be determined in a reasonable time, especially when Analyzing of a large amount of gas oil fractions should make it
compared to for example the heavier gas oil fractions. For very possible to identify the most important components. The latter
heavy feedstocks, such as vacuum gas (VGO’s), a stochastic then constitute the molecular library for this type of oil fraction.
method such as those presented by Neurock et al. (1994) and
Verstraete, Revellin, Dulot, and Hudebine (2004), or a group 2.3. The state with maximum shannon entropy
contribution method as proposed by Quann and Jaffe (1996) is
more suited because it becomes more and more difficult to obtain Shannon (1948) states that: “To find the most informative
a detailed experimental composition that can serve as a basis for state, the change in entropy with regard to the initial state
the molecule library. must reach a maximum”. This corresponds to maximizing the
Table 1 contains the list of selected components present in entropy-equation specified in Eq. (2) where the probabilities πi
the molecular library for naphtha feedstocks. These components are replaced by mole fractions xi of the components in the library.
are selected based on the detailed molecular composition of We consider the global solution of the following nonconvex
30 reference naphthas with widely varying characteristics. problem:
The range of the naphthas used is specified in Table 2. Their
average molecular weight varies between 79 and 101 g mol−1 . 
N
MAX S(xi ) = − xi ln xi (3)
Only those components with a weight fraction higher than
i=1
1 wt.% in at least one of the reference mixtures are selected.
This leads to a total of 37 selected components in the molecular

N
library that cover on average more than 90 wt.% of the entire Subject to : fj − fi,j xi = 0, j = 1, . . . , J
feedstock composition. The library includes eight n-paraffins, i=1
 
five aromatics, nine naphthenes, twelve iso-paraffins and three  N 
 
olefins. If, instead of applying the proposed selection criteria, fk − xi fi,k  , ≤ σk k = 1, . . . , K
all the components observed in one of the 30 reference naphthas  
i=1
would be included, the molecular library would expand to 173 0 ≤ xi ≤ 1
with fj the value of constraint j, fi,j the coefficient of molecule i
Table 2 for constraint j, N the number of molecules in the library, σ k the
Range of the commercial indices used for determining the library of molecules uncertainty on constraint k, J the number of exact linear con-
Commercial index Range straints and K the number of linear constraints with uncertainty.
If the specific density of the mixture is known, the following
Specific density, 15/4◦ 0.66–0.73
exact constraint is imposed:
PIONA-analysis (wt.%)
Paraffins 20–48 1 
N
xi M i
Iso-paraffins 18–44 − =0 (4)
Olefins 0–2 d exp di
i=1
Naphthenes 11–33
Aromatics 3–14 Table 3 gives an overview of the different boundary condi-
Average molecular weight (g mol−1 ) 79–101
tions resulting from the commercial indices that are considered
in this article.
1024 K.M. Van Geem et al. / Computers and Chemical Engineering 31 (2007) 1020–1034

Table 3 case, the optimization function can be transformed from a non-

Overview of the constraints resulting from the different boundary conditions
Commercial indices fi,j
Molecular weight Mexp − Mi
H/C-ratio ((H/C)exp − (H/C)i ) Mi
Specific density ((1/di ) − (1/d exp )) Mi
If component i ∈ PIONA class k
exp
%Fk /100 − 1 Mi
PIONA weight fraction
If component i ∈
/ PIONA class k
exp
%Fk /100) Mi
If boiling point i < boiling point k
exp
1 − (%Gk /100) (Mi /di )
ASTM boiling point
If boiling point i > boiling point k
exp
−(%Gk /100) (Mi /di )
Finding the global optimum of nonconvex problems such
as the one defined in Eq. (3) is not straight forward and can
be very time consuming (Floudas & Pardalos, 1996). From the
complexity point of view these type of global optimization prob-
lems belong to the class of NP-hard problems implying that as
the input size of the problem increases, the computational time
required to solve the problem is expected to grow exponentially
(Floudas & Pardalos, 1990). Although standard nonlinear pro-
gramming algorithms will usually obtain a local minimum or a
stationary point to a global optimization problem, such a local
minimum will only be global when certain conditions are sat-
isfied (such as S(xi ) is quasi-convex and the feasible domain is
convex) (Floudas & Pardalos, 1996). In general, several local
minima may exist and the corresponding function values may
differ substantially. Active research during the past decades has
produced a variety of methods for determining global solu-
tions to nonconvex nonlinear optimization problems (Biegler
& Grossmann, 2004). Neumaier, Shcherbina, Huyer, and Vinko
(2005) tested a number of existing state of the art solvers for
global constrained optimization and found that there are now
about a dozen solvers for constrained global optimization that
claim to solve global optimization and/or constraint satisfac-
tion problems to global optimality by performing a complete
search, e.g. BARON (Ryoo & Sahinidis, 1996). However, in the
maximum entropy field (Maxent) these methods are only rarely
used. Generally the optimization problem is solved using the
linear equation in the N mole fractions xi into a nonlinear
fj equation in J parameters λj . However, this imposes a restric-
0
tion on the commercial indices and affects which commercial
0 indices can and cannot be used. Information such as the Bureau
0 of Mines Correlation Index, the Watson characterization fac-
tor or the vapor pressure cannot be used as commercial indices.
0 The following list of commercial indices can be used in the infor-
mation entropy maximization method of this work: the average
0 molecular weight, the specific density, the H/C-ratio, the PIONA
weight fractions, a set of ASTM D2887 or D2892 boiling points,
0 NMR spectra and a detailed distribution of the hydrocarbons
per carbon atom determined via GC–MS. All these commer-
0
cial indices are obtained via a number of standardized methods
(ASTM methods).
The information entropy maximization method has the ten-
dency to favor the most extreme molecules in comparison to
molecules with average properties (Hudebine, 2003). On the
other hand, in refining, experimental data show that most com-
positions approach a Gaussian distribution of the mole fractions,
with a large amount of molecules with average properties and
only very little extreme molecules. Hence, an extra constraint
is necessary that restricts the method to mole fractions with a
Gaussian distribution as a function of the carbon number. The
only commercial index that is directly related to the number
of carbon atoms and is considered in Table 3 is the aver-
 is the standard deviation of the
age molecular weight. If σm
molecular weight the following extra boundary condition can be
introduced:

N
 2
(σm ) − xi (Mi − M exp )2 = 0 (6)
i=1
The solution of the optimization problem depends on the speci-
fied constraints. In the most general case, the problem includes
J exact linear constraints, K linear constraints with uncertainties
and a single constraint taking into account the normal distri-
bution for the molecular weight. The criterion that needs to be
optimized is of the following form (Hudebine, 2003; Hudebine
& Verstraete, 2004):
 

N 
N

Lagrange multiplier method (Guiasu & Shenitzer, 1985; Kapur ξ(x) = − xi ln xi + μ 1− xi

& Kesavan, 1992). The Lagrange multiplier method is used to i=1 i=1
 
find the optima of a function f(x) under the constraints gj (x) that 
J 
N
equal zero. A new function which incorporates the function f(x) + λj f1 − xi fi,j
and all its constraints is introduced: j=1 i=1
 N
ξ(x) = f (x) + 
J+K
1 (fk −
2
λj gj (x) (5) i=1 xi fi,k )
+ −
j
k=J+1
2 σk2
with λj a constant variable called the Lagrange multiplier. The  

N
 2 exp 2
optimization problem is then reduced to finding the optima of + v1 (σm ) − xi (Mi − M ) (7)
ξ(x) in xi and λj (Guiasu & Shenitzer, 1985; Kapur & Kesavan, i=1
1992). Solving the optimization problem can be drastically sim-
plified when all the constraints are linear in the variables xi The optimum of this function is found by setting the deriva-
(Guiasu & Shenitzer, 1985; Kapur & Kesavan, 1992). In this tives of this equation with respect to the mole fractions
 K.M. Van Geem et al. / Computers and Chemical Engineering 31 (2007) 1020–1034 1025

xi equal to 0: the problem (number of variables) (Floudas & Pardalos, 1990),

the problem becomes less difficult to solve when the number
∂ξ  J
of variables is decreased. As stated previously, the assumption
= −1 − ln xi − μ − λj fi,j
∂xi of using only linear constraints excludes the Bureau of Mines
j=1
Correlation Index (BMCI), the Watson characterization factor

J+K
fi,k and the vapor pressure. Note that if Raoult’s law holds, which
+ εk − v1 (Mi − M exp )2 , ∀i ∈ N (8) is not always the case, the vapor pressure can be used as con-
σk
k=J+1 straint. If for example the BMCI should be considered, the global
with: optimum should be found by evaluating the maximalisation cri-
N terion under all the constraints. However, this would require
fk − i=1 xi fi,k significantly more time, making the method less attractive for
εk = , ∀k ∈ K (9)
σk industrial practice. Although this is a limitation of the feedstock
Re-ordering the terms in Eq. (8) leads to the following expres- reconstruction method, it is not of huge importance because
sion: the BMCI and the Watson characterization factor are nowadays
⎡ ⎤ only rarely used in the petrochemical industry for characterizing
J  fi,k
J+K
hydrocarbon fractions.
e1+μ xi = exp ⎣− λj fi,j + εk − v1 (Mi − M exp )2 ⎦ In this work the Rosenbrock method (Rosenbrock, 1960) is
σk
j=1 k=J+1
used to locate the optimum of the function given in Eq. (13).
(10) The Rosenbrock method is a 0th order search algorithm. Yet, it
approximates a gradient search thus combining advantages of
Summation over all components i makes it possible to eliminate
0th order and 1st order strategies (Rosenbrock, 1960). In the
the Lagrange multiplier μ, resulting in an equation for the mole
first iteration, it is a simple 0th order search in the directions of
fraction xi :
   the base vectors of an n-dimensional coordinate system. In the
 fi,k case of a success, which is an attempt yielding a new minimum
exp − Jj=1 λj fi,j + J+Kk=J+1 σk εk − v1 (Mi − M
exp )2
xi = value of the target function, the step width is increased, while
Z in the case of a failure it is decreased and the opposite direction
× ∀i ∈ N (11) will be tried. Once a success has been found and exploited in
each base direction, the coordinate system is rotated in order to
with:
⎡ ⎤ make the first base vector point into the direction of the gradient.

N 
J  fi,k
J+K The creation of a new rotated coordinate system is done using a
Z= exp ⎣− λj fi,j + εk − v1 (Mi − M exp )2 ⎦ Gram-Shmidt orthogonalization procedure. Then all step widths
σk
i=1 j=1 k=J+1 are initialized and the process is repeated using the rotated coor-
(12) dinate system. Note that the nonconvexities in the optimization
criteria can affect the convergence and optimality with certain
If a normal distribution is imposed on the mole fractions, then the optimization routines. For this problem Nelder–Mead Methods,
number of parameters becomes larger than the number of com- so-called non-linear simplex methods (Nelder & Mead, 1965),
mercial indices. To determine the values of the parameters λj , although they are very similar to the Rosenbrock Algorithm
εk and v1 necessary to calculate the mole fractions, the entropy (Rosenbrock, 1960), should be avoided as there is no proof of
criterion needs be optimized. Substitution of Eq. (11) in Eq. (7) convergence for these algorithms. In fact, Nelder–Mead Meth-
leads to the following expression for the entropy criterion: ods can fail to find a local optimum of a very simple optimization
problem based on the minimization of a polynomial (Bazaraa,

J
E(λ, ε, v) = ln Z + λj fj Sherali, & Shetty, 1993). In our case convergence problems are
j=1 indeed encountered when Nelder–Mead methods are used. The

 Rosenbrock algorithm has also been proven to always converge

J+K
1 2 fk (Bazaraa, Sherali, & Shetty, 1993), i.e. global convergence to
+ v1 σ  1
2
+ ε − εk (13)
2 k σk a local optima is assured. Initializing the step widths to rather
k=J+1
high values enables the strategy to leave local optima and to go
The values of λj , εk and v1 at the optimum permit to calculate on with search for more global minima (Schwefel, 1981). It has
the most probable mole fractions of the molecules. One of the turned out that this simple approach is more stable than many
main advantages of this method over other methods is the limited sophisticated algorithms and it requires much less calculations
time necessary to determine a detailed molecular composition. of the target function than higher order strategies (Schwefel,
As only linear constraints are considered, the optimization func- 1981).
tion can be transformed from a nonlinear equation in the N mole In the absence of any information about the mixture, the only
fractions xi into a nonlinear equation in J + K + 1 parameters λj , restriction on the system is that the sum of the mole fractions
εk and v1 . Because N is in the order of 102 to 106 and J + K + 1 is has to equal 1. Again, the optimum of this function (13) can
maximum 13, the gain in the optimization level can be consid- be found by setting the derivatives of the function to the mole
erable. As the number of local optima increases with the size of fractions xi equal to zero. The mole fractions of the N library
1026 K.M. Van Geem et al. / Computers and Chemical Engineering 31 (2007) 1020–1034
components are then equal to:
1
xi = , ∀i ∈ N
N
This result verifies the first hypothesis of the information entropy
maximization method. In the absence of any information, con-
straint or analysis, it is impossible to favor one molecule with
regard to another, and thus the distribution of the mole fractions
is uniform. Every commercial index that leads to a constraint
specifies some extra information about the mixture, resulting in
a loss of the uniform character of the mole fractions.
To use the previously defined method, the physical properties
of the molecules in the library have to be known. For some
commercial indices like the average molecular weight, the values
of the individual components can be easily obtained. For the
calculation of the density, this is not so straightforward. The
density at 298 K of every individual component considered in the
molecular library is calculated by the group contribution method
of Fedors (1974). Fedors states that the solubility parameters and
the molar volumes of the molecules can be used to estimate the
thermodynamic properties of the molecules, such as the density.
A general system to determine the molar volume V of a molecule
developed by Fedors (1974) is based on the following relation:
 all the molecules that have a normal boiling point inferior to the
V = ni vi (15)
i
where vi represents the group contributions of the molar vol-
ume and ni the number of times that the atom or group i appears
in the molecule. These contributions are applicable at a temper-
ature of 298 K. Based on Eq. (15) and the molecular weight M
of the molecule, the density at 298 K is calculated:
M are determined and the commercial indices of the reconstructed
d=
i i vi
n
Fedors showed that the difference between the estimated and
the experimentally measured densities deviate within less than
10%.
To calculate the boiling point of a component, the group inter-
action contribution (GIC) method from Marrero, Pardillo, and
Fernandez (1999) is used. This method is specifically developed
for hydrocarbons and is known for its high statistical accuracy.
Fig. 2. Overview of the feedstock module SimCO used for the reconstruction of naphtha fractions implemented in the simulation package CRACKSIM.
A GIC method considers the contribution of the interactions
between two bonding groups instead of the contribution of sim-
(14) ple groups.
3. Results and discussion
3.1. Software module
The computer program predicting the detailed feedstock
composition is part of a complete software package for the
simulation of the steam cracking process containing the kinetic
module CRACKSIM (Clymans & Froment, 1984) as its core. A
general overview of the structure of the feedstock module sim-
ulated composition (SimCO) is given in Fig. 2. As input, the
commercial indices and the database containing the molecules
that are included in the library and their properties (molecular
weight, density, boiling point) are required.
First, the library is reduced by a number of filters based on
the actual values of the user-defined commercial indices. Several
filters are built in, for example, a PIONA filter that eliminates all
the components from a specific PIONA fraction if this fraction
is not experimentally observed. Another important filter is based
on the results of the ASTM boiling point curve and eliminates
initial boiling point or superior to the end boiling point of the
distillation. Next the entropy criterion is maximized. As stated
previously, for the optimization of Eq. (13) the robust Rosen-
brock optimization routine (Rosenbrock, 1960) is implemented.
Based on the values obtained for the Lagrange multipliers, the
mole fractions of the library components can be calculated using
Eqs. (11) and (12). Finally, the detailed PIONA weight fractions
(16) mixture are calculated.
The computer program is written in FORTRAN as compiled
codes written in Fortran or C for numerical calculations are very
fast (Alfradique & Castier, 2005). The FORTRAN code is com-
piled with a traditional Visual FORTRAN compiler, creating
a DOS executable. The user-friendliness of the DOS applica-
tions can be improved by the use of a graphical user interface
(GUI). Also to meet the CAPE-Open standard (Barrett & Yang,
2005; Köller & Töbermann, 2002) this FORTRAN code should
 K.M. Van Geem et al. / Computers and Chemical Engineering 31 (2007) 1020–1034 1027

Fig. 3. The results screen of the GUI for a mixture with an average molecular weight of 100.4.

be “migrated”. A wrapping approach seems to be most suited In comparison to the method of Liguras and Allen (1989a,b),
in this case implying that around the compiled DOS program Allen and Liguras (1991) and Hudebine and Verstraete (2004)
another program, the GUI, is built. The GUI creates the neces- the gain in time is considerable. The method of Liguras and
sary input files for the DOS program, enables to run the program Allen (1989a,b), Allen and Liguras (1991) is very time consum-
in the background (invisible to a user), reads the output and ing because of the large amount of analytical work necessary to
visualizes the results. In this work Visual Basic.Net is used to determine the 190 constraints imposed on the molecular library.
develop the GUI. The .Net framework allows applications to Gathering the complete compound class analysis accurately,
develop into a distributed, cross platform simulation and process using a combination of chromatographic separation in combi-
control environment that can be easily extended to incorporate nation with mass spectrometry and nuclear magnetic resonance
novel chemical process computing applications (Barrett & Yang, spectroscopy, as done by Liguras and Allen (1989a,b), Allen and
2005). The core of the program still remains the FORTRAN Liguras (1991), can take several days. Our method uses easily
source code, compiled with a Visual Fortran compiler creating an obtainable commercial indices, which are all obtained within
executable. The executable is called from the Visual Basic.Net the hour using standard ASTM methods. Also compared to the
source code and runs in the background. The program is cur-
rently not available to the general community as an open-source Table 4
application, but a trial version can be obtained on request. The Simulation time and number of function evaluations as function of the number
results screen of the GUI for a simple demonstration case is of imposed commercial indices
shown in Fig. 3. These results are obtained for a mixture with Number of imposed Number of function Simulation
an average molecular weight of 100.4 F. The simulation time for commercial indices evaluations time (s)
obtaining this detailed composition is 0.02 s on an Intel Pentium
1 269 0.02
III processor of 1.0 GHz. In Table 4, the simulation time and 5 2134 0.08
the number of function evaluations as function of the number 10 12006 0.25
of imposed commercial indices are given. These results show 13 32756 0.56
that the simulation time increases rapidly with the number of Commercial indices: Mm = 95 g mol−1 ; C/H = 2.1; ρ = 0.69; P = 43.7 wt.%;
imposed commercial indices and the number of function evalu- I = 38.4 wt.%; O = 0.7 wt.%; N = 13.2 wt.%, A = 4.0 wt.%; Tb ,0.5 = 300 K;
ations increases rapidly with the number of imposed commercial Tb ,10 = 320 K; Tb ,30 = 340 K; Tb ,50 = 370 K; Tb ,70 = 380 K; Tb ,90 = 420 K;
indices. Tb ,99.5 = 470 K.
1028 K.M. Van Geem et al. / Computers and Chemical Engineering 31 (2007) 1020–1034

method of Hudebine and Verstraete (2004) our method is signif- Table 6

icantly faster. If the method of Hudebine and Verstraete (2004) Distribution of the weight fractions of the paraffinic components as function the
uncertainty on the molecular weight (σ k )
is used, the gain in time is not situated in the analytical part
but in the time necessary for the stochastic reconstruction of the Weight fractions (wt.%) σ k = 0.5 σk = 2 σ k = 10 Base case
molecular library. For example, the time to generate a molecular n-Butane 16.4 14.8 10.9 16.5
library of an LCO gasoil using the stochastic method is typically n-Pentane 15.9 14.8 12.0 16.0
20 h. In our case the library already exists and no time is required n-Hexane 14.8 14.3 12.7 14.9
to generate a new library. n-Heptane 13.5 13.3 13.0 13.5
n-Octane 12.0 12.4 13.2 12.0
n-Nonane 10.5 11.2 13.0 10.4
3.2. Effect of introducing uncertainty and a Gaussian n-Decane 9.1 10.1 12.8 9.0
distribution n-Undecane 7.8 9.0 12.4 7.7
Calculated Mm 90.2 92.3 98.2 90.0
One advantage of using the method based on the maximiza- Simulation time (s) 0.07 0.07 0.10 0.02
tion of the entropy is that next to the boundary conditions Base case: Mm = 90; P = 100; σ k = 0. Simulation time on an Intel Pentium III
originating from the commercial indices other criteria can be processor of 1 GHz.
used to improve the simulation results. For example, experimen-
tal data show that most naphtha fractions have a large amount of  should not be higher than 3. Com-
the standard deviation σm
molecules with average properties and only very little molecules
parison with experimental data further shows that values of the
with extreme properties. Hence, using a Gaussian distribution  , lower than 1.5 are unrealistic as the distri-
standard deviation σm
of the mole fractions as function of the number of carbon atoms
bution of the mole fractions becomes too narrow. A comparison
could improve the predictions of the feedstock reconstruction
between experimental and simulated naphtha compositions of
module. As shown in Eq. (9), an extra constraint on the system
30 different naphtha fractions with widely varying characteris-
can be introduced to restrict the solution to a normal distribu- 
tics intervals shows that the value of the standard deviation σm
tion of the mole fractions as function of the carbon number.
always lies within the [1.5,3] interval. However, no unique ideal
Table 5 shows the effect of using a Gaussian distribution as extra  can be determined. The optimal
value of standard deviation σm
boundary condition for a mixture containing only n-paraffins.  varies strongly from case to
value of the standard deviation σm
Evidently, for very small values of the standard deviation of
case, from one naphtha to another naphtha. Therefore, imposing
the Gaussian distribution, the distribution of mole fractions of the
 ) of the a Gaussian distribution is very difficult for industrial practice.
paraffins is also narrow. As the standard deviation (σm
The simple demonstration case used in the previous para-
Gaussian distribution increases, the distribution first becomes
graph can also be used to illustrate the effect of introducing
more and more uniform. For very high values of the stan-
 , the distribution of the mole fractions of the uncertainty on one of the boundary conditions. In Table 6, the
dard deviation σm
effect of using uncertainty on the molecular weight is shown.
paraffins flips. The mole fractions of components with average
As the uncertainty on this value increases, the mole fractions
properties decrease, while the components on the edge of the
become more and more uniformly distributed. Moreover, the
system, i.e. components with either a very high or a very low
deviation of the calculated molecular weight deviates from the
molecular weight, become more and more important. This last
specified value of 90. Using an uncertainty on one particular
situation is of course unrealistic as experimental data show that
commercial index has the advantage that the importance of that
most naphtha compositions approach a Gaussian distribution of
commercial index can be reduced compared to other more accu-
the mole fractions. Practically, this implies that the values of
rately known indices. However, this does not imply that the
value of this commercial index is completely ignored as can
Table 5 be seen in Table 6. Only for very high values of the uncer-
Distribution of the weight fractions of the paraffinic components as function tainty, i.e. σ k > 5, the specified molecular weight is completely
 ) of the Gaussian distribution introduced on the
of the standard deviation (σm
ignored.
specified average molecular weight
Weight fractions (wt.%)  = 1.5
σm  = 2.5
σm  =5
σm Base case
3.3. Reconstruction of naphtha feedstocks
n-Butane 0.1 2.9 17.9 16.5
n-Pentane 7.2 15.2 15.6 16.0 The main objective of the feedstock reconstruction method
n-Hexane 53.8 33.8 13.7 14.9
n-Heptane 36.6 32.2 12.2 13.5
is to generate in a minimum of time a detailed molecular
n-Octane 2.3 13.3 11.0 12.0 composition with the desired characteristics as specified by
n-Nonane 0 2.4 10.2 10.5 the commercial indices, not to precisely reconstruct the mix-
n-Decane 0 0.2 9.8 9.1 ture’s composition. It is a priori clear that the composition with
n-Undecane 0 0 9.6 7.8 maximum Shannon entropy will only rarely correspond to the
Calculated Mm 90 90 90 90 analytically determined one. The composition with maximum
Simulation time (s) 0.03 0.02 0.02 0.02 Shannon entropy is the statistically favored one, since this com-
 . Simulation time
Base case: Mm = 90; P = 100 wt.%; no standard deviation σm position maximizes the uncertainty on the missing information.
on an Intel Pentium III processor of 1 GHz. The error introduced by the missing information is minimized
 K.M. Van Geem et al. / Computers and Chemical Engineering 31 (2007) 1020–1034 1029

but nothing more than that. However, it can be expected that Table 8
this novel method is able to generate a detailed molecular com- Simulated detailed PIONA weight fractions (wt.%) with the commercial indices
specified in Table 6
position that corresponds reasonably well to the analytically
determined one. In particular, when more and more informa- wt.% P I O N A Sum
tion about the mixture is known, i.e. more commercial indices C4 0.2 0.2 – – – 0.4
are specified, the agreement should become better and better. In C5 7.2 0.1 0.0 2.8 – 10.1
the next paragraph, this will be illustrated in more detail. C6 6.7 13.2 0.1 5.1 2.4 27.5
Gas chromatography (GC) is considered to be the most C7 6.0 11.1 0.1 6.5 2.6 26.2
C8 5.2 6.0 0.0 6.2 5.5 22.9
appropriate method for determining the detailed molecular com- C9 4.5 4.8 – 2.9 0.3 12.5
position of naphtha fractions. The calibration factors used for the C10 0.1 0.1 – – – 0.2
quantitative analysis are those proposed by Dietz (1967). Cal- C11 0.1 – – – – 0.1
ibration factors for components not mentioned in this article Sum 30.1 35.4 0.2 23.6 10.8 100.0
are calculated using the group contribution method of Dierickx,
Simulation qime of 0.05 s on an Intel Pentium III processor of 1 GHz.
Plehiers, and Froment (1986). Obviously, the quantitative anal-
ysis of the mixture should be preceded by a qualitative analysis.
The latter is carried out using GC–MS data, Kovats retention of the low amount of olefins in this feedstock no olefinic fraction
indices and reference data from previously studied naphthas. The as function of the carbon number is shown in Fig. 4. Over-
molecular composition of the naphtha fraction serves as a basis all a good agreement can be observed between the simulated
to determine the detailed PIONA weight fractions used for vali- and analytically obtained PINA weight fractions. Fig. 4 further
dation purposes. In Table 7, the commercial indices obtained for shows that if more information is specified about the mixture,
a specific naphtha fraction, Naphtha 1, are specified. For Naph- such as the molecular weight and some more ASTM-boiling
tha 1, only the specific density, the PIONA weight fractions and points, a better agreement with the analytically determined PINA
three ASTM boiling points are known. Although limited, the weight fractions is observed. In this case, these last extra values
available information is sufficient for the software to simulate a have been calculated from the detailed analytically determined
possible composition with the desired commercial indices. The molecular composition. However, they also could have been
commercial indices specified in Table 7 are all considered as determined analytically.
exact linear constraints for the optimization problem. Neither a The differences seen in Fig. 4 are caused by several reasons.
Gaussian distribution of the weight fraction, nor error bars on First of all the specified commercial indices in Table 7 are rather
one of the commercial indices is considered. If no error bars are limited. More information about the mixture seriously improves
specified on the PIONA weight fractions it is very important that the agreement as illustrated in Fig. 4. Secondly, the inaccuracy of
the sum of these weight fractions exactly equals 100%. Other- certain commercial indices can cause significant deviations. A
wise contradicting constraints are specified for the system and manually performed sensitivity analysis shows that the method is
no optimum can be found. in particular sensitive to changes of some ASTM-boiling points.
In Table 8, the simulated detailed PIONA weight fractions This is obvious because the ASTM-boiling point curve is one
are given. The commercial indices of the simulated mixture of the only indices that is directly related with the carbon num-
meet the values specified in Table 7 except for the IBP, where ber distribution of the molecules, and thus strongly influences
a minor difference is observed. The value obtained via simu- the distribution of the molecules in a specific PIONA fraction.
lation was 299 K while a value of 305 K was specified. Fig. 4 Finally, as stated previously, this is a statistical method without
gives an overview of the agreement between the simulated and any correlations. This method selects a single composition with
analytically obtained detailed paraffin, isoparaffin, naphthenes, maximum Shannon entropy out of a number of compositions
aromatics (PINA) weight fractions per carbon number. Because that meet all the boundary conditions. Hence, it is inherent to
the method that some differences exist. The composition with
maximum Shannon entropy is not necessary the one obtained
Table 7 experimentally. The same conclusions have been found using
Commercial indices of Naphtha 1
other naphtha samples.
Commercial indices A possible solution for the problems encountered with the
Specific density, 15/4◦ 0.71 ASTM boiling points is to impose uncertainties on the con-
straints linked to the boiling points. If the commercial indices
PIONA-analysis (wt.%)
Paraffins 30.0 from Table 7 are used as input data for the program after addi-
Iso-paraffins 35.3 tion of an uncertainty of 3 on each of the boiling points, even
Olefins 0.2 larger deviations from the actual composition are observed. The
Naphthenes 23.5 method of maximization of the entropy tends to a distribution
Aromatics 11.0
of the mole fractions that is as uniform as possible. The input
ASTM boiling points (K) of the uncertainties gives the method extra opportunity to make
IBP 305 the distribution more uniform. Another option to overcome the
50% 365
FBP 434
shortcomings related to the boiling points is to ignore these com-
mercial indices completely. Only the specific density and the
1030 K.M. Van Geem et al. / Computers and Chemical Engineering 31 (2007) 1020–1034

Fig. 4. Comparison of the detailed PIONA weight fractions. (a) Weight fractions of paraffinic components, (b) weight fractions of iso-paraffinic components, (c)
weight fractions of naphthenic components, (d) weight fractions of aromatic components (||simulated with commercial indices from Table 7,  experimental results,
( ) simulated with commercial indices from Table 7 +Mw = 92.2; Tb,10 = 311 K; Tb,30 = 334 K; Tb,70 = 383 K; Tb,90 = 414 K).

PIONA weight fractions remain as boundary conditions for the
optimization of the entropy criterion. As in the situation with
imposed uncertainties, less constraints increase the degrees of
freedom, resulting in a more uniform distribution of the mole
fractions.
Note that prior to the start of a new feedstock simulation
the imposed commercial indices are subjected to several tests,
mainly to verify if the specified commercial indices can corre-
spond to a naphtha fraction. Three different tests on the com-
mercial indices are performed. The average molecular weight
must be comprised between 140 and 60 g mol−1 . The highest
specified ASTM boiling point should be lower than 500 K. Also
the specific density should lie between 0.65 and 0.75. If these
criteria are met, our method will reconstruct a molecular naph-
tha composition with practically the same commercial indices as
those specified, even when the imposed commercial indices are
outside of the range of those specified in Table 2. Only when this
method is used for entirely different feedstocks, such as gasoils
or gas condensates it cannot produce a composition that meets all
the boundary conditions set by the commercial indices. There-
fore, these tests are carried out and different warning messages
are given when the commercial indices are outside the specified
range.
3.4. Simulation of a set of pilot plant experiments
The results in the previous paragraph show that the simula-
tion module is able to reconstruct a feedstock composition based
on the available commercial indices. However, the most impor-
tant test for our feedstock reconstruction program SimCO is to
compare the simulation results of the complete software pack-
age for steam cracking, i.e. feedstock and reactor modeling, with
experimental data obtained from the pilot plant installation of
the Laboratorium voor Petrochemische Techniek (LPT) in Ghent
University (Van Damme & Froment, 1982; Van Geem, Reyniers,
& Marin, 2005). This experimental setup allows fundamental
studies of the kinetics of the cracking reactions (Wauters &
Marin, 2002) as well as practical issues such as coke deposition
in both the radiant coil (Reyniers & Froment, 1995) and the trans-
fer line exchanger (TLE) (Dhuyvetter, Reyniers, Froment, &
Marin, 2001). The pilot plant installation consists of three parts:
a feed section, the furnace containing the suspended reactor coil
and the analysis section. The tubular reactor used in this set of
experiments has a length of 23.14 m and has an internal diame-
ter of 10 mm. These dimensions are chosen to achieve turbulent
flow conditions in the coil. The temperature and pressure profile
along the reactor can be measured and regulated. As feedstock,
 K.M. Van Geem et al. / Computers and Chemical Engineering 31 (2007) 1020–1034 1031

Table 9 reactor model for simulating the steam cracking process can
Characteristics of the naphtha feedstocks used for the SimCO simulations and lead to errors, e.g. when modeling ethane cracking in a Lum-
the range of experimental conditions in the LPT pilot plant installation
mus SRT-I reactor (Van Geem, Heynderickx, & Marin, 2004).
Commercial indices The source of these errors is that in industrial single coil reac-
Naphtha 2 Naphtha 3 Naphtha 4 Naphtha 5 tors for ethane cracking important radial temperature gradients
(>100 K) exist (dreac = 100 mm) (Van Geem et al., 2004). As the
Specific density, 15/4◦ 0.68 0.69 0.67 0.68
tube diameter in the pilot plant reactor used in this work is much
PIONA-analysis (wt.%) smaller (i.e. dreac = 10 mm), the radial temperature gradients are
Paraffins 43.9 34.0 40.1 39.9
far less pronounced (<15 K) thereby allowing a reasonable accu-
Iso-paraffins 38.6 38.1 50.5 38.2
Olefins 0.7 0.1 0.4 0.2 rate simulation using a 1-dimensional reactor model. The reactor
Naphthenes 13.3 20.3 5.8 16.4 model is coupled to a radical kinetic model for the cracking of
Aromatics 3.6 7.5 3.2 5.3 light fractions consisting of 60 molecular and 68 radical species
Molar H/C ratio 2.23 – – – and over 1200 reactions (Clymans & Froment, 1984). The par-
ity plots for the yields of the main cracking products methane,
Experimental conditions ethene, propene and benzene in Fig. 5 show that the combination
CIP (MPa) 2.4–2.1 of the feedstock module SimCO with the 1-dimensional reactor
COP (MPa) 1.8–1.5 model is able to provide accurate simulation results for naphtha
CIT (K) 823–873 fractions over a wide range of process conditions. Note that the
COT (K) 953–1170 differences in the parity plots can almost entirely be attributed to
Dilution (kgsteam /kgHydrocarbon ) 0.2–0.8
Hydrocarbon flow rate (kg/s) 0.6 × 10−3 –1.2 × 10−2
the microkinetic model. The same parity plots are obtained if the
P/E-ratio (wt.%/wt.%) 0.9–0.6 set of pilot plant experiments is simulated with the analytically
determined molecular composition instead of the one obtained
using SimCO.
four different naphtha fractions Naphtha 2–5 are used. In Table 9, As stated earlier a crucial element for the success of the com-
both the commercial indices of these naphtha fractions as well plete simulation package (feedstock reconstruction + reactor
as the experimental conditions used in the different experiments modeling) is the judicious selection of candidate feedstock
are specified. In total, 50 pilot plant experiments are simulated molecules in the molecular library. Using a more extensive
for which the experimental conditions vary over a broad range. library does not necessary lead to better simulation results. On
The flow rate of the hydrocarbon feedstock is varied between the contrary, the simulation results become worse as can be
0.6 and 1.2 g s−1 , while the coil outlet temperature varies from seen in Table 11. Using a more extensive molecular library
953 to 1170 K. The dilution varies from 0.2 kgsteam /kgnaphtha to (Sim1) containing the 173 components observed during the
0.8 kgsteam /kgnaphtha . The coil outlet pressure varies from 0.18 to analysis of the molecular compositions of the 30 reference naph-
0.15 MPa. These conditions correspond to a P/E-range (propene thas leads to worse results than the results obtained with the
to ethene ratio) from 0.9 to 0.6. reduced library containing 37 key molecules (Sim2). This, in
First, the detailed feedstock composition is simulated using particular, is the case for the two main products, ethene and
the commercial indices of the different naphthas as input. The propene. Also for other steam cracking experiments this phe-
simulated detailed PIONA weight fractions of Naphtha 2 are
given in Table 10. The determined composition exactly meets Table 11
all the specified boundary conditions specified in Table 9. Once Simulated and experimentally determined product yields obtained for Naphtha
the detailed molecular composition of the naphtha feedstock 2 in the LPT pilot plant reactor
is determined, the reactor simulations can be performed. The Product yields (wt.%) Sim1 Sim2 Sim3 Experimental
experiments carried out in the pilot plant reactor are modeled
using a 1-dimensional reactor model. Using a 1-dimensional Hydrogen 0.6 0.6 0.6 0.6
Methane 10.1 10.2 10.3 10.4
Ethyne 0.2 0.2 0.2 0.2
Table 10 Ethene 20.1 21.7 21.5 21.3
Simulated detailed PIONA weight fractions for Naphtha 2 (wt.%) with the Ethane 3.0 3.0 3.0 4.1
commercial indices specified as in Table 9 Propene 18.3 17.3 17.4 17.3
wt.% P I O N A Sum 1,3-Butadiene 4.5 5.0 4.9 4.1
1-Butene 3.9 3.7 3.4 2.9
C4 9.2 5.9 – – – 15.1 2-Butene 1.0 1.3 1.4 1.5
C5 8.2 4.7 0.2 3.7 – 16.8 Isobutene 3.3 3.2 3.0 3.1
C6 6.8 10.7 0.2 8.2 – 26.1 Benzene 2.5 2.6 3.0 3.3
C7 5.7 8.3 0.2 0.8 0.1 15.1
P/E-ratio (wt.%/wt.%) 0.9 0.8 0.8 0.8
C8 4.7 4.4 0.0 0.4 0.8 10.3
C9 3.8 3.3 – 0.1 2.6 9.9 Sim1: molecular library of 173 naphtha components (supplementary material
C10 3.1 1.2 – – – 4.3 Annex 1); Sim2: molecular library of 37 key components; Sim3: sim-
C11 2.5 – – – – 2.5 ulation with analytically determined composition (Annex 2) (simulation
Sum 43.9 38.6 0.7 13.3 3.6 100.0 conditions: CIT = 823 K; COT = 1073 K; CIP = 0.22 MPa; COP = 0.18 MPa;
F = 1.2 × 10−3 kg s−1 ; δ = 0.2 kg/kg).
1032 K.M. Van Geem et al. / Computers and Chemical Engineering 31 (2007) 1020–1034
Fig. 5. Parity plot for the ethene yield (simulation conditions: CIT: 823–873 K; COT: 953–1170 K; CIP: 0.24–0.21 MPa; COP: 0.18–0.15 MPa; F: 0.6–1.2 g s−1 ; δ:
0.2–0.8 kg/kg).
nomenon is observed. The main reason for this behavior is that
the entropy method is too insensitive to predict mole fractions
of non-important components accurately because it differenti-
ates between components based on the difference in physical
properties of these components. The molecular library of 173
components in supplementary material Annex 1 is an extension
of the molecular library of 37 components with for 95% iso-
mers of the 37 components. These isomers have very similar
physical properties. As any method for feedstock reconstruction
belonging to the same family, i.e. Liguras and Allen (1989a,b),
Allen and Liguras (1991), Neurock et al. (1994), Hudebine and
Verstraete (2004), differentiates between components based on
these physical properties, it seems useless to include isomers
with almost identical physical properties as it is almost impossi-
ble to distinguish between them based on average properties of
the mixture, such as the commercial indices. This is for example
the case for isomers of branched paraffinic and naphthenic com-
ponents. Although the physical properties of these isomers are
quite similar their cracking behavior can differ significantly, and
hence, incorrectly estimating the distribution of the different iso-
mers can affect the simulation results. By carefully constructing
the molecular library and by selecting only those components
with weight fractions higher than 1 wt.% in at least one of the
reference mixtures this effect can be minimized. As illustrated
in Table 11, the simulation results obtained with the molecular
library containing 37 key molecules (Sim2) practically coin-
cides with the results obtained with the analytically determined
composition (Sim3). These results are confirmed by simulation
results obtained for other pilot plant experiments.
4. Conclusions
A new method for feedstock reconstruction for naphtha feed-
stocks is applied using the analytically determined commercial
indices as input. This method is based on Shannon’s entropy
criterion and creates a molecular composition that meets all the
boundary conditions set by the available commercial indices.
One of the advantages of this method is that not only exact con-
straints can be used, also uncertainties can be built in. Indeed,
some of the commercial indices are error-prone because they are
based on analyses of petroleum fractions with a certain uncer-
tainty. Therefore, uncertainties can be specified by the user,
resulting in a higher flexibility to adapt the model to one’s needs.
Furthermore it is also possible to take into account the fact that
petroleum fractions possess particular distributions.
 K.M. Van Geem et al. / Computers and Chemical Engineering 31 (2007) 1020–1034
One of the key elements for success is a good selection of the
library of molecules because no correlations are implemented
at any stage of a simulation. Based on the analysis of a large
number of reference feedstocks the most important components
traditionally present in a naphtha fraction are selected as possi-
ble feedstock components. For other feedstocks than naphtha, a
new library has to be selected because every type of petroleum
fraction has its own molecular characteristics. In this way, the
method can be easily extended to other fractions such as gas
oils.
One of the main advantages of this method over other
methods is the limited time necessary to determine a detailed
molecular composition. This is because only linear constraints
are considered. With an Intel Pentium III processor of 1.0 MHz
even when all the commercial indices are known it takes less
than one second to obtain a detailed feedstock composition. If
less commercial indices are specified the simulation time drops
rapidly. Comparing the calculated detailed PIONA analysis with
the actual analytically determined one shows that there exists a
reasonable correspondence between the two compositions; how-
ever, these results also bring out some shortcomings. Important
is the accuracy of the specified commercial indices because the
simulation results can be very sensitive to, for example, the
specified ASTM-boiling points. Furthermore, it is obvious that
the number of analytically determined commercial indices also
affects the simulation results. If these are accurately determined,
more commercial indices lead to better simulation results. It is
also important to state that not all the shortcomings are caused
by inaccuracies of the indices or by inconsistency between the
experimental and the calculated values; also some part is inher-
ent to the method itself. It is obvious that without introducing any
correlations it is hard to obtain a completely accurate detailed
molecular composition.
The fast reconstruction of a molecular composition makes
the feedstock module SimCO very attractive for implementation
in the simulation package for steam cracking. The combina-
tion of these two simulation tools makes it possible to obtain
simulation results even faster than before because no time has
to be wasted for determining a detailed molecular composition
of the naphtha feedstock. A graphical user interface improves
the user-friendliness of the simulation package and minimizes
further loss of valuable time. Comparison of the simulation
results with pilot plant data obtained from a set of 50 pilot plant
experiments performed in the LPT-pilot plant installation shows
a good agreement between the simulated and experimentally
determined product yields. This result is remarkable consider-
ing that only a very limited amount of information is specified
for the considered naphtha fractions. One of the key elements
for success is a judicious choice of the components included in
the molecular library.
Simulation results show that the quality of the simulations
improves if a library containing only key components of
naphtha fractions is used instead of an extensive library. The
entropy method is too insensitive to predict mole fractions of
non-important components accurately based on the generally
available commercial indices (ASTM boiling points, specific
density, PIONA weight fractions) because it differentiates
1033
between components based on the difference in physical
properties of these components.
Acknowledgments
Kevin M. Van Geem was a holder of a PhD grant of
the Institute for the Promotion of Innovation by Science
and Technology in Flanders (IWT-Vlaanderen) and is cur-
rently a Postdoctoral researcher of the Fonds Wetenschappelijk
Onderzoek-Vlaanderen.
Appendix A. Supplementary data
Supplementary data associated with this article can be
found, in the online version, at doi:10.1016/j.compchemeng.
2006.09.001.
References
Alfradique, M. E., & Castier, M. (2005). Modeling and simulation of reac-
tive distillation columns using computer algebra. Computers and Chemical
Engineering, 29, 1875–1884.
Allen, D. T., & Liguras, D. (1991). Structural models of catalytic cracking chem-
istry: A case of a group contribution approach to lumped kinetic modeling.
New York: Mobil Workshop: Chemical Reactions in Complex Mixtures.
Barrett, W. M., & Yang, J. (2005). Development of a chemical process mod-
eling environment based on CAPE-OPEN interface standards and the
Microsoft.NET framework. Computers and Chemical Engineering, 30,
191–201.
Bazaraa, M. B., Sherali, H. D., & Shetty, C. M. (1993). Nonlinear programming:
Theory and algorithms (2nd ed.). New York: Wiley.
Biegler, L. T., & Grossmann, I. E. (2004). Retrospective on optimization. Com-
puters and Chemical Engineering, 28(8), 1169–1192.
Clymans, P. J., & Froment, G. F. (1984). Computer generation of rate equations
in the thermal cracking of normal and branched paraffins. Computers and
Chemical Engineering, 8, 137–142.
Dente, M., & Ranzi, E. (1979). Detailed prediction of olefin yields from
hydrocarbon pyrolysis through a fundamental simulation program SPYRO.
Computers and Chemical Engineering, 3, 61.
Dente, M., Bozzano, G., Sugaya, M., & McGreavy, C. (1988). The characteriza-
tion of residual hydrocarbon fractions with model compounds—Retaining
the essential information. ACS National Meeting, Boston, 43, 653–
657.
Dhuyvetter, I., Reyniers, M. F., Froment, G. F., & Marin, G. B. (2001). The
influence of dimethyl disulfide on naphtha steam cracking. Industrial &
Engineering Chemistry Research, 40, 4353–4362.
Dietz, W. A. (1967). Response factors for gas chromatographic analyses. Journal
of Gas Chromatography, 5, 68.
Dierickx, J. L., Plehiers, P. M., & Froment, G. F. (1986). Online gas-
chromatographic analysis of hydrocarbon effluents—Calibration factors and
their correlation. Journal of Chromatography, 362, 155.
Fedors, R. (1974). A method for estimating both the solubility parameters
and molar volumes of liquids. Polymer Engineering and Science, 14,
147.
Floudas, C. A., & Pardalos, P. M. (1990). A collection of test problems for
constrained global optimization algorithms. New York: Springer-Verlag.
Floudas, C. A., & Pardalos, P. M. (1996). State of the art in global optimization:
Computational methods & applications. New York: Kluwer.
Froment, G. F. (1992). Kinetics and reactor design in the thermal cracking for
olefin production. Chemical Engineering Science, 47, 2163.
Guiasu, S., & Shenitzer, A. (1985). The principle of maximum entropy. The
Mathematical Intelligencer, 7, 42–48.
Hudebine, D. (2003). Reconstruction Moléculaire de Coupes Pétrolières, PhD
Thesis, Ecole Normale Supérieure de Lyon.
1034 K.M. Van Geem et al. / Computers and Chemical Engineering 31 (2007) 1020–1034
Hudebine, D., Vera, C., Wahl, F., & Verstraete, J. J. (2002). Molecular represen-
tation of hydrocarbon mixtures from overall petroleum analyses. In AIChE
2002 Spring Meeting. Paper 27a.
Hudebine, D., & Verstraete, J. J. (2004). Molecular reconstruction of LCO
gasoils from overall petroleum analyses. Chemical Engineering Science,
59, 4755–4763.
Joo, E., Park, S., & Lee, M. (2001). Pyrolysis reaction mechanism for industrial
naphtha cracking furnaces. Industrial & Engineering Chemistry Research,
40, 2409–2415.
Kapur, J. N., & Kesavan, H. K. (1992). Entropy optimization principles with
applications. San Diego: Academic Press.
Köller, J., & Töbermann, J. C. (2002). Global cape-open. Migration Cokebook,
http://www.colan.org.
Liguras, D. K., & Allen, D. T. (1989a). Structural models for catalytic crack-
ing. 1. Model compound reactions. Industrial & Engineering Chemistry
Research, 28, 665.
Liguras, D. K., & Allen, D. T. (1989b). Structural models for catalytic cracking.
2. Reactions of simulated oil mixtures. Industrial & Engineering Chemistry
Research, 28, 674.
Marrero, M. J., Pardillo, F. E., & Fernandez, B. S. (1999). Estimation of hydrocar-
bon properties from group-interaction contributions. Chemical Engineering
and Communication, 176, 161.
Nelder, J. A., & Mead, R. (1965). A simplex method for function minimization.
The Computer Journal, 7, 308–313.
Neumaier, A., Shcherbina, O., Huyer, W., & Vinko, T. (2005). A comparison
of complete global optimization solvers. Mathematical Programming, 103,
335–356.
Neurock, M., Nigam, A., Trauth, D., & Klein, M. T. (1994). Molecular
representation of complex hydrocarbon feedstocks through efficient char-
acterization and stochastic algorithms. Chemical Engineering Science, 49,
4153.
Quann, R. J., & Jaffe, S. B. (1996). Building useful models for complex reaction
systems in petroleum refining. Chemical Engineering Science, 51, 1615.
Reyniers, M. F., & Froment, G. F. (1995). Influence of Metal surface and sul-
fur addition on coke deposition in the thermal cracking of hydrocarbons.
Industrial & Engineering Chemistry Research, 34, 773–785.
Rosenbrock, R. R. (1960). An automatic method for finding the greatest or least
value of a function. The Computer Journal, 3, 175–184.
Ryoo, H. S., & Sahinidis, N. V. (1996). A branch-and-reduce approach to global
optimization. Journal of Global Optimization, 8, 107–139.
Shannon, C. E. (1948). A mathematical theory of communications. The Bell
System Technical Journal, 27, 379.
Schwefel, H. P. (1981). Numerical optimization of computer models. Chichester:
Wiley.
Van Damme, P. S., & Froment, G. F. (1982). Putting computers to
work—Thermal cracking computer control in pilot plants. Chemical Engi-
neering Program, 78, 77–82.
Van Geem, K. M., Heynderickx, G. J., & Marin, G. B. (2004). A comparison
of one and two-dimensional reactor models for steam cracking: Effect on
yields and coking rate. AIChE Journal, 50, 173–183.
Van Geem, K. M., Reyniers, M. F., & Marin, G. B. (2005). Two severity indices
for scale-up of steam cracking coils. Industrial & Engineering Chemistry
Research, 44, 3402–3411.
Verstraete, J. J., Revellin, N., Dulot, H., & Hudebine, D. (2004). Molecular recon-
struction of vacuum gas oils. Preparations of Paper—American Chemical
Society Division of Fuel Chemistry, 49(1), 20–21.
Wauters, S., & Marin, G. B. (2002). Kinetic modeling of coke formation dur-
ing steam cracking. Industrial & Engineering Chemistry Research, 41,
2379–2391.
Wahl, F., Hudebine, D. & Verstraete, J. J. (submitted for publication). Recon-
struction of the molecular composition of FCC Gasolines from overall
petroleum analyses. Oil and Gas Science and Technology—Review IFP.