Computers and Mathematics with Applications 65 (2013) 1674–1697

Contents lists available at SciVerse ScienceDirect

Computers and Mathematics with Applications

journal homepage: www.elsevier.com/locate/camwa

Complexities in modeling of heterogeneous catalytic

reactions
Frerich J. Keil ∗
Hamburg University of Technology, Institute of Chemical Reaction Engineering, Eissendorfer Str. 38, 21073 Hamburg, Germany

article info abstract

Keywords: Catalysts are omnipresent in the chemical industry. Understanding of catalytic chemical
Catalysis
reactions is crucial for a rational development of catalysts. The present paper describes
Pore structure
approaches for simulating phenomena in heterogeneous catalysis on a molecular level.
Adsorption
Diffusion Modeling of porous structures and their fractal surfaces will be presented. Simulation of
Reaction adsorption and diffusion of reactants and products inside the pores by means of various
Multiscaling Monte Carlo and Molecular Dynamics algorithms is described, followed by quantum
chemical methods for calculating reactions on the active centers. Optimization of pore
structures and multiscaling procedures complete the picture.
© 2012 Elsevier Ltd. All rights reserved.

1. Introduction

Most chemical products are synthesized by means of catalysts. They are used in about 90% of all chemical processes
worldwide [1]. Fields of catalyst application are refining, chemicals, polymerization, and environmental protection. Catalyst
manufacturing accounts for many billion dollars in sales, but the value derived from products based on catalytic processes
is many times higher. What is a catalyst? Catalysts are materials that accelerate reactions and convert reactants preferably
to desired products without being consumed themselves. After a catalytic cycle, the catalyst is restored to its initial state.
A catalytic cycle comprises adsorption of reactants, often additionally dissociation of the adsorbed molecules, reaction, and
desorption of products from the catalyst surface. The ideal catalyst should convert the reactants to the desired products
whereby no undesired by-products are formed. In practice, these ideal requirements cannot be fully met, but one has to
find materials that will come as close to optimal performance as possible. At present, catalyst development is dominated
by experimental work. New catalysts are very rarely found, mostly by accident. The majority of experiments are directed
towards optimization of existing catalysts with respect to given criteria.
Various spectroscopic methods [2–5] and simulation tools [6–24] can already give a deep insight into the details
of catalytic reaction mechanisms. The majority of industrial catalysts contain an active component in the form of
nanoparticles <20 (nm) in size that are dispersed onto high-surface catalyst supports. The activity and selectivity are
affected by the local size, shape, crystal face and chemical composition of catalyst particles employed. The electronic and
thermodynamic interactions of the catalyst nanoparticles with their support (for example, alumina (Al2 O3 ), silica (SiO2 ))
are also important. In so-called single-site catalysts just one or very few atoms are the catalytic active clusters. The ultimate
goal is understanding catalytic phenomena occurring on nanomaterials to a point where a rational design of new catalysts is
possible. Here, we confine the considerations to porous heterogeneous catalysts. By means of Fig. 1 the problems a catalyst
modeler has to deal with will be explained.
The catalytic active site, the single atoms or nanoparticles, are inside the pore network of the catalyst support. The
supports are piled up in tubular reactors. These supports can be crystals with a well-defined pore structure or amorphous

∗ Tel.: +49 40 42878 3042; fax: +49 40 42878 2145.

E-mail address: [email protected].

0898-1221/$ – see front matter © 2012 Elsevier Ltd. All rights reserved.
doi:10.1016/j.camwa.2012.11.023
 F.J. Keil / Computers and Mathematics with Applications 65 (2013) 1674–1697 1675

Fig. 1. (a) Phenomena on catalytic surfaces and inside porous catalyst supports; (b) Knudsen diffusion; (c) molecular diffusion; (d) surface diffusion; (e)
single-file diffusion.

materials. The pore walls inside amorphous supports are often of multifractal structure. Therefore, one has to describe
the porous structure as the reactants are transported from the bulk gas phase to the active centers inside the pores, and
the products diffusing back to the bulk gas. The next problem is the description of multicomponent transport phenomena
inside pores as the total reaction time is not only determined by the intrinsic reaction at the active sites but also by the time
for gas transport inside the support. Additionally, diffusion controls the local concentration of the reactants at the active
sites. As can be realized from Fig. 1(b)–(e), diffusion phenomena inside pores may be classified into Knudsen diffusion (only
collisions with the pore walls), molecular diffusion (only molecular–molecular collisions), surface diffusion, and single-
file diffusion (the pores are so narrow that molecules cannot bypass each other). Sometimes pore entrance effects can be
important. As molecules, in particular at low pressure in the bulk gas phase, adsorb at the pore walls, one has also to model
multicomponent adsorption. The next important task is the calculation of the reactions itself. Chemical reactions are genuine
quantum chemical occurrences. Here, the wide experiences of quantum chemical calculations come into play [7–16,25–28].
As these calculations result in data of observable measurements at zero Kelvin, a further tool, statistical mechanics, has to be
introduced [23,29–32] to convert these data to real catalytic operating conditions. Another reason for employing statistical
mechanics approaches is that not only single molecules are involved in catalysis but molecule ensembles. The quantum
chemical calculations give results for the electronic energies, molecular structures (bond lengths, angles), electronic density
distributions and spectroscopic properties. These data are fundamentals for calculations of rate constants and give insights
into electronic processes during bond cleavages and bond formations. The rate constants are calculated by means of various
approaches of transition state theory [33,34]. Dynamic reaction phenomena may be simulated by quasi-classical trajectory
Molecular Dynamics (QCT-MD) [35,36] or Car–Parinello approaches [37,38]. Classical molecular simulation methods, like
Monte Carlo (MC) [17–19] and/or Molecular Dynamics (MD) [20–22] are suitable for converting data obtained from quantum
chemical calculations (0 (K)) into real operating conditions of catalysts. Additionally, MC approaches may be employed
for multicomponent adsorption simulations and other properties at thermodynamic equilibrium conditions. Molecular
dynamics is the standard approach for diffusivity calculations. Reactive fluid flow approaches, like computational fluid
dynamics [39], are the proper tool for calculating the flow of the bulk gas phase. Finally, the entire chemical reactor is
modeled using ordinary or partial differential equations which result in composition, temperature, and pressure profiles
along a tubular reactor [40]. The short compilation of the integral parts of heterogeneous catalytic reactor models reveals its
large scope and complexity. In the following paragraphs some of important developments in modeling catalytic phenomena
will be discussed. As several thousand papers have been issued on this subject over the last few years, only a small part can
be included in this review, but references to other reviews in related fields of research will be given.
1676 F.J. Keil / Computers and Mathematics with Applications 65 (2013) 1674–1697

Fig. 2. (a) Zeolite ZSM-5; (b) structure of MOFs.

Fig. 3. Structure of amorphous catalyst supports.

Fig. 4. (a) Parallel straight pores; (b) curved pores; (c) model by Johnson and Stewart.

2. Modeling of porous structures

The catalytic active centers are placed inside porous supports which are either crystalline or amorphous. The crystalline
supports are mostly zeolites or metal–organic frameworks (MOFs). Zeolites are microporous aluminosilicate minerals which
have the ability to selectively sort molecules based on a size exclusion process (shape selectivity). This property is due to a
very regular pore structure of molecular dimensions. An example of a zeolite pore structure (ZSM-5) is presented in Fig. 2(a).
The maximum size of the molecules that can enter the pores of a zeolite is controlled by the dimensions of the channels.
At present about two hundred unique zeolite frameworks have been identified, and more than forty naturally occurring
zeolite frameworks are known [41,42]. Zeolites can accommodate a wide variety of cations, such as K+ , Na+ , Ca2+ , Mg2+ ,
and many others. They are used as adsorbents and catalyst supports [43]. Their crystal structures are available from
databases [41]. Therefore, the description of zeolite pore structures is determined by their crystallographic data. The same
holds for metal–organic frameworks (MOFs). MOFs are composed of metal ions or clusters of metal ions and organic linker
molecules which are typically mono-, di-, tri-, or tetravalent ligands (Fig. 2(b)) [44]. The choice of metals and linkers has
significant effects on the structure and properties of the MOF. The metals in the MOF structure often act as Lewis acids. The
linkers can also act as catalytic sites [43,45,46]. Additionally, MOFs are used as adsorbers [45,47]. Their porous structure is
also obtained from crystallographic data.
More difficult is the modeling of porous structures of amorphous catalyst supports, like alumina (Al2 O3 ) or silica (SiO2 ).
Electron-microscopical pictures show three-dimensional cavities of arbitrary structure with rough wall surfaces which
resemble cauliflower. Catalyst supports are made from microporous powders which are pelletized to spheres or cylinders,
amongst other shapes. The interspaces between the powder particles are macroporous (Fig. 3). Simple models are parallel
straight (Fig. 4(a)) or curved (Fig. 4(b)) pores [48].
A more realistic model was introduced by Johnson and Stewart [49] which employs randomly oriented pores (Fig. 4(c)).
Beeckman and Froment [50] and Reyes and Jensen [51] have introduced Bethe lattices (Fig. 5(a)/(b)) [52] which are cycle-free
graphs where each node is connected to z neighbors. The term z denotes the coordination number.
As can be seen from Fig. 5(a)/(b), Bethe lattices are tree-like structured emanating from a central node, with all nodes
arranged in shells around the central node. The number of nodes in the kth shell is given by Nk = z (z − 1)k−1 (k > 0). The
 F.J. Keil / Computers and Mathematics with Applications 65 (2013) 1674–1697 1677

Fig. 5. (a) Bethe lattice with coordination number z = 3; (b) corresponding pore structure; (c) Voronoi polyhedra micro/macro networks.

Fig. 6. Three-dimensional pore structures, (a) macropores span the entire pellet; (b) pores are partially blocked.

simple models cannot describe the so-called connectivity of pores which corresponds to the average coordination number
of the nodes. Furthermore, these models have no closed loops of the pores. A totally random topology can be described by
Voronoi polyhedra (Fig. 5(c)) [53,54]. Voronoi polyhedra allow the creation of isotropic networks by subdividing the space
by a so-called Voronoi tessellation. For that purpose, a set of points placed at random in space is enclosed with the smallest
polygon that can be formed by intersecting planes that bisect the lines connecting the point and its neighbor.
Closer to reality come three-dimensional pore networks (Fig. 6). Fig. 6(a) presents a macro/micropore network whereby
the macropores span the entire pellet, in accordance with a real catalyst support (see Fig. 3(b)).
Random three-dimensional networks are advantageous in several respects [55]:
– Any type of network can be modeled (e.g. regular, irregular, Voronoi).
– The effect of pore connectivity can be taken into account.
– Any pore size distribution and any pore geometry (e.g. cylindrical, slit-like) can be used.
– Local heterogeneities, for example, spatial variation in mean pore size, can be modeled.
– Any distribution of catalytic active centers may be taken into account.
– The pore walls can be smooth, irregular, or fractal.
– Commonly used fitting parameters, like tortuosity, are avoided.
– Percolation phenomena [56] can be described. This is particularly useful for modeling deactivation phenomena which
are accompanied by temporal changes of pore structure, like hydrodesulfurization or hydrodemetallation of natural oil
[57,58].
Jerauld et al. [59,60] and Winterfeld et al. [61] have shown that as long as the average coordination number of a
topologically disordered system is equal to the coordination number of a regular network, the transport properties of the
systems are essentially identical. With respect to three-dimensional networks one has to link the porous structure of the
catalyst support and the network model. That means one has to find out textural parameters of the support. Standard
procedures are mercury porosimetry and nitrogen adsorption [62–66] which give pore radii distributions, total surface
area, and connectivities [67,68]. Nuclear magnetic resonance (NMR) [69] and small-angle X-ray scattering (SAXS) [70] may
be also used for exploring pore structures. The total pore volume (void fraction) has to be measured also. Based on these
data, networks are designed which show the same pore radii distribution and pore connectivity as the real catalyst support.
The pore radii distribution is generated by using a probability density function (pore radii distribution function) f (r ) and a
random number ρi (0 ≤ ρi ≤ 1) which result in the radius Ri of an arbitrary pore i as a solution of the equation
 Ri
ρi = f (r )dr . (1)
0

An important problem is the description of the pore wall structure. For crystalline materials the atom locations are well-
defined and known from crystallographic data. Pore walls of amorphous catalyst supports initially were modeled as smooth
tubes. The main reason for this situation has been the nature of the extremely complex geometry of catalyst support surfaces.
The consequence is that it becomes exceedingly difficult to study geometry–efficiency relations, both phenomenologically
1678 F.J. Keil / Computers and Mathematics with Applications 65 (2013) 1674–1697

Fig. 7. (a) Menger sponge; (b) fractal catalytic surface.

and in a predictive fashion. Progress has been made after Mandelbrot [71] has introduced the concept of fractals [72–74].
Fractals are self-similar patterns, that means they have no characteristic length scale. Mandelbrot stated that fractals are ‘‘a
rough or fragmented geometric shape that can be split into parts, each of which is (at least approximately) a reduced-size
copy of the whole’’ [74]. An example of a self-similar object, a Menger sponge, is presented in Fig. 7(a). A fractal catalytic
surface is given in Fig. 7(b) [75].
Farin and Avnir [76,77] have found for many catalysts a relation a = RDR between catalytic activity a (mol · time−1 ·
particle−1 ), and particle size, 2R, of dispersed catalysts, whereby DR is the reaction dimensionality and R is the radius of the
particle. Avnir [77] gives the following reasons for fractal properties of catalytic support surfaces:
– Most values of surface dimensions Da , allude to a fractal geometry, as values 2 ≤ Da ≤ 3 have been found.
– For materials that could be expected to be plane, values of Da = 2 ± 0.15 have been found.
– Graphite has a value of Da = 2, whilst microporous carbon has a value of Da = 3. Heating graphite under anaerobic
conditions, one observes a continuous transition from Da = 2 to Da = 3.
– The fractal dimensions can be measured by various approaches which result in nearly the same values for Da .
– Evaluations of scanning tunneling microscope (STM) images show a fractal dimension.
A random walk on fractal surfaces [78] gives the following expression for the mean square displacement

r 2 (t ) ∝ t 2/DW ,
 
(2)
whereby DW is the fractal dimension of the distance covered by the random walker. The influence of fractal surfaces
on the performance of catalysts has been described, for example, by Gutfraind et al. [79]. The effects of catalyst surface
morphology and surface active-sites distribution on diffusion limited reactions were studied by these authors. The
complex distribution of reaction probabilities of the active sites were analyzed by multifractal approaches. Specific scaling
assumptions of the multifractal formalism could be corroborated. Systematic experimental investigations on fractality of
chemical reactions were executed by Kopelman [80]. Classical reaction kinetics has been found to be unsatisfactory when
the reactants are spatially constrained on the microscopic level by either walls, phase boundaries, or force fields [80].
Among the practical examples of ‘‘fractal-like kinetics’’ are chemical reactions in pores of membranes, excitation trapping
in molecular aggregates, excitation fusion in composite materials, and charge recombination in colloids and clouds.
Coppens [81] discussed the effect of fractal surface roughness on Knudsen diffusion. An analytical expression for the
Knudsen diffusivity has been derived and the residence time distribution of the molecules is obtained from Monte Carlo
simulations. The equations for diffusion and reaction in fractal pores and in porous catalysts with a fractal internal surface
were given. Practical examples, like naphtha reforming, illustrated the significant influence of fractal roughness on product
yields and distributions of industrial processes, from the microscale up to the observable reactor scale [81]. Guo and
Keil [82] investigated the kinetics of N2 O catalytic decomposition over three-dimensional fractals by a Monte Carlo random
walk algorithm. The simulation results have shown that usual kinetic equations based on the mean field theory [83] are
not applicable to the reaction over fractals, and a different rate law containing a fractional reaction order is more suitable
for describing the reaction kinetics over fractals. The mean-field approximation assumes that all adsorbed species are
distributed randomly over the surface, and there is no interaction between adsorbed species. This approximation is seldom
fulfilled. For low coverages and repulsive interactions between the adsorbed molecules this approximation works usually
well. At high coverages, adsorbate interactions are always present, such that pre-exponential factors and activation energies
are coverage dependent [83].
The deactivation of catalysts due to sintering or deposition of coke or metals can be described by percolation
theory [56,84]. Applications of percolation theory are outlined by Sahimi [85]. Percolation theory shows when a system
is macroscopically susceptible to a given phenomenon. For example, it indicates when there is a flow of molecules possible
from one side of a catalyst pellet to the opposite. For describing blockage phenomena in porous materials bond percolation
is useful (Fig. 8). Initially all pores are open (Fig. 8(a)).
Owing to deposition of coke or metal (hydrodemetallation) the pores are blocked stepwise in the course of chemical
reactions. At a certain point, the so-called percolation threshold, the connection between one side of the network and
 F.J. Keil / Computers and Mathematics with Applications 65 (2013) 1674–1697 1679

Fig. 8. Bond percolation (a) all bonds are open; (b) some bonds are closed; (c) only isolated clusters of open bonds exist.

Fig. 9. (a) Normalized accessible pore volume vs. time, ten network realizations; (b) normalized pore volume distribution density vs. pore radius, averaged
over ten network realizations.

the other gets interrupted. From there on only isolated open clusters exist (Fig. 8(c)). Percolation thresholds and cluster
counting on large lattices can be executed numerically by an algorithm developed by Hoshen and Kopelman [86,56]. For
very large lattices it gives the distribution of open cluster sizes and the percolation threshold. This algorithm has been used,
for example, by Rieckmann et al. [57] carrying out a detailed analysis of geometric parameters of a catalyst pellet for a
hydrodemetallation reaction. A three-dimensional network of interconnected cylindrical pores was taken as a pore model.
The change of the pore space with time was analyzed by the Hoshen–Kopelman algorithm [86]. It turned out that the reaction
continues for a long time after the percolation threshold has passed. The inner pore volume accessible from the pore surface
is the relevant parameter. A complicated interplay of reaction kinetics and diffusion determine the accessibility of the pore
space. In Fig. 9(a), the normalized accessible pore volume vs. time for ten network realizations is presented. The decrease
of the mole fluxes during the first few days corresponds to a decrease of the accessible pore space of 10%. After this initial
decrease the mole fluxes and the accessible pore volume remain nearly constant for a long time. They drop to zero, as soon as
the pore mouths of the macropores get blocked. Fig. 9(b) shows the change of the pore radii distribution with time averaged
over the network realizations. One observes the stepwise reduction of the pore volume.
After having modeled the pore structure, the next steps are multicomponent adsorption and diffusion of reactants and
products inside the pores.

3. Adsorption and diffusion in catalyst supports

3.1. Adsorption

Reactants adsorb at the outer pore mouths of catalyst supports. The material in the adsorbed state is defined as the
‘‘adsorbate’’ and that in the bulk phase the ‘‘adsorptive’’. Adsorption can result from van der Waals forces (physisorption) or
from far stronger chemical interactions (chemisorption). For chemical reactions at least one of the reactants has to adsorb
strongly at the active site. Adsorption leads to a decrease in entropy and free energy of the adsorbed fluids compared to the
bulk phase gases owing to a reduction of translatory, and partially rotatory, motion. Adsorption is an exothermic process.
In microporous materials, like, for example, zeolites, all atoms of the adsorbent can interact with the adsorbates. After the
reactants have adsorbed at the pore mouths, they diffuse into the pores towards the active centers. An important concept
is the ‘‘adsorption isotherm’’ which is the equilibrium relation between the amount of adsorbed materials and the pressure
or concentration in the bulk phase. Various empirical adsorption isotherms have been developed. A compilation of well-
known isotherms is given in a book by Do [87]. The first researcher who introduced a clear concept of the monomolecular
adsorption and derived his now famous isotherm was Irving Langmuir [88]:
k·P
θ= , (3)
1 + kP
1680 F.J. Keil / Computers and Mathematics with Applications 65 (2013) 1674–1697

whereby θ is the fractional coverage of the surface, P is the gas pressure above the surface, and k is a constant. This equation
can be derived in various ways [89–92]. Besides the many empirical isotherms, statistical thermodynamics and molecular
modeling approaches are in use [17,30,31,90] for calculating adsorption isotherms. In equilibrium the following conditions
hold:
µbulk = µads ; Tbulk = Tads , (4)
whereby µ is the chemical potential and T the temperature. Additionally, the volume is fixed. Norman and Filinov [93]
have developed a Monte Carlo approach in the grand canonical ensemble (µVT ) for calculating adsorption isotherms. One
gives the chemical potential and temperature inside the pores, whereby the number of particles inside the pores fluctuates
during the simulation till the first condition in Eq. (4) is fulfilled. Via an equation of state one can calculate the pressure
corresponding to the given chemical potential:
P = {exp(µ/kB T )kB T }/Λ3 , (5)
whereby kB is the Boltzmann constant, and Λ the de Broglie thermal wavelength

Λ= h2 /(2π mkB T ) (6)
h is the Planck constant.
The chemical potential is usually determined using an approach by Widom [94]. The excess chemical potential is the
difference between the actual value and that of the equivalent ideal gas system (µ = kB T ln((N /V )Λ3 )):
µexcess = −kB T ln exp[−U (r test )/(kB T )] .
 
(7)
U (r test ) is the potential energy of the inserted test particle. The brackets ⟨⟩ indicate an average value.
The bulk gas phase is considered as a reservoir. In a grand canonical Monte Carlo (GCMC) simulation, three types of
moves are performed, namely displacement of particles, insertion of a particle, and removal of a particle. For the first case
a particle is selected at random and given a new position. For all three cases suitable acceptance rules are introduced [17].
Although the GCMC approach is useful for atoms and simple molecules, one runs into problems for complex molecules like
long alkanes/alkenes, polymers and molecules with side chains, because the probability of achieving a successful insertion
or removal step is often very small. Insertion steps fail because the fluid is so dense that it is difficult to insert a new
particle without causing overlaps with neighboring particles. Removal fails because the particles in a fluid or narrow pore
experience significant attractive interactions. A solution of this problem has been the ‘‘Configurational Bias Monte Carlo’’
scheme (CBMC) originally introduced by Siepmann [95] and improved later on in various papers [96,97]. The CBMC scheme
is based upon the Rosenbluth sampling algorithm [17,98]. The essence of the CBMC method is that a growing molecule is
preferentially directed (biased) towards acceptable structures. The biases have to be removed by suitable modifications of
the acceptance rules. The CBMC approach can be applied to lattice and non-lattice models [99]. The molecule chains are
inserted bead by bead into the voids of the porous material. The CBMC approach has been demonstrated being useful in
very many applications. An overview of applications is given, for example, by Smit and Maesen [100] and Keil [101].
Classical molecular simulation methods need as most important inputs inter- and intramolecular potentials describing
the interactions between the catalyst supports and the atoms or molecules of the adsorbing reactants. The accuracy of
the simulation results is primarily determined by the potentials. In principle one can calculate these potentials from
quantum chemical calculations, but often the computing time is prohibitive for these applications, in particular for long-
chained molecules. Therefore, empirical force field expressions, which have adaptable constants, are fitted to measured
quantities, like adsorption isotherms [102]. The term ‘‘force field’’ in molecular simulation refers to the mathematical form
and parameters of functions employed to describe the potential energy of an ensemble of particles. Typical isotherms for
n-heptane in MFI at various temperatures are presented in Fig. 10, which reveal inflection points.
This phenomenon has been explained by Smit and Maesen [106] in terms of ‘‘commensurate freezing’’, that means n-
heptane has a length commensurate with the zig–zag channel of MFI (see Fig. 2(a)). At high pressures the molecules are
exclusively localized in the channels and not in the intersections, leaving the straight channels free for further filling of the
pores which leads to a step in the isotherm.
The mathematical expression of the force field leading to the isotherms in Fig. 10 is split into an internal and external
term of the total potential energy U:
U = U int + U ext . (8)
The internal part splits into a bond, a bending, and a torsion contribution:
 1
U bond = k1 (r − r0 )2 (9)
bonds
2
 1
U bend = k2 (cos θ − cos θ0 )2 (10)
bends
2
n =5
 
U torsion = ηn cosn φ (11)
torsions n=0
 F.J. Keil / Computers and Mathematics with Applications 65 (2013) 1674–1697 1681

Fig. 10. Isotherms of n-heptane in MFI at various temperatures [103].

Source: Experimental data are taken from Sun et al. [104] and Eder [105].

Fig. 11. Lennard-Jones potential.

k1 and k2 are bond energy and bend energy constants, respectively. r0 and θ0 are the reference bond length and bend angle,
respectively. φ is the dihedral angle, and the ηn are the six torsion parameters.
The external energy, U ext , consists of a molecule–molecule intermolecular energy U mm , a catalyst support (here: zeolite)
molecule interaction energy U zm , and an intramolecular Lennard-Jones interaction energy, U intra , for beads in a chain
separated by more than three bonds:

U ext = Uijmm + Uijzm + Uijintra (12)

with
 6 
σij σij
12 
mm,zm,intra

Uij = 4εij − − Ecut (13)
LJ-pairs
rij rij

mm,zm,intra
where rij is the distance between site i and site j, Ecut is the energy at the cutoff radius rcut , whereby Uij = 0
when rij > rcut . The Lennard-Jones potential falls off very rapidly with distance. At 2.5 σ the LJ potential has just 1% of its
value at σ . Therefore, for non-bonded interactions one introduces a cutoff radius where the interactions between all pairs
of atoms that are further apart than the cutoff value are set to zero. One can correct for the systematic error by adding a tail
contribution to the potential [17]. Switching functions are used for switching the force smoothly off at the cutoff radius. The
Lennard-Jones potential consists of two parameters, σ is the size parameter, and ε is the strength parameter (Fig. 11). The
parameters in the potential expressions are iteratively fitted to experimental data [103] or, if possible, are obtained from ab
initio calculations [102].
Parametrization of a force field is not a trivial task. Depending on the properties that should be reproduced by the
empirical force field, one fits the parameters to thermodynamic data, vibrational frequencies, equilibrium conformations,
adsorption isotherms, transport properties, etc. Numerical least-squares techniques are employed for fitting the parameters
to the experimental data. Force fields should satisfy the following requirements: robustness, accuracy, simplicity, and
1682 F.J. Keil / Computers and Mathematics with Applications 65 (2013) 1674–1697

Fig. 12. (a) Adsorption isotherms of 1-butene and ethene by means of molecular simulation and IAST [108]; (b) fugacities of ethene and 1-butene as a
function of pressure [109].

transferability. Robustness means that force fields should be valid in varying environments. For example, phase behavior
can only be described correctly if the same model represents all phases of the same quality. The accuracy of a force field
should be such that experimental data can be reproduced with required precision. The simplicity requirement means that
force fields should include as few terms as possible. Transferability is the requirement that the underlying principles of
constructing the model and the parameters obtained should be applicable for various molecular problems. It is often the
non-additivity of constituent terms, and the omission of important contributions, that makes the terms non-transferable.
Since most force fields contain parameters adjusted to measurements, one obtains effective interaction parameters, that
means a deviation or omission in one term is compensated by other terms, which are then not precise enough in different
molecular environments or various physical conditions. The dominant and most relevant omission in the usual force fields
is the incorporation of electronic polarizability, which is a non-additive electrical interaction.
Multicomponent adsorption Monte Carlo simulations are extremely time consuming as for each bulk reactant
composition one has to calculate the isotherms. Fortunately, there is an approach available which allows one to calculate
mixture isotherms from pure component isotherms. This approach is the ‘‘Ideal Adsorption Solution Theory’’ (IAST) [107].
An example is given in Fig. 12(a)/(b).
Simulated adsorption isotherms of a binary mixture of ethene and 1-butene for a fixed equimolar bulk phase composition
at 400 (K) are presented based on molecular simulations (symbols) and IAST. As can be seen from Fig. 12(a) a maximum for
1-butene is obtained [108]. At higher pressures less 1-butene is adsorbed but the total amount of adsorbed gases increases
monotonically with increasing pressure. This ‘‘squeezing out’’ effect has also been observed for other mixtures. The reason
for this is an entropy effect, that means the packing density is higher for the smaller ethene. Interestingly this ‘‘squeezing out’’
effect is also calculated by IAST. Another reason for a maximum can be found in nonideality effects [109]. Fig. 12(b) shows
the fugacities of ethene and 1-butene as a function of pressure in an equimolar mixture. The slope of the 1-butene fugacity
decreases significantly for higher pressures in contrast to the ethene fugacity. As has been shown by Heyden et al. [108], a
difference in the component fugacities can lead to a maximum in the isotherm, because, in addition to the potential energies,
exerted by the molecules, the ratio of the fugacities decides which component ‘‘wins the competition’’ inside the pores.
Some remarks should be made about adsorption phenomena at the active center or close to it [16]. At least one of the
reactants in a catalytic reaction has to dissociate. A catalytic reaction is initiated when adsorbed molecular fragments are
produced on the surface as a result of dissociation. In a simplified manner the approach of a molecule to a metal surface
can be divided into three cases: adsorption via a precursor, direct adsorption, or reflexion (see Fig. 13(a)). In any case the
impinging molecule has to dissipate its kinetic energy to the catalytic nanoparticle and the catalyst support. In the so-called
dynamical precursor state [110], the molecule does not have sufficient kinetic energy perpendicular to the surface to scatter
back in the gas phase. It bounces back and forth along the surface whereby it dissipates energy to the catalyst surface until
it accommodates at a site. If the kinetic energy of the incoming molecule is low, a direct adsorption may be possible. In
this case the molecule often is steered to the attractive paths toward the chemisorption states. For high kinetic energies a
reflexion of the incoming molecule may occur. The adsorption process is also influenced by precovered surfaces. It has been
demonstrated experimentally that cold H2 molecules impinging on an almost completely hydrogen-covered Pd(111) do not
adsorb dissociatively in a hydrogen dimer vacancy, but three or more vacancies are required to dissociate hydrogen [111].
The presence of hydrogen leads to formation of activation barriers [112].
Fig. 13(b) presents the trajectories of hydrogen atoms on a Pd(100) surface after dissociation of a hydrogen
molecule [113]. The trajectories pass along several surface sites before they come to rest. The hydrogen atoms approach
each other again after an initial increase in the interatomic distance before they separate again. This example demonstrates
that the dynamics of adsorption in catalysis is a very complicated process. Investigations of dynamic processes are still in
their infancy.
In general the ability of a material to adsorb gases varies with whether the material is a metal, semiconductor, or insulator,
and the position of the material within the periodic table. Quite often one finds that metals in the middle of the periodic
table are most reactive for adsorption of some gases. If one goes to high enough temperatures and pressure, most molecules
 F.J. Keil / Computers and Mathematics with Applications 65 (2013) 1674–1697 1683

Fig. 13. (a) Adsorption on lattice sites; (b) trajectories of hydrogen atoms after dissociative adsorption [113].

will dissociate on transition metal surfaces, whereby some molecules dissociate easily, while others are more difficult to
dissociate. For dissociative adsorption the adsorbate has to cross the activation barrier to dissociation before desorption
occurs. A certain molecule can bind in a variety of different configurations, and the preferred configuration varies with
both metal and face. The species may rearrange to form a variety of different intermediates. Most adsorbates can adsorb on
multiple sites. Random site filling and ordered adsorption are possible. In the former case the surface sites do not fill up in any
particular order, whereas in the latter case the sites fill in such a way that the adsorbate forms an ordered structure on the
surface at moderate coverages. Generally, random site filling occurs when the interaction between adsorbed molecules on
adjacent sites is small with respect to kB T . Quite often one gets ordered adsorption when there is a mixture of attractive and
repulsive forces so that some sites are filled with higher probability while others are filled with reduced probability. If there
are strong attractive forces amongst the adsorbed molecules, cluster formation might occur. The adsorption of molecules
sometimes leads to a surface reconstruction.

3.2. Diffusion

The adsorbed reactants diffuse to the active centers. In mesoporous catalyst supports various diffusion mechanisms are
possible depending on the local pore structure and loading (see Fig. 1(b)–(d)). Knudsen diffusion assumes that the molecules
exclusively collide with the walls, but no molecule–molecule collision occurs [114]. The scattering at the wall takes place in
such a way that the molecules are scattered in every direction of the half-space with the same probability (cosine law). A
derivation of the cosine law may be found in Ref. [115]. The Knudsen diffusivity of component i may be calculated according
to:

DiK = 9.7 · 103 r T /Mi [cm2 /s].


(14)
The term r is the effective pore radius in centimeter, T is the absolute temperature, and Mi is the molecular mass of
component i. As can be realized from Eq. (14), there is neither an expression about the details of the structure of the diffusing
molecules nor about the geometric structure of the pore walls. As expected, Eq. (14) cannot give accurate results in all cases.
A modification of the original equation has, for example, been given by Jakobtorweihen et al. [116] (see also [117]). Molecular
diffusion assumes molecule–molecule scattering, but no scattering at the walls. A well-founded theory of molecular diffusion
was developed by Maxwell [118] and Boltzmann [119]. Simplified equations, which are accurate up to 2%, are listed in a book
by Poling et al. [120]. As an example the formula by Wilke and Lee for binary diffusion coefficients will be given:
√
3.03 − (0.98/ MAB )(10−3 )T 3/2
DAB √ [cm2 /s]. (15)
P MAB 2σAB
2
ΩD
MAB is the averaged molar mass of the components A and B, with MAB = 2[(1/MA ) + (1/MB )]−1 ; T is the absolute
temperature, P is the pressure, σAB represents an averaged Lennard-Jones parameter, and ΩD is a collision integral.
The transition between Knudsen diffusion and molecular diffusion can be described by the Bosanquet equation [121].
This equation holds only approximately.
A survey of surface diffusion was presented by Doll and Voter [122] and Kapoor et al. [123]. Krishna [124] developed
an equation for surface diffusion in the framework of Maxwell–Stefan equations [125]. Surface diffusion is a hopping
process where the molecules hop from one adsorption site to another. It is an activated process, such that the surface
diffusivities show an exponential behavior with regard to the temperature. Multicomponent diffusion in porous media
has been described very often by means of the dusty-gas model (DGM) [126]. This model is based on kinetic theory of
gases. However, Chapman–Enskog solutions [127] of the Boltzmann equation do not have to be repeated all over again in a
somewhat different context. Thus, thermal diffusion, pressure diffusion, and forced diffusion are all easily included into the
1684 F.J. Keil / Computers and Mathematics with Applications 65 (2013) 1674–1697

Fig. 14. Dusty-gas model.

DGM. The porous medium is treated as an array of heavy (dust) molecules that are motionless and uniformly distributed
in space (Fig. 14). By treating the dust particles as giant molecules, results of the Chapman–Enskog kinetic theory can be
employed. The DGM comprises two independent parts: a diffusion part, consisting of a set of Stefan–Maxwell equations,
and a viscous flow part, consisting of an equation of motion for the gas mixture as a whole. The geometric structure of the
medium is isolated as an independent problem. Kerkhof and Geboers [128,129] investigated the DGM and found a number of
inconsistencies, for example, the ‘‘double counting’’ of viscous contributions. These authors have developed a binary fraction
model [128,129] which avoids inconsistencies. Young and Todd [130] investigated the foundations of diffusion in porous
media in a very clear way. These authors follow Kerkhof’s arguments and reject the DGM as a suitable foundation on which
to construct a viable theory. A particular problem is the transition regime between Knudsen and molecular diffusion. Young
and Todd suggested an interpolation of flow at arbitrary Knudsen number, which is the ratio of the molecular mean free path
length to the pore radius. The porous structure has been described by a porosity–tortuosity factor which was not included
into the diffusivities like in previous models, but in a separate equation. Young and Todd called their model ‘‘cylindrical pore
interpolation model’’ (CPIM). The present author follows Kerkhof’s arguments, in particular because he could demonstrate
the deficiencies of the DGM.
Based on a pore network model (see Fig. 6(a)) Rieckmann and Keil [131] have developed a multicomponent diffusion and
reaction model which includes all mentioned mechanisms of diffusion and arbitrary reaction kinetic expressions. For each
pore of the network the following differential equation was set up:

dJ(c) 2
− ν r(c, T ) = 0. (16)
dw rp
The term J comprises the diffusive fluxes (molecular, Knudsen, surface) and viscous fluxes, w is the pore length
coordinate, c is the vector of component concentrations, ν is the stoichiometric matrix, rp is the pore radius, r represents
the system of kinetic equations, and T is the absolute temperature. For the inner nodes Kirchhoff’s law (total inflow equals
total outflow) holds:

Jk · rp2,k = 0; k ∈ Π. (17)
k

The term Π represents the set of pores. At the outer boundary layer the following fluxes are found:
J = β(cb − cs ), (18)
whereby cb represents the vector of bulk gas phase concentrations, cs is the vector of surface concentrations, and β is a
transport coefficient.
The entire pore network leads to a large system of non-linear equations, depending on the number of nodes, which was
solved numerically using a finite difference scheme and a Schur-complement–Newton approach [132,133]. This approach
needs only minor storage space. The linear subproblem was solved by a direct method for sparse matrices [134]. The
convergence was improved by homotopy techniques. Finally, one obtains the concentration profiles of all components inside
the pore network.
Diffusion in nanoporous materials, for zeolites, carbon nanotubes (CNTs), MOFs, can often not be described by simple
equations like molecular diffusion or Knudsen diffusion, except for some cases. For example, Keipert and Baerns [135]
employed a temporal analysis of products reactor (TAP) system to determine the intracrystalline diffusion coefficients of
isobutane and 2-methylpentane in a H-ZSM-5 sample. The diffusion phenomena inside the zeolite could be described by
the Knudsen diffusivity equation. Delgado et al. [136] have developed a detailed transport model for a TAP-like system
(Multitrack setup) which allows analysis of adsorption, diffusion and catalysis phenomena. The effects of viscous flow
and thermal transpiration have also been analyzed. The Knudsen diffusion equation could successfully be used. Nijhuis
 F.J. Keil / Computers and Mathematics with Applications 65 (2013) 1674–1697 1685

et al. [137] used the Multitrack system for investigating adsorption and diffusion phenomena in microporous materials. The
diffusivities obtained based on the Knudsen equation were in good agreement with those determined using microscopic
techniques such as NMR. The pore radii of nanoporous materials are in the range of the sizes of the diffusing molecules.
Therefore, the molecules ‘‘feel’’ all the structural details of the pore walls, such that one has to revert to molecular simulation
methods. Molecular Dynamics (MD) [17,20–23,138,139] is the method of choice. This approach solves Newton’s equations
of motion by employing symplectic integrators, like the velocity Verlet algorithm [23,140,141]. An important feature of
Newton’s equations is that they can be viewed as a Hamiltonian system, i.e. they can be put in the form q̇ = ∂ H /∂ p and
ṗ = −∂ H /∂ q, where q denotes the position coordinates, p the momentum coordinates, and H (p, q) is the Hamiltonian.
The set of position and momentum coordinates (q, p) are called canonical coordinates. The time evolution of Hamilton’s
equations is a symplectomorphism, that means it conserves the symplectic two-form dpΛdq [142]. A numerical integrator
is a symplectic integrator if it also conserves this two-form. The standard Runge–Kutta algorithm is not symplectic. As
in MD simulations several hundred thousand time integration steps are run, a non-symplectic integrator would result in
a shift of the total energy of the system. This can be avoided by means of symplectic integrators. Like for Monte Carlo
calculations, MD simulations need accurate force fields [102,143]. A force field should be transferable between different
molecules, and valid for a wide range of environments and conditions. For this purpose one has to include the physically
important contributions to the potentials. In practice this can be a very difficult task. A list of common force fields is given
by Keil [144]. For calculating large reacting systems, so-called reactive force fields were developed [145]. The Reax FF force
field [145] uses the bond order/bond energy concept of Pauling [146] which describes the number of shared electrons in de
bond between two atoms, and is linked by a continuous function to the bond length. At short distances, this bond order/bond
distance relation approaches a maximum while at infinite distance the bond order goes to zero. Molecular simulations of
diffusion in zeolites and MOFs use quite often Kiselev-type potentials [147]. In a Kiselev-type model, the zeolite atoms are
kept fixed at their crystallographic positions. Interactions between the molecules and the zeolite are modeled by placing
Lennard-Jones sites and partial charges on all atoms of the framework and the sorbate molecules. A rigid lattice has the
advantage that one does not need a model for the zeolite–zeolite interactions. Furthermore, the computing time for flexible
lattices is larger by about two orders of magnitude. Demontis and Suffritti [148] reviewed force fields for flexible zeolites.
Investigations have shown that the effect of zeolite flexibility on the diffusion process is complex. For example, Zimmermann
et al. [149] have shown that for the same system, depending on which force field model is used for the zeolite interactions,
either an increase or a decrease of diffusion coefficients is observed. Further results of flexible zeolite lattices are reviewed
in Ref. [100].
Diffusion coefficients in zeolites may vary as much as ten orders of magnitude. The larger diffusion coefficients can be
computed directly from MD simulations. For the very slowly diffusing molecules so-called rare event simulations [100,150]
and kinetic Monte Carlo (kMC) schemes have been developed. In kMC a list of all possible process types is set up. A physical
time can be assigned to the simulation steps [151,152]. In order to make use of this advantage, one has to provide as input
the rates of all relevant individual processes. These data can be obtained from MD simulations or preferably from quantum
chemical calculations combined with transition state theory. The hopping rate changes with zeolite loading.
The MD simulations result in trajectories of all molecules of the system. These trajectories are then evaluated with respect
to certain properties. In equilibrium molecular dynamics (EMD) simulations the self-diffusion coefficient, Dsi , is obtained,
which is a measure of the average mean-squared displacement over all particles in direction x:
 Ni

1 d 
Dsi,x = lim (rji (t ) − rji (0))2 (19)
2Ni t →∞ dt j =1

where Ni is the number of molecules of component i, t is the time, and rji is the center of mass of molecule j of component
i. Eq. (19) is known as the Einstein equation. Alternatively, Dsi,x can be expressed by the Green–Kubo relation
N 
 ∞ i
1 
Dsi,x = v (t ) · v (0) dt
i
j
i
j (20)
Ni 0 j =1

which is the time integral of the velocity autocorrelation function. At very short time scales the mean-squared displacement
shows a quadratic dependence on time. This is called the ballistic regime, where particles on average do not yet collide.
When particles start colliding an intermediate regime comes up owing to confinement, as the particles collide only with
nearby particles. The diffusional regime is reached when particles move along the entire lattice. Then the mean-squared
displacement becomes linear with time. For y and z directions the same calculations can be done, such that the overall
self-diffusion coefficient is obtained from
Dsx + Dsy + Dsz
Ds = . (21)
3
For practical applications the transport diffusion or Fick coefficient is more important, which is related to the macroscopic
flux of particles, J, induced by a concentration gradient:

J(c ) = −DT (c )∇ c (22)

1686 F.J. Keil / Computers and Mathematics with Applications 65 (2013) 1674–1697

where c is the concentration (moles per unit volume). As the driving force is the gradient in the chemical potential, µ, one
can rewrite Eq. (22) as:
L( c )
J(c ) = − ∇µ = −Dc (c )∇µ (23)
kB T
where L(c ) is the Onsager transport coefficient, which can be expressed in terms of the autocorrelation function:
 
 ∞ N N
1  
L= vi (t ) ·
′
vj (0) dt ′ (24)
3V 0 i=1 j=1

where vi (t ) is the velocity of the particle i at time t , V is the volume of the system. The termkB is the Boltzmann constant,
T is the absolute temperature, and Dc (c ) is the so-called collected diffusivity. Other names are corrected diffusivity or
Maxwell–Stefan diffusivity. To link the Maxwell–Stefan diffusion coefficient to the transport diffusion coefficient, one has to
convert the concentration gradient into a gradient of the chemical potential. This can be done by a thermodynamic correction
factor Γ :
1 ∂µ
Γ = . (25)
kB Tc ∂ ln c
This leads to
∇µ = Γ ∇ c (26)
and
DT (c ) = Γ Dc (c ) = L(c )Γ . (27)
From MD simulations one obtains the transport diffusion coefficient in one dimension as follows:
 2 
N
Γ d 
T
D = lim (rj (t ) − rj (0)) (28)
2N t →∞ dt j =1
or
  
N N
Γ ∞
  
T
D = vj (t ) vj (0) dt . (29)
N 0 j=1 j =1

The thermodynamic factor may be written in terms of fugacity, f , as given below:

∂ ln f
 
Γ = . (30)
∂ ln c T
Alternatively one can calculate Γ by a method developed by Reed and Ehrlich [153] during a MC simulation in the grand
canonical ensemble:
⟨N ⟩
Γ =  . (31)
− ⟨N ⟩2

N2
Omission of Γ in Eqs. (28)/ (29) results in the corrected diffusivities. Extensions of the above formulas for multicomponent
systems and applications may be found in Refs. [100,154–159].
The number of simulations of diffusion in nanoporous materials increased sharply over the last few years. The results
have been reviewed [100,154–159]. Out of the very many results one important phenomenon will be discussed. Skoulidas
and Sholl [160] found that the diffusivities of methane in different zeolites may change in various ways as a function of
loading (Fig. 15(a)).
Beerdsen et al. [161,162] analyzed the behavior of diffusivities by the dynamically corrected transition state theory
(dcTST) which can be employed for finite loadings in nanoporous materials [163]. DcTST computes free energy barriers
of a tagged particle. The other particle–particle correlations can be taken into account by a proper definition of an effective
hopping rate of a single particle. From this hopping rate the self-diffusivities can be calculated. The analysis by Beerdsen
et al. [161,162] employing dcTST gave insight into the surprising diffusion behavior of different zeolites. This behavior could
be assigned to the different pore structures of zeolites (Fig. 15(b)). Cage-type zeolites (SAS, LTA, ERI, CHA) consist of large
cages, connected by narrow windows forming large energy barriers. For this type additional molecules lower the energy
barrier that leads to an increase of the diffusion coefficient. Inside tube-like zeolites (AFI, TON, MTW) the free energy in the
interior of the cage rises when new molecules are added, but the effect on the free energy is even larger at the barriers.
The result is a decreasing diffusivity with loading. The behavior of diffusivities in zeolites with two- or three-dimensional
intersecting channels (MFI, ISV, BOG, BEC) is more complex as it involves effects as non-simultaneous freezing in vertical
and horizontal channels, owing to differences in channel diameter and length, causing a varying degree of commensurability
of the molecules with pore length as a function of loading and direction.
After adsorption and diffusion of reactants have been simulated, the reaction itself can be calculated.
 F.J. Keil / Computers and Mathematics with Applications 65 (2013) 1674–1697 1687

Fig. 15. (a) MD simulations of the loading dependence of the corrected diffusivity of methane in various zeolites [160]; (b) free energy profiles of three
categories of zeolites, (1) cage-type (SAS), (2) tube-type (AFI), (3) intersecting type [161,162].

Fig. 16. Interaction of the hydrogen σg and σu∗ levels with a d-band metal surface.

4. Reactions in porous media

Reactions involve bond breaking and bond formation which are genuine quantum chemical processes. Therefore,
quantum chemistry is an important ingredient in reaction modeling. Any catalytic reaction consists of a series of elementary
steps at various active sites. Local geometries have a crucial effect on reaction mechanisms and rates. Inside porous supports
metal nanoparticles or other active sites supplying atoms are distributed, for example, single site Brønsted centers. The
metals of groups VIII and Ib dominate heterogeneous catalysis. Transition metals with unfilled d-bands turned out to be the
most effective catalysts, whereby the position of the d-band center is a good measure of the ability of the atoms to form
bonds to an adsorbate [164]. Taking the center of the d-band is justified because the d-band is usually quite narrow owing
to the strong localization and small overlap between the wave functions of the d-electrons. For transition metals, the Fermi
level lies within the d-band. In the following the dissociative adsorption of hydrogen on a transition metal will be described
qualitatively [16] (Fig. 16).
Hydrogen in the gas phase (see Fig. 16 on the rhs) has a fully occupied bonding σg state and an empty anti-bonding σu∗
state. Interacting with the d-band of the metal, both states split into a bonding and an anti-bonding state with respect to the
surface-molecule bond. The splitting is unsymmetric, as the up-shift of the anti-bonding state is larger than the down-shift of
the bonding state. This is caused by the orthogonalization of the states which raises the kinetic energy and, therefore, leads to
an energetic cost. Thus, if both the bonding and the anti-bonding are occupied, the total energy is raised leading to repulsion.
For a situation like in Fig. 16, both the σg and the σu∗ -derived bonding levels become occupied, while the anti-bonding states
(above the Fermi level) stay empty. This causes an attractive interaction and is the reason for the high reactivity of transition
metals. Owing to the occupation of the σu∗ -derived bonding level, the H–H bond will be weakened and eventually broken.
Besides the electronic effects, geometrical effects come into play from different surface geometries providing different
configurations to the molecule for bonding. In 1969 Boudart [165] introduced a classification of reactions into structure-
sensitive and structure insensitive reactions. In Boudart’s definitions structure-sensitive reactions are those whose rate
varies with surface structure, and structure-insensitive reactions are reactions whose rate does not depend on surface
structure. There seems to be no reaction which is structure-insensitive under all conditions. That the surface structure has
1688 F.J. Keil / Computers and Mathematics with Applications 65 (2013) 1674–1697

Fig. 17. Sabatier volcano curve; α1 = 0.89 eV, α2 = −0.5 eV, β1 = 1.34 eV, β2 = 0.8 eV, 1E = −1 eV [170].

an important effect on reactions was already experimentally detected in 1925 by Pease and Stewart [166], and Taylor [167].
Pease and Stewart found that if they adsorbed a small amount of mercury onto a copper catalyst, the rate of adsorption
of hydrogen went down by a factor of 200 even though the equilibrium constant for the adsorption only changed by an
order of magnitude. Pease and Stewart proposed that adsorption was not occurring uniformly over the solid surface. Taylor
showed that it was possible to poison the reaction without poisoning the adsorption process. Therefore, he suggested that
most reactions occur only on special sites which he termed active sites. One of the ways to detect if a reaction is structure-
sensitive is to run the reaction on a supported catalyst and see if the reaction rate varies with a metal particle size. The
distribution of sites in a supported catalyst varies with particle size. If a variation of rate with particle size is observed, one
knows that the reaction is structure sensitive. An example is the ammonia synthesis over an iron catalyst [168]. One has to
be careful with the interpretation of those measurements, because it is not obvious that if one varies the particle size, one
samples all possible surface structures. Another effect is that if one varies the particle size, one does not vary the distribution
of all of the sites in the same way. Another way to observe structure sensitivity is to run reactions on a variety of faces of
single crystals and look for a variation in rate with crystal face. The size of the geometrical effect depends on the nature of
the transition state of reactions. If the transition state is quite extended, e.g. for N2 dissociation, and several atoms can be
used to stabilize it, then a geometric effect will be found comparing close-packed with more open surfaces. On the other
hand breaking C–H bonds, the transition state is much less extended, shows only a weak geometry-dependence [169].
Bligaard et al. [170] utilized the Brønsted–Evans–Polanyi (BEP) relation to introduce a so-called ‘‘Sabatier analysis’’. The
BEP describes a linear relation between the activation energy of surface reactions and the reaction energy:
EAct = α 1Ereac + β, (32)
whereby α and β are constants. The following simple reaction scheme may be considered [169]:
A2 + ∗ ⇔ A2 ∗ (33)
A2 ∗ +∗ ⇔ 2A ∗ (34)
A ∗ +B ⇔ AB + ∗. (35)
The ‘‘∗’’ represents a surface site, that means A2 ∗ signifies that the molecule A2 is bound on a surface site. If this precursor
state is only weakly adsorbed, the adsorption (Eq. (33)) and the activation (Eq. (34)) of the reactant, A2 , can be viewed as a
single dissociation chemisorption step with an effective activation barrier of EA1 relative to the gas phase reactant:
A2 + 2∗ ⇔ 2A ∗ . (36)
Reaction (Eq. (35)) is written as if the gas phase molecule B reacts directly with an adsorbed atom A with an activation
barrier EA2 . Using the BEP relation, the reaction kinetics depends on two coupled linear relations:
EA1 = α1 1E + β1 (37)
EA2 = α2 1E + β2 (38)
where in most cases α1 is positive and α2 is negative. This means, as the surface becomes more reactive (more negative 1E)
the barrier for activation of the reactants becomes smaller and the barrier for desorption of products becomes larger. This
behavior leads to a Sabatier volcano curve for the over-all reaction rate (Fig. 17).
The Sabatier volcano curve traces out the maximal rate curves. The top of this curve coincides with the top of the
full solution to the microkinetic model. It is a general observation that the Sabatier volcano curve describes the kinetic
well, as long as the reaction is far from equilibrium and the surfaces of optimal reactivity are not poisoned by reactants
or products [169]. The Sabatier analysis can be employed to classify different types of structure sensitivity in catalytic
reactions [169].
 F.J. Keil / Computers and Mathematics with Applications 65 (2013) 1674–1697 1689

Fig. 18. Arrhenius plot of measured thermal sticking coefficients of N2 on a clean Ru(0001) surface and the same surface with gold covered steps [171].

The description of the metal surface, including the active site, is also a problem. Quite often plane crystal faces are
modeled, but steps and corners may be the more important active centers, particularly there are always some defects on the
surface. As catalysts are operated at higher temperatures, the surfaces change with time anyhow. An interesting example
has been published by Dietrich et al. [171] (Fig. 18). The figure shows thermal sticking coefficients of N2 on a clean Ru(0001)
surface with approximately 1% step sites. Blocking the steps with gold results in a reduction of the sticking by about nine
orders of magnitude, which clearly demonstrates the effect of step sites.
In zeolites single atom sites, like, for example, Brønsted sites, are often in use. The advantage of single-site active centers
is that one has a unique well-defined active site. Many other phenomena on surfaces are compiled in a book by Somorjai
and Li [2] and a recent review by Gao and Goodman [172].
In chemical engineering, the Langmuir–Hinshelwood (LH) or Eley–Rideal (ER) rate expressions are very often employed
to describe the kinetics of heterogeneously catalyzed reactions in mesoporous and microporous catalyst supports. Based
on some experimental data, obtained from gaschromatographic measurements in differential or integral reactors under
stationary conditions, several ‘‘reaction mechanisms’’ are proposed that could be in agreement with the experimental
findings. These reaction mechanisms are then transferred into Langmuir–Hinshelwood terms [168], whose parameters are
fitted to the experiments. In a book by Mezaki and Inoue [173], nearly 98% of the kinetic expressions of LH-type could be
fitted to published kinetic data of industrial processes. Quite often more than one mechanism fits well the same data set.
Sometimes one or more LH expressions can be excluded by executing additional experiments under instationary conditions
by using, for example, mass spectrometry. The LH expressions are based on the assumptions of the Langmuir isotherm:
no lateral interactions amongst the adsorbed atoms or molecules, all adsorption sites show the same heat of adsorption
(no heterogeneity), the heat of adsorption is coverage independent. None of these assumptions is fulfilled for a catalyst in
operation. Furthermore details of the surface geometry, modes of adsorption (direct, precursor-mediated, competitive, etc.)
are not taken into account. In LH kinetics it is assumed that all species are chemisorbed and are in thermal equilibrium
with the surface before they take part in any reactions. The Eley–Rideal mechanism assumes that one of the reactants reacts
directly out of the gas phase, without being accommodated at the surface. The fitted Langmuir–Hinshelwood expressions
can be used for reactor design within the range of measurements included in the data fit. Ostrovskii [174] has executed a
critical investigation of surface heterogeneity and has come to the conclusion that no surface heterogeneity reveals itself
in catalysis and chemisorption at metals. Ternel et al. [175] have compared phenomenological kinetics with kinetic Monte
Carlo calculations of CO oxidation on ruthenium. The authors have found that the phenomenological kinetics did not agree
with the kMC results. Inspired by the book by Dumesic et al. [176] microkinetic modeling has been introduced by many
groups. Based on various spectroscopic approaches and simulations a detailed reaction mechanism is set up, which explicitly
includes surface species. No rate-determining step is assumed. The rate expressions for all steps involved are needed. The
procedures are reviewed by Stoltze [177].
The qualitative discussions on reaction phenomena on surfaces have to be completed with mathematical modeling
approaches. Here quantum chemistry comes into play [7–15,178–185].
The stationary, non-relativistic Schrödinger equation is the starting point of most quantum chemical methods.
Born–Oppenheimer approximation separating electronic and nuclear degrees of freedom is employed. The time-
independent Schrödinger equation is of the form:

Ĥ ψ = E ψ̂ (39)
1690 F.J. Keil / Computers and Mathematics with Applications 65 (2013) 1674–1697

where the Hamiltonian Ĥ, for a Coulombic system is given by

1  ZA  1
Ĥ = − ∇i2 − + (40)
i
2 i,A
|ri − rA | i>j
rij

with a set of nuclei with charges ZA and positions rA and number of electrons N. One has to minimize the energy over all
possible anti-symmetric wave functions ψ(x1 , x2 , . . . , xN ), where xi contains the spatial coordinate ri and spin coordinate
σi . The minimum ψ corresponds to the ground state energy. The spin is yet another quantum coordinate associated with
each electron, which can have two values: spin up or spin down. The Pauli exclusion principle states that there can be only
two electrons in the same orbital, and they have to be of opposite spin. An orbital is a quadratic integrable one-electron
function. To include the Pauli exclusion principle into the wave functions Slater determinants are introduced. Starting with
orbitals φi (xi ) = φ(ri )σ (s), the following functions are obtained:
 φ1 (x1 ) φ2 (x1 ) ··· φN (x1 ) 
 
1  φ1 (x2 ) φ2 (x2 ) ··· φN (x2 ) 

ψ(x1 , x2 , . . . , xN ) = √  .. . (41)
N!  .


φ (x ) φ2 (xN ) ··· φ (x )
1 N N N

Slater determinants are anti-symmetric products of orbitals. Every anti-symmetric N-electron function can be expanded
in Slater determinants. The factor (N !)−0.5 is a normalization factor. This Slater determinant describes N electrons occupying
N spin orbitals (φ1 , φ2 , . . . , φN ), without specifying which electron is in which orbital. A short-hand notation for Slater
determinants, including the normalization factor, is
ψ(x1 , x2 , . . . , xN ) = |φ1 (x1 )φ2 (x2 ) · · · φN (xN )⟩. (42)
If one takes only one Slater determinant for the description of an N-electron system, so the expectation value of the
corresponding Hamilton operator Ĥ is a function of the N orbitals:
 
E (φ1 , φ2 , . . . , φN ) = ψ|Ĥ |ψ̂ . (43)

Calculating the energy values, E, for all possible sets of N normalized, pair-wise orthogonal orbitals, then there is a
smallest (infimum) value for E, which is called the Hartree–Fock energy:
def.
E HF = inf E [φ1 , φ2 , . . . , φN ]. (44)
φ1 ,...,φN

The variational principle guarantees, that E HF is a value above the ground state energy

E ≤ E HF . (45)
This property gives the condition

δ E HF [φ1 , . . . , φN ]/δφj = 0, j = 1, 2, . . . , N . (46)

The Ritz variational approach leads to the result that the energetically ‘‘best’’ N-orbital eigenfunctions must be
eigenfunctions of an effective Hamilton operator, the Hartree–Fock operator, ĥHF :

ĥHF φj = εj φj ; j = 1, 2, . . . , N . (47)
E HF
is not the sum of the orbital energies εj . The Hartree–Fock eigenvalue equation (Eq. (47)) replaces the linear eigenvalue
problem of the full N-electron Hamilton operator Ĥ by a non-linear one-electron problem with the effective Hamilton
operator ĥHF . This operator has the form:

ĥHF = ĥ(i) + (Ĵj (i) + K̂j (i)) (48)
j

where ĥ(i) is given by

N
1  ZA
ĥ(i) = − ∇i2 − . (49)
2 r
A=1 iA

The operators Ĵ (i) and K̂ (i) are the Coulomb and exchange operator, respectively, which are defined by:
 
Ĵi φj (1) = dx2 φi (2)r12
∗ −1
φi (2) φj (1) (50)
 
K̂i φj (1) = dx2 φi∗ (2)r12
−1
φj (2) φi (1). (51)
 F.J. Keil / Computers and Mathematics with Applications 65 (2013) 1674–1697 1691

The operators Ĵ and K̂ are invariant with respect to any unitary transformation of the spin orbitals. Unlike the local
Coulomb operator, the exchange operator is said to be a non-local operator, since there does not exist a simple potential
K̂i (x1 ) uniquely defined at a local point (x1 ). The result of operating with K̂i (x1 ) on φj (x1 ) depends on the value of φj
throughout all space, as is evident from Eq. (51). For the practical solution of Eq. (47), the orbitals are expanded in a basis
set of exponential functions, mostly Gaussian type orbitals:

g (r ) = f (x, y, z ) exp(−ζ r 2 ) (52)

where f (x, y, z ) is a polynomial. Any product of two Gaussian-type orbitals is a linear combination of Gaussian-type orbitals.
Finally, one can derive the Roothaan self-consistent field equations, which are eventually an eigenvalue equation whose
solutions (eigenvalues) give the orbital energies, from which the total electronic energy may be calculated [9,26]. Energy
expectation values based on one Slater determinant incorporate exchange correlation, which means that the motion of
two electrons with parallel spins is correlated, but not the motion with opposite spins. Therefore, single determinant wave
functions are referred to as uncorrelated wave functions. Quantum chemistry offers a variety of approaches which include,
at least partially, the correlation energy, defined as the difference between the exact (non-relativistic) energy and the
Hartree–Fock energy in the basis set limit [9,26].
Here, we will mention only the very often used density function theory (DFT). Hohenberg, Kohn, and Sham established a
theoretical basis for justifying the replacement of the many-body wave function by one-electron wave functions [186,187].
The first Hohenberg and Kohn theorem states that for any system of electrons in an external potential, Uext , the Hamiltonian
is determined uniquely by the ground-state charge density alone. In other words, given a charge density ρ , it is possible to
obtain a unique Hamiltonian and associated ground-state wave function, ψ . The second Hohenberg–Kohn theorem provides
the means for obtaining this reverse mapping: The ground-state density of a system in a particular extended potential can
be found by minimizing an associated energy functional. This statement is formal, but later Kohn and Sham provided a
computational method [187] based on the following result: For each interacting electron system, with external potential,
V0 , there is a local potential, Vl , which results in a density ρ equal to that of the interacting system. For a non-interacting
system, the Hamiltonian is introduced as:
N N
1 
Ĥ KS = − ∇i2 + Veff (ri ). (53)
2 i i

Since this Hamiltonian operator does not contain any electron–electron interactions, it indeed describes a non-interacting
system, whereby its ground state wave function is represented by a Slater determinant (see Eq. (41)) with Kohn–Sham
orbitals, ϕi , instead of Hartree–Fock spin orbitals. One obtains an eigenvalue equation:

fˆ KS ϕi = ε KS ϕi (54)
KS
with the one-electron Kohn–Sham operator, f , defined as

1
fˆ KS = − ∇ 2 + Veff (r). (55)
2
The resulting coupled integro-differential equations are [10–15]:
  
1 N 
|ϕj (r2 )|2 M
 ZA
− ∇ + 2
dr2 + Vxc (r1 ) − ϕi = εi ϕi . (56)
2 j
r12 A=1
r1A

Once one knows the various contributions in the square brackets, one can calculate Veff which one has to insert into the
one-particle equations, which in turn determine the Kohn–Sham orbitals, and hence the ground-state electron density and
the ground-state energy. Note that Veff already depends on the electron density. Therefore, like for the Hartree–Fock scheme,
the Eqs. (55) are solved iteratively. Problems stem from the term Vxc , the potential owing to the exchange–correlation Exc .
How the explicit mathematical form of the corresponding potential is unknown, Vxc is simply defined as the functional
derivative of Exc with respect to the electron density:

δ Exc
Vxc = . (57)
δρ
If the exact forms of Vxc and Exc were known, the Kohn–Sham scheme would lead to the exact (non-relativistic) energy,
i.e. the correct eigenvalues of the Hamilton operator Ĥ of the Schrödinger equation. That means that the Kohn–Sham
approach is in principle exact. It should be pointed out that the effective potential Veff (r) is local, as it depends only on
the spatial variable r, but not on other points in space. This is in contrast to the non-local exchange contribution in the
Hartree–Fock approach. The eigenvalues, εI , connected to the Kohn–Sham orbitals, do not have a strict physical meaning
(e.g. ionization energies), except the eigenvalue corresponding to the highest occupied orbital which equals the negative
of the exact ionization energy. Another case is the Slater–Janak theorem [188]. In computational schemes, the Kohn–Sham
1692 F.J. Keil / Computers and Mathematics with Applications 65 (2013) 1674–1697

Fig. 19. Transition state on a potential energy surface.

orbitals are expanded in basis functions, like Gaussian basis functions, or, for periodic systems, in plane waves (exp[ikr]).
One finally ends up with an eigenvalue equation similar to Hartree–Fock:

FKS C = SCε (58)

where F KS is the Kohn–Sham operator


KS
F̂µν = χµ (r1 )fˆ KS χν (r1 )dr1 . (59)

C is the matrix of the basis functions expansion vectors, ε is a diagonal matrix of Kohn–Sham orbital energies, and S is an
overlap matrix

Sµν = χµ (r1 )χν (r1 )dr1 . (60)

Problematic are the exchange and correlation functionals Vxc . Various expressions for Vxc are available in the litera-
ture [10–15]. Which functional one uses depends on the problem at hand. Although DFT has been extremely successful in
solid state physics and quantum chemistry there are limitations [189,190]. An important problem should be mentioned:
the calculation of dispersive forces. Many functionals do not include dispersion forces. One approach is a simple correction
by an additional dispersion term [191]. Other, more theory based approaches are also available. Numerical techniques for
solving the eigenvalue problems of the involved Hermitian operators are given in Refs. [178,192].
The results from quantum chemical calculations are electronic energies, charge densities, vibrational frequencies and
other molecular properties, as, for example, NMR shifts. The electronic energies are used to calculate activation energies of
chemical reactions. Here, the problem of finding transition states in high-dimensional spaces occurs. Transition states (TS)
are first-order saddle points on one spin potential energy surface (PES) (Fig. 19).
One algorithmic approach for finding TSs is based on interpolation between a reactant and a product minimum, another
method uses only local information. Both algorithms can be combined. Interpolation methods generate a sequence of
approximate minimum energy paths (MEPs) by interpolating between a reactant and a product state. The highest energy
configuration along a MEP is a first-order saddle point. Both reactant and product states must be known, so that these
approaches cannot reveal unexpected chemical pathways with multiple intermediates. To confirm a reaction mechanism,
it is necessary to prove that the particular transition state found by the above mentioned algorithms connects the desired
reactants and products. This can be done by following the path of steepest descent downhill from the TS toward the reactants
and toward the products. Following the reaction path can also reveal whether the mechanism involves intermediate states
between reactants and products. Any coordinate system can be used to explore the mechanism of a reaction. Quite often
mass-weighted Cartesian coordinates are used [27]. The path of steepest descend in this coordinate system is called the
intrinsic reaction coordinate (IRC). The interpolation algorithms convert a saddle point search in configuration space to a
minimization problem in discretized path space. Interpolation algorithms include, for example, the nudged elastic band
(NEB) [193] and the string method [194]. Peters et al. [195] have shown that the growing string method, an interpolation
method that does not require an initial guess for the initial pathway, needs significantly fewer gradient calculations to
find the saddle point than the standard NEB and string method. Local surface-walking algorithms explore the PES using
local gradient and usually second derivative information. These methods can be initiated anywhere on the PES. It is
recommendable to employ an interpolation algorithm like the growing string method to generate a starting point for the
local surface walking algorithm. Two of the most used local surface walking algorithms are the partitioned rational function
optimization (P-RFO) method by Baker [196] and the dimer method by Henkelman and Jónsson [197] or its improved version
by Heyden et al. [198].
 F.J. Keil / Computers and Mathematics with Applications 65 (2013) 1674–1697 1693

Fig. 20. Optimization of a catalyst support for a hydrodemetallation reaction; the molar fluxes are presented as a function of micropore radii and
microporosity.

The next step in characterizing a reaction is to calculate the rate of reaction. At the simplest level, one can use the
harmonic transition state theory (TST). A comprehensive review of this subject was presented by Hänggi et al. [34]. The
rate coefficients for elementary reactions on a catalyst surface are obtained by conventional TST in the following way:

kB TQTS (T )
kTST (T ) = exp(−E + /(RT )) (61)
hQR (T )
where kB is the Boltzmann constant, h the Planck constant, T is the absolute temperature and E + is the difference in electronic
energies between the transition state and the reactant state, respectively. The Qi are the partition functions of the TS and
the reactants, respectively. R is the gas constant. In order to set up the description of an entire reaction mechanism on
catalyst surfaces, one has to calculate the rate coefficients for all relevant elementary reaction steps. This requires chemical
intuition and cooperation with experimentalists [176]. An important point is the thermodynamic consistency of the reaction
mechanism. The equilibrium constants of reaction mechanisms are directly tied to thermodynamic quantities. Incorrect
equilibrium constants lead to incorrect predictions of equilibrium limited processes. Furthermore, under non-isothermal
conditions inconsistencies between kinetic parameters and thermodynamic properties result in incorrect solution of the
energy balance and thus temperatures and compositions. The rate constants are inserted into rate equations of various types,
like power laws, Langmuir–Hinshelwood kinetics, amongst others. These expressions are described in many textbooks on
chemical kinetics [199–207].
It should be mentioned that there are examples where the harmonic rigid-rotator approximation to TST fails in describing
reaction kinetics. Simulations based on the static approach can sometimes lead to completely incorrect prediction of the
reaction mechanism. For example, in catalytic transformations of short alkanes, entropy plays an important role. Entropy
can stabilize some otherwise unstable reaction intermediates, opening unexpected alternative reaction channels competing
with the mechanism deduced from a static TST search. Therefore, one has to explore the free-energy surface and not just
of the PES in configuration space. For example, Bucko and Hafner [208] employed transition path sampling techniques
[209,210] to overcome the static TST problems. Quasi-classical MD [35,36] simulations are a further approach to include
reaction dynamics into the simulations (see also [211]). Other general aspects computing catalytic reactions are discussed
by Broadbelt and Snurr [212].
After having developed a reaction scheme and its rate equations, one can optimize the pore structure of amorphous
catalyst supports. For this purpose one can use a three-dimensional pore network model (see Eqs. (16)–(18)). An example is
presented in Fig. 20 [58] for a hydrodemetallation reaction, which shows a maximum in the molar fluxes of metal-bearing
molecules as a function of the micropore radii and microporosity. A further example is given by Wang et al. [213].
Problems of multiscaling in heterogeneous catalysis were reviewed by Raimondeau and Vlachos [214,215], Salciccioli
et al. [216], Keil [144], Chung et al. [217], and Hansen and Keil [218]. The article by Hansen and Keil [218] describes a
complete example of a hierarchical model for benzene alkylation from the active center to the reactor.
Outlook.
Improved theoretical methods based in quantum and statistical mechanics, improvements in intermolecular force fields,
rapid increase in computer speed and memory, more efficient algorithms has enabled unprecedented insights into the
details of catalytic phenomena. Numerical approaches combined with supercomputers and parallel algorithms has led to a
shift from analytical theories, such as partial differential equations, to direct numerical solutions which bypass many of the
approximations necessary in analytical solutions. In the future, more sophisticated dynamical methods, beyond statistical
TST, will give deep insights into the dynamics of chemical reactions on catalytic surfaces. Statistical mechanics will allow for
calculating properties under realistic temperature and pressure conditions. By means of suitable multiscale approaches
1694 F.J. Keil / Computers and Mathematics with Applications 65 (2013) 1674–1697

phenomena on larger time and length scales will be simulated. A combination of in-situ spectroscopical methods and
theoretical simulations will lead to a comprehensive understanding of catalytic processes, and, hopefully, also to rational
catalyst design in the future.

References

[1] C.H. Bartholomew, R.J. Farrauto, Fundamentals of Industrial Catalytic Processes, second ed., Wiley-AIChE, New York, 2005.
[2] G.A. Somorjai, Y.M. Li, Introduction to Surface Chemistry and Catalysis, second ed., Wiley & Sons, New York, 2010.
[3] G. Ertl, Reactions at Solid Surfaces, Wiley & Sons, New York, 2009.
[4] D.P. Woodruff, T.A. Delchar, Modern Techniques of Surface Science, second ed., Cambridge University Press, Cambridge, 1994.
[5] J. Niemantsverdriet, Spectroscopy in Catalysis: An Introduction, second ed., Wiley & Sons, New York, 2000.
[6] C.T. Campbell, Studies of model catalysts with well-defined surfaces combining ultrahigh-vacuum surface characterization with medium-pressure
and high-pressure kinetics, Adv. Catal. 36 (1989) 1–54.
[7] R.A. van Santen, P. Sautet (Eds.), Computational Methods in Catalysis and Materials Science, Wiley-VCH, Weinheim, 2009.
[8] O. Deutschmann (Ed.), Modeling and Simulation of Heterogeneous Catalytic Reactions, Wiley-VCH, Weinheim, 2011.
[9] T. Helgaker, P. Jørgensen, J. Olsen, Molecular Electronic Structure Theory, Wiley & Sons, New York, 2002.
[10] R. Martin, Electronic Structure—Basic Theory and Practical Methods, Cambridge University Press, Cambridge, 2004.
[11] J. Kohanoff, Electronic Structure Calculations for Solids and Molecules, Cambridge University Press, Cambridge, 2006.
[12] E. Engel, R.M. Dreizler, Density Functional Theory: An Advanced Course, Springer, Heidelberg, 2011.
[13] W. Koch, M.C. Holthausen, A Chemist’s Guide to Density Functional Theory, second ed., Wiley-VCH, Weinheim, 2001.
[14] D. Sholl, Density Functional Theory: A Practical Introduction, Wiley-Interscience, New York, 2009.
[15] U. van Barth, Basic density functional theory—an overview, Phys. Scr. T 109 (2004) 9–39.
[16] A. Groß, Theoretical Surface Science: A Microscopic Perspective, Springer, Heidelberg, 2003.
[17] D. Frenkel, B. Smit, Understanding Molecular Simulation, second ed., Academic Press, San Diego, 2001.
[18] K. Binder, D.W. Heermann, Monte Carlo Simulation in Statistical Physics: An Introduction, fifth ed., Springer, 2010.
[19] M.E.J. Newman, G.T. Barkema, Monte Carlo Methods in Statistical Physics, Oxford University Press, Oxford, 1999.
[20] D.C. Rapaport, The Art of Molecular Dynamics Simulation, second ed., Cambridge University Press, Cambridge, 2004.
[21] M.P. Allen, D.J. Tildesley, Computer Simulation of Liquids, Oxford University Press, Oxford, 1989.
[22] M. Griebel, S. Knapek, G. Zumbusch, Numerical Simulation in Molecular Dynamics: Numerics, Algorithms, Parallelization, Applications, Springer,
Heidelberg, 2009.
[23] M.E. Tuckerman, Statistical Mechanics: Theory and Molecular Simulation, Oxford University Press, Oxford, 2010.
[24] J.M. Thijssen, Computational Physics, Cambridge University Press, Cambridge, 1999.
[25] G.C. Schatz, M.A. Ratner, Quantum Mechanics in Chemistry, Dover Publications, New York, 2002.
[26] A. Szabo, N.S. Ostlund, Modern Quantum Chemistry: Introduction to Advanced Electronic Structure Theory, Dover Publications, New York, 1989.
[27] L. Piela, Ideas of Quantum Chemistry, Elsevier, Amsterdam, 2007.
[28] I.N. Levine, Quantum Chemistry, sixth ed., Pearson-Prentice Hall, London, 2009.
[29] D.Al. McQuarrie, Statistical Mechanics, University Science Books, Sausolito, 2000.
[30] T.L. Hill, An Introduction to Statistical Mechanics, Dover Publications, Mineola, New York, 1987.
[31] D. Chandler, Introduction to Modern Statistical Mechanics, Oxford University Press, Oxford, 1987.
[32] W. Ebeling, I.M. Sokolov, Statistical Thermodynamics and Stochastic Theory of Nonequilibrium Systems, World Scientific, Singapore, 2005.
[33] D.G. Truhlar, B.C. Garrett, S.J. Klippenstein, Current status of transition-state theory, J. Phys. Chem. 100 (1996) 12771–12800.
[34] P. Hänggi, P. Talkner, M. Borkovec, Reaction rate theory: fifty years after Kramers, Rev. Modern Phys. 62 (1990) 251–341.
[35] M. Karplus, R.N. Porter, R.D. Sharma, Exchange reactions with activation energy. I. Simple barrier potential, J. Chem. Phys. 43 (1965) 3259–3287.
[36] R.N. Porter, L.M. Raff, W.H. Miller, Quasiclassical selection of initial coordinates and momenta for rotating Morse oscillator, J. Chem. Phys. 63 (1975)
2214–2218.
[37] R. Car, M. Parinello, Unified approach to molecular dynamics and density-functional theory, Phys. Rev. Lett. 55 (1985) 2471–2474.
[38] D. Marx, J. Hutter, Ab Initio Molecular Dynamics: Basic Theory and Advanced Methods, Cambridge University Press, Cambridge, 2012.
[39] T.J. Chung, Computational Fluid Dynamics, second ed., Cambridge University Press, Cambridge, 2010.
[40] G. Froment, K.B. Bischoff, J. De Wilde, Chemical Reactor Analysis and Design, second ed., Wiley & Sons, New York, 2010.
[41] Database issued by the international zeolite association (IZA-SC): www.iza-structure.org/databases.
[42] Ch. Baerlocher, L.B. McCusker, D.H. Olson, Atlas of Zeolite Framework Types, sixth ed., Elsevier, Amsterdam, 2007.
[43] J. Čejka, A. Corma, S. Zones (Eds.), Zeolites and Catalysis: Synthesis, Reactions and Applications, Vol. 2, Wiley-VCH, Weinheim, 2010.
[44] D.J. Franchemontagne, J.L. Mendoza-Cortés, M. O’Keefe, O.M. Yagi, Secondary building units, nets and bonding in the chemistry of metal–organic
frameworks, Chem. Soc. Rev. 38 (2009) 1257–1283.
[45] J. Čejka (Ed.), Metal-Organic Frameworks—Applications from Catalysis to Gas Storage, Wiley-VCH, Weinheim, 2011.
[46] J.Y. Lee, O.K. Farha, J. Roberts, K.A. Scheidt, S.t. Nguyen, J.T. Hupp, Metal–organic framework materials as catalysts, Chem. Soc. Rev. 38 (2009)
1450–1459.
[47] T. Düren, Y.-S. Bae, R.Q. Snurr, Using molecular simulation to characterise metal–organic frameworks for adsorption applications, Chem. Soc. Rev.
38 (2009) 1237–1247.
[48] A. Wheeler, Reaction rates and selectivity in catalyst pores, Adv. Catal. 3 (1951) 249–327.
[49] M.F.L. Johnson, W.E. Stewart, Pore structure and gaseous diffusion in solid catalysts, J. Catal. 4 (1965) 248–252.
[50] J.W. Beeckman, G.F. Froment, Catalyst deactivation by site coverage and pore blockage. Finite rate of growth of the carbonaceous deposit, Chem. Eng.
Sci. 35 (1980) 805–812.
[51] S. Reyes, K.F. Jensen, Estimation of effective transport coefficients in porous solids based on percolation concepts, Chem. Eng. Sci. 40 (1985)
1723–1734.
[52] H.A. Bethe, Statistical theory of superlattices, Proc. R. Soc. Lond. Ser. A 150 (1935) 552–575.
[53] G. Voronoi, Novel applications of continuous parameters on the theory of quadratic forms, J. Reine Angew. Math. 134 (1908) 198–287.
[54] A. Okabe, B. Boots, K. Sugihara, Spatial Tessellations: Concepts and Applications of Voronoi Diagrams, Wiley &Sons, New York, 1992.
[55] F.J. Keil, Modeling reactions in porous media in [8], pp. 149–186.
[56] D. Stauffer, A. Aharoni, Introduction to Percolation Theory, second ed., Taylor & Francis, London, 1992.
[57] C. Rieckmann, T. Düren, F.J. Keil, Interaction of diffusion, reaction and geometric structure of pore networks in catalyst supports—a percolation
theoretical approach to hydrodemetallation, Hung. J. Ind. Chem. 25 (1997) 137–145.
[58] F.J. Keil, C. Rieckmann, Optimization of three-dimensional catalyst pore structures, Chem. Eng. Sci. 49 (1994) 4811–4822.
[59] G.R. Jerauld, J.C. Hatfield, L.E. Scriven, H.T. Davis, Percolation and conduction on Voronoi and triangular networks: a case study in topological disorder,
J. Phys. C 17 (1984) 1519–1529.
[60] G.R. Jerauld, L.E. Scriven, H.T. Davis, Percolation and conduction on the 3D Voronoi and regular networks: a second case study in topological disorder,
J. Phys. C 17 (1984) 3429–3439.
[61] P.H. Winterfeld, L.E. Scriven, H.T. Davis, Percolation and conductivity of random two-dimensional composites, J. Phys. Chem. 14 (1981) 2361–2376.
[62] S.J. Gregg, K.S. Sing, Adsorption, Surface Area and Porosity, second ed., Academic Press, London, 1982.
 F.J. Keil / Computers and Mathematics with Applications 65 (2013) 1674–1697 1695

[63] J.M. Thomas, W.J. Thomas, Principles and Practice of Heterogeneous Catalysis, Wiley-VCH, Weinheim, 1996.
[64] W.C. Conner, S. Christensen, H. Topsoe, M. Ferrero, A. Pullen, The estimation of pore-network dimensions and structure: analysis of sorption and
comparison with porosimetry, Stud. Surf. Sci. Catal. 87 (1994) 151–163.
[65] J.C. Groen, L.A.A. Peffer, J. Pérez-Ramirez, Pore size determination in modified micro- and macroporous materials. Pitfalls and limitations in gas
adsorption data analysis, Microporous and Mesoporous Mater. 60 (2003) 1–17.
[66] J.R.H. Ross, Heterogeneous Catalysis—Fundamentals and Applications, Elsevier, Amsterdam, 2011.
[67] N.A. Seaton, Determination of the connectivity of porous solids from nitrogen sorption measurements, Chem. Eng. Sci. 46 (1991) 1895–1909.
[68] H. Liu, L. Zhang, N.A. Seaton, Determination of the connectivity of porous solids from nitrogen sorption measurements: II. Generalisation, Chem. Eng.
Sci. 47 (1992) 4393–4404.
[69] S. Naumov, R. Valiullin, J. Kärger, R. Pithcumani, M.-O. Coppens, Tracing pore connectivity and architecture in nanostructured silica SBA-15,
Microporous Mesoporous Mater. 110 (2008) 37–40.
[70] J.M. Calo, P.J. Hall, The application of small angle scattering techniques to porosity characterization in carbons, Carbon 42 (2004) 1299–1304.
[71] B.B. Mandelbrot, The Fractal Geometry of Nature, Freeman, San Francisco, 1982.
[72] H. Takayasu, Fractals in the Physical Sciences, Manchester University Press, Manchester, 1990.
[73] M. Barnsley, Fractals Everywhere, Academic Press, San Diego, 1988.
[74] D. Avnir (Ed.), The Fractal Approach to Heterogeneous Chemistry, John Wiley, New York, 1992.
[75] W.-G. Ackermann, H. Spindler, Die fraktale Dimension als Katalysatorkenngröße—I. Eine mathematische Einführung, Z. Phys. Chem. (Leipzig) 269
(1988) 1000.
[76] D. Farin, D. Avnir, The reaction dimension in catalysis on dispersed metals, J. Am. Chem. Soc. 110 (1988) 2039–2045.
[77] D. Avnir (Ed.), The Fractal Approach to Heterogeneous Chemistry: Surfaces, Colloids, Polymers, Wiley, Chichester, 1989.
[78] J.C. Russ, Fractal Surfaces, Plenum Press, New York, 1994.
[79] R. Gutfraind, M. Sheintuch, D. Avnir, Fractal and multifractal analysis of the sensitivity of catalytic reactions to catalyst structure, J. Chem. Phys. 95
(1991) 6100–6111.
[80] R. Kopelman, Fractal reaction kinetics, Science 241 (1988) 1620–1626.
[81] M.-O. Koppens, The effect of fractal surface roughness on diffusion and reaction in porous catalysts: from fundamentals to practical applications,
Catal. Today 53 (1999) 225–243.
[82] X.-Y. Guo, F.J. Keil, Kinetics of N2 O catalytic decomposition over three-dimensional fractals, Chem. Phys. Lett. 330 (2000) 410–416.
[83] I. Chorkendorff, J.W. Niemantsverdriet, Concepts of Modern Catalysis and Kinetics, Wiley-VCH, 2003.
[84] G. Grimmett, Percolation, second ed., Springer, Heidelberg, 2010.
[85] M. Sahimi, Applications of Percolation Theory, Taylor & Francis, London, 1994.
[86] J. Hoshen, R. Kopelman, Percolation and cluster distribution. I. Cluster multiple labeling technique and critical concentration algorithm, Phys. Rev. B
14 (1976) 3438–3445.
[87] D.D. Do, Adsorption Analysis: Equilibria and Kinetics, Imperial College Press, London, 1998.
[88] I. Langmuir, The constitution and fundamental properties of solids and liquids, part I. Solids, J. Am. Chem. Soc. 38 (1916) 2221–2295.
[89] A.W. Adamson, A.P. Gast, Physical Chemistry of Surfaces, Wiley, New York, 1997.
[90] H.T. Davis, Statistical Mechanics of Phases, Interfaces and Thin Films, Wiley-VCH, Weinheim, 1996.
[91] W. Steele, The Interaction of Gases with Surfaces, Pergamon, New York, 1974.
[92] W. Rudzinski, D.H. Everett, Adsorption of Gases on Heterogeneous Solids, Academic Press, London, 1992.
[93] G.E. Norman, V.S. Filinov, Investigations of phase transitions by a Monte Carlo method, High Temp. (USSR) 7 (1969) 216–222.
[94] B. Widom, Topics in the theory of fluids, J. Chem. Phys. 39 (1963) 2808–2812.
[95] J.I. Siepmann, A method for the direct calculation of chemical potentials for dense chain systems, Mol. Phys. 70 (1990) 1145–1158.
[96] J.I. Siepmann, D. Frenkel, Configurational Bias Monte Carlo: a new sampling scheme for flexible chains, Mol. Phys. 75 (1992) 59–70.
[97] T. Vlugt, M. Martin, B. Smit, J. Siepmann, R. Krishna, Improving the efficiency of the configurational-bias Monte Carlo algorithm, Mol. Phys. 94 (1998)
727–733.
[98] M.N. Rosenbluth, A.W. Rosenbluth, Monte Carlo calculation of the average extension of molecular chains, J. Chem. Phys. 23 (1955) 356–359.
[99] D. Frenkel, G. Mooij, B. Smit, Novel scheme to study structural and thermal properties of continuously deformable molecules, J. Phys.: Condens.
Matter 4 (1992) 3053–3076.
[100] B. Smit, T.L.M. Maesen, Molecular simulations of zeolites: adsorption, diffusion, and shape selectivity, Chem. Rev. 108 (2008) 4125–4184.
[101] F.J. Keil, Molecular simulation of adsorption in zeolites and carbon nanotubes, in: L.J. Dunne, G. Manos (Eds.), Adsorption and Phase Behaviour in
Nanochannels and Nanotubes, Springer, Heidelberg, 2010, pp. 9–40.
[102] A.R. Leach, Molecular Modeling—Principles and Applications, second ed., Prentice Hall, New York, 2001.
[103] D. Dubbeldam, S. Calero, T.J.H. Vlugt, R. Krishna, T. Maesen, B. Smit, United atom force field for alkanes in nanoporous materials, J. Phys. Chem. B 108
(2004) 12301–12313.
[104] M.S. Sun, D. Shah, H. Xu, O. Talu, Adsorption equilibria of C1 to C4 alkanes, CO2 and SF6 , J. Phys. Chem. B 102 (1998) 1466–1473.
[105] L. Eder, Thermodynamic siting of alkane adsorption in molecular sieves, Ph.D.-Thesis, University of Twente, 1996.
[106] B. Smit, T.L. Maesen, Commensurate freezing of alkanes in the channels of a zeolite, Nature 374 (1995) 42–44.
[107] A.L. Myers, J.M. Prausnitz, Thermodynamics of mixed-gas adsorption, AIChE J. 11 (1965) 121–127.
[108] S. Jakobtorweihen, N. Hansen, F.J. Keil, Molecular simulation of alkene adsorption in zeolites, Mol. Phys. 103 (2005) 471–489.
[109] A. Heyden, T. Düren, F.J. Keil, Study of molecular shape and non-ideality effects on mixture adsorption isotherms of small molecules in carbon
nanotubes: a Monte Carlo simulation, Chem. Eng. Sci. 57 (2002) 2439–2448.
[110] C. Crespos, H.F. Busnengo, W. Dong, A. Salin, Analysis of H2 dissociation dynamics on the Pd(111) surface, J. Chem. Phys. 114 (2001) 10954–10963.
[111] T. Mitsui, M.K. Rose, E. Fomin, D.F. Ogletree, M. Salmeron, Hydrogen adsorption and diffusion on Pd(111), Surf. Sci. 540 (2003) 5–11.
[112] N. Lopez, Z. Lodziana, F. Illas, M. Salmeron, When Langmuir is too simple: H2 dissociation on Pd(111) at high coverage, Phys. Rev. Lett. 93 (2004)
Paper 146103.
[113] A. Groß, Ab initio molecular dynamics study of hot atom dynamics after dissociative adsorption of H2 on Pd(100), Phys. Rev. Lett. 103 (2009) Paper
246101.
[114] M. Knudsen, The Kinetic Theory of Gases, Methuen, Londen, 1934.
[115] A. Argönül, F.J. Keil, An alternative procedure for modeling of Knudsen flow and surface diffusion, Period. Polytech. 52 (2008) 37–55.
[116] S. Jakobtorweihen, C.P. Lowe, F.J. Keil, B. Smit, Diffusion of chain molecules and mixtures in carbon nanotubes: the effect of host lattice flexibility
and theory of diffusion in the Knudsen regime, J. Chem. Phys. 127 (2007) Paper 024904.
[117] W. Steckelmacher, Knudsen flow 75 years on: the current state of the art for flow of rarefied gases in tubes and systems, Rep. Progr. Phys. 49 (1986)
1083–1107.
[118] J.C. Maxwell, Illustrations of the dynamical theory of gases, Phil. Mag. 19 (1860) 19–32.
[119] L. Boltzmann, Vorlesungen über Gastheorie, Ambrosius Barth, Leipzig, 1896–1898, Reprint: Lectures on Gas Theory, Dover Publ., New York, 1995.
[120] B. Poling, J.M. Prausnitz, J.P. O’Connell, The Properties of Gases and Liquids, McGraw-Hill, New York, 2001.
[121] R. Krishna, J.M. van Baten, Investigating the validity of the Bosanquet formula for estimation of diffusivities in mesopores, Chem. Eng. Sci. 69 (2012)
684–688.
[122] J.D. Doll, A.F. Voter, Recent developments in the theory of surface diffusion, Annu. Rev. Phys. Chem. 38 (1987) 413–431.
[123] A. Kapoor, R.T. Yang, C. Wong, Surface diffusion, Catal. Rev.-Sci. Eng. 31 (1989) 129–214.
1696 F.J. Keil / Computers and Mathematics with Applications 65 (2013) 1674–1697

[124] R. Krishna, Multicomponent surface diffusion of adsorbed species: a description based on the generalized Maxwell–Stefan equations, Chem. Eng. Sci.
45 (1990) 1779–1791.
[125] R. Taylor, R. Krishna, Multicomponent Mass Transfer, Wiley, New York, 1993.
[126] E.A. Mason, A.P. Malinauskas, Gas Transport in Porous Media: The Dusty-Gas Model, Elsevier, Amsterdam, 1983.
[127] S. Chapman, T.G. Cowling, The Mathematical Theory of Non-Uniform Gases, third ed., Cambridge University Press, Cambridge, 1970.
[128] P.J.A.M. Kerkhof, A modified Maxwell–Stefan model for transport through inert membranes: the binary friction model, Chem. Eng. J. 64 (1996)
319–343.
[129] P.J.A.M. Kerkhof, M.A.M. Geboers, Toward a unified theory of isotropic molecular transport phenomena, AIChE J. 51 (2005) 79–121.
[130] J.B. Young, B. Todd, Modelling of multicomponent gas flows in capillaries and porous solids, Int. J. Heat Mass Transfer 48 (2005) 5338–5353.
[131] C. Rieckmann, F.J. Keil, Multicomponent diffusion and reaction in three-dimensional networks: general kinetics, Ind. Eng. Chem. Res. 36 (1997)
3275–3281.
[132] W. Hoyer, J.W. Schmidt, Newton-type decomposition methods for equations arising in network analysis, Z. Angew. Math. Mech. 63 (1984) 397–405.
[133] J.W. Schmidt, W. Hoyer, Ch. Haufe, Consistent approximation in Newton-type decomposition methods, Numer. Math. 47 (1985) 413–425.
[134] I.S. Duff, Direct methods for solving sparse systems of linear equations, SIAM J. Sci. Comput. 5 (1982) 605–619.
[135] O.P. Keipert, M. Baerns, Determination of the intra crystalline diffusion coefficients of alkanes in H-ZSM-5 zeolite by transient technique using the
temporal-analysis-of-products (TAP) reactor, Chem. Eng. Sci. 53 (1998) 3623–3634.
[136] J.A. Delgado, T.A. Nijhuis, F. Kapteijn, J.A. Moulijn, Modeling of fast pulse responses in multitrack; an advanced TAP reactor, Chem. Eng. Sci. 57 (2002)
1835–1847.
[137] T.A. Nijhuis, L.J.P. van den Broeke, J.J.G. Linders, M. Makkee, F. Kapteijn, J.A. Moulijn, Modeling of the transient sorption and diffusion processes in
microporous materials at low pressure, Catal. Today 53 (1999) 189–205.
[138] M.E. Tuckerman, G.J. Martyna, Understanding modern molecular dynamics: techniques and applications, J. Phys. Chem. B 104 (2000) 159–178.
[139] M.E. Tuckerman, D.A. Yarne, S.O. Samuelson, A.L. Hughes, G.H. Martyna, Exploiting multiple levels of parallelism in molecular dynamics based
calculations via modern techniques and software paradigms on distributed memory computers, Comput. Phys. Comm. 128 (2000) 333–376.
[140] B. Leimkuhler, S. Reich, Simulating Hamiltonian Mechanics, Cambridge University Press, Cambridge, 2005.
[141] E. Hairer, C. Lubick, G. Wanner, Geometric Numerical Integration: Structure-Preserving Algorithms for Ordinary Differential Equations, second ed.,
Springer, 2006.
[142] S. Hassani, Mathematical Physics, Springer, Heidelberg, 1999.
[143] H.J.C. Berendsen, Simulating the Physical World-Hierarchical Modeling from Quantum Mechanics to Fluid Dynamics, Cambridge University Press,
Cambridge, 2007.
[144] F.J. Keil, Multiscale modelling in computational heterogeneous catalysis, Top. Curr. Chem. 307 (2012) 69–108.
[145] K.D. Nielson, A.C.T. van Duin, J. Oxgaard, W.-Q. Deng, W.A. Goddard III, Development of the reax FF reactive force field for describing transition metal
catalyzed reactions, with application to the initial stages of the catalytic formation of carbon nanotubes, J. Phys. Chem. A 109 (2005) 493–499.
[146] L. Pauling, Atomic radii and interatomic distances in metals, J. Am. Chem. Soc. 69 (1947) 542–553.
[147] A.G. Bezus, A.V. Kiselev, A.A. Lopathin, P.Q.J. Du, Molecular statistical calculation of thermodynamic adsorption characteristics using atom–atom
approximation. 1. Adsorption of methane by zeolite NaX, J. Chem. Soc. Faraday Trans. II 74 (1978) 367–379.
[148] P. Demontis, G. Suffritti, Structure and dynamics of zeolites investigated by molecular dynamics, Chem. Rev. 97 (1997) 2845–2878.
[149] N.E.R. Zimmermann, S. Jakobtorweihen, E. Beerdsen, B. Smit, In-depth study of the influence of host-framework flexibility on the diffusion of small
gas molecules in one-dimensional zeolitic pore systems, J. Phys. Chem. C 111 (2007) 17370–17381.
[150] P.G. Bolhuis, C. Dellago, Trajectory-based rare event simulations, in: K.B. Lipkowitz, D.B. Boyd (Eds.), Reviews in Computational Chemistry, Vol. 27,
Wiley, New York, 2011.
[151] K.A. Fichthorn, W.H. Weinberg, Theoretical foundations of dynamical Monte Carlo simulations, J. Chem. Phys. 95 (1991) 1090–1096.
[152] K. Reuter, First-principles kinetic Monte Carlo simulations for heterogeneous catalysis: concepts, status, and frontiers in [8], pp. 71–111.
[153] D.A. Reed, G. Ehrlich, Surface diffusivity and the time correlation of concentration fluctuations, Surf. Sci. 105 (1981) 603–628.
[154] F.J. Keil, R. Krishna, M.-O. Coppens, Modeling of diffusion in zeolites, Rev. Chem. Eng. 16 (2000) 71–197.
[155] D. Dubbeldam, R.Q. Snurr, Recent developments in the molecular modeling of diffusion in nanoporous materials, Mol. Simul. 33 (2007) 305–325.
[156] J. Kärger, D. Ruthven, Diffusion in Zeolites and Other Microporous Solids, Wiley, New York, 1992.
[157] J. Kärger, D.M. Ruthven, D.N. Theodorou, Diffusion in Nanoporous Materials, Vol. 2, Wiley-VCH, Weinheim, 2012.
[158] R. Krishna, Diffusion in porous crystalline materials, Chem. Soc. Rev. 41 (2012) 3099–3118.
[159] D.S. Sholl, Understanding macroscopic diffusion of adsorbed molecules in crystalline nanoporous materials via atomistic simulations, Acc. Chem.
Res. 39 (2006) 403–411.
[160] A.I. Skoulidas, D.S. Sholl, Molecular dynamics simulations of self-diffusivities, corrected diffusivities, and transport diffusivities of light gases in four
silica zeolites to assess influences of pore shape and connectivity, J. Phys. Chem. A 107 (2003) 10132–10141.
[161] E. Beerdsen, D. Dubbeldam, B. Smit, Understanding diffusion in nanoporous materials, Phys. Rev. Lett. 96 (2006) Paper 044501.
[162] E. Beerdsen, D. Dubbeldam, B. Smit, Loading dependence of the diffusion coefficient of methane in nanoporous materials, J. Phys. Chem. B 110 (2006)
22754–22772.
[163] D. Dubbeldam, E. Beerdsen, T.J.H. Vlugt, B. Smit, Molecular simulation of loading-dependent diffusion in nanoporous materials using extended
dynamically corrected transition state theory, J. Chem. Phys. 122 (2005).
[164] B. Hammer, J.K. Nørskov, Electronic factors determining the reactivity of metal surfaces, Surf. Sci. 343 (1995) 211–220.
[165] M. Boudart, Catalysis by supported metals, Adv. Catal. 20 (1969) 153–166.
[166] R.N. Pease, L. Stewart, The catalytic combination of ethylene and hydrogen in the presence of metallic copper. III carbon monoxide as a catalyst
poison, J. Am. Chem. Soc. 47 (1925) 1235–1240.
[167] M. Boudart, A. Delboville, J.A. Dumesic, S. Khammouma, H. Topsoe, Surface, catalytic and magnetic properties of small iron particles: I. Preparation
and characterization of samples, J. Catal. 37 (1975) 486–502.
[168] H.S. Fogler, Elements of Chemical Reaction Engineering, Prentice Hall, New York, 2005.
[169] J.K. Nørskov, T. Bligaard, B. Hvolbaek, F. Abild-Pedersen, I. Chorkendorff, C.H. Christensen, The nature of the active site in heterogeneous metal
catalysis, Chem. Soc. Rev. 37 (2008) 2163–2171.
[170] T. Bligaard, J.K. Nørskov, S. Dahl, J. Matthiesen, C.H. Christensen, J. Sehested, The Brønsted–Evans–Polanyi relation and the volcano curve in
heterogeneous catalysis, J. Catal. 224 (2004) 206–217.
[171] H. Dietrich, P. Geng, K. Jacobi, G. Ertl, Sticking coefficient for dissociative adsorption of N2 on Ru single-crystal surfaces, J. Chem. Phys. 104 (1996)
375–380.
[172] F. Gao, D.W. Goodman, Model catalysts: simulating the complexities of heterogeneous catalysts, Annu. Rev. Phys. Chem. 63 (2012) 265–286.
[173] R. Mezaki, H. Inoue (Eds.), Rate Equations of Solid-Catalyzed Reactions, University of Tokyo Press, Tokyo, 1991.
[174] V.E. Ostrovskii, Paradox of heterogeneous catalysis: paradox or regularity? Ind. Eng. Chem. Res. 43 (2004) 3113–3126.
[175] B. Ternel, H. Meskine, K. Reuter, M. Scheffler, H. Metiu, Does phenomenological kinetics provide an adequate description of heterogeneous catalytic
reactions? J. Chem. Phys. 126 (2007) Paper 204711.
[176] J.A. Dumesic, D.F. Rudd, L.M. Aparicio, J.E. Rekoske, A.A. Treviño, The Microkinetics of Heterogeneous Catalysis, American Chemical Society,
Washington, 1993.
[177] P. Stoltze, Microkinetic simulation of catalytic reactions, Prog. Surf. Sci. 65 (2000) 65–150.
[178] Y. Saad, J.R. Chelikousky, S.M. Shontz, Numerical methods for electronic structure calculations of materials, SIAM Rev. 52 (2010) 3–54.
[179] L.-W. Wang, Novel computational methods for nanostructure electronic structure calculations, Annu. Rev. Phys. Chem. 61 (2010) 19–39.
 F.J. Keil / Computers and Mathematics with Applications 65 (2013) 1674–1697 1697

[180] P. Huang, E.A. Carter, Advances in correlated electronic structure methods for solids, surfaces, and nanostructures, Annu. Rev. Phys. Chem. 59 (2008)
261–290.
[181] J. Hafner, Ab-initio simulations of materials using VASP: density-functional theory and beyond, J. Comput. Chem. 29 (2008) 2044–2078.
[182] S. Kummel, L. Kronik, Orbital-dependent density functionals: theory and applications, Rev. Modern Phys. 80 (2008) 3–60.
[183] P. Nachtigall, J. Sauer, Applications of quantum chemical methods in zeolite science, Stud. Surf. Sci. Catal. 168 (2007) 701–736.
[184] P. Pyykkö, Relativistic effects in chemistry: more common than you thought, Annu. Rev. Phys. Chem. 63 (2012) 45–64.
[185] J. Autschbach, Perspective: relativistic effects, J. Chem. Phys. 136 (2012) Paper 150902.
[186] P. Hohenberg, W. Kohn, Inhomogeneous electron gas, Phys. Rev. B 136 (1964) 864–871.
[187] W. Kohn, L.J. Sham, Self-consistent equations including exchange and correlation effects, Phys. Rev. A 140 (1965) 1133–1138.
[188] J.F. Janak, Proof that ∂ e/∂ ni = εi in density functional theory, Phys. Rev. B 18 (1978) 7165–7168.
[189] A.J. Cohen, P. Mori-Sánchez, W. Yang, Challenges for density functional theory, Chem. Rev. 112 (2012) 289–320.
[190] A. Rudenko, F.J. Keil, M.I. Katsnelson, A.I. Lichtenstein, Adsorption of cobalt on graphene: electron correlation effects from a quantum chemical
perspective, Phys. Rev. B 86 (2012) Paper 075422.
[191] L. Goerigk, S. Grimme, Accurate dispersion-corrected density functionals for general chemistry applications, in: P. Comba (Ed.), Modeling of Molecular
Properties, Wiley-VCH, Weinheim, 2011, pp. 3–16.
[192] M.C. Payne, M.P. Teter, D.C. Allan, T.A. Arias, J.D. Joannopoulos, Iterative minimization techniques for ab initio total-energy calculations: molecular
dynamics and conjugate gradients, Rev. Modern Phys. 64 (1992) 1045–1097.
[193] G. Henkelman, H. Jónsson, Improved tangent estimate in the nudged elastic band method for finding minimum energy paths and saddle points,
J. Chem. Phys. 113 (2000) 9978–9985.
[194] W. E, W. Ren, E. Vanden Eijnden, String method for the study of rare events, Phys. Rev. B 66 (2002) Paper 052301.
[195] B. Peters, A. Heyden, A.T. Bell, A. Chakraborty, A growing string method for determining transition states: comparison to the nudged elastic band
and string methods, J. Chem. Phys. 120 (2004) 7877–7886.
[196] J. Baker, An algorithm for the location of transition states, J. Comput. Chem. 7 (1986) 385–395.
[197] G. Henkelman, H. Jónsson, A dimer method for finding saddle points on high-dimensional potential surfaces using only first derivatives, J. Chem.
Phys. 111 (1999) 7010–7022.
[198] A. Heyden, A.T. Bell, F.J. Keil, Efficient methods for finding transition states in chemical reactions: comparison of improved dimer method and
partitioned rational function optimization method, J. Chem. Phys. 123 (2005) Paper 224101.
[199] M. Boudart, G.D. Mariadassou, Kinetics of Heterogeneous Catalytic Reactions, Princeton University Press, Princeton, 1984.
[200] P.L. Houston, Chemical Kinetics and Reaction Dynamics, Dover Publ., New York, 2006.
[201] J.I. Steinfeld, Chemical Kinetics and Dynamics, Prentice Hall, Upper Saddle River, NJ, 1998.
[202] G.B. Marin, G.S. Yablonsky, Kinetics of Chemical Reactions, Wiley-VCH, Weinheim, 2011.
[203] R.I. Masel, Chemical Kinetics & Catalysis, Wiley-Interscience, New York, 2001.
[204] K.J. Laidler, Chemical Kinetics, third ed., Harper & Row Publ., Cambridge, 1987.
[205] L. Arnaut, S. Formosinho, H. Burrows, Chemical Kinetics—From Molecular Structure to Chemical Reactivity, Elsevier, Amsterdam, 2007.
[206] D. Murzin, T. Salmo, Catalytic Kinetics, Elsevier, Amsterdam, 2005.
[207] F.G. Helfferich, Kinetics of Multistep Reactions, second ed., in: N.J.B. Green (Ed.), Chemical Kinetics, vol. 40, Elsevier, Amsterdam, 2004, pp. 1–488.
[208] T. Bucko, J. Hafner, Entropy effects in hydrocarbon conversion reactions: free energy integrations and transition path sampling, J. Phys.: Condens.
Matter 22 (2010) Paper 384201.
[209] L.R. Pratt, A statistical method for identifying transition states in high dimensional problems, J. Chem. Phys. 85 (1986) 5045–5048.
[210] P.G. Bolhuis, C. Dellago, Trajectory-based rare event simulations, in: K.B. Lipkowitz, D.B. Boyd (Eds.), Reviews in Computational Chemistry, Vol. 27,
Wiley, New York, 2011, pp. 111–210.
[211] W. E, E. Vanden-Eijnden, Transition-path theory and path-finding algorithms for the study of rare events, Annu. Rev. Phys. Chem. 61 (2010) 391–420.
[212] L.J. Broadbelt, R.Q. Snurr, Applications of molecular modeling in heterogeneous catalysis research, Appl. Catal. A 200 (2000) 23–46.
[213] G. Wang, M.-O. Coppens, C.R. Kleijn, A tailored strategy for PDE-based design of hierarchically structured porous catalysts, Int. J. Multiscale Comput.
Eng. 6 (2008) 179–190.
[214] S. Raimondeau, D.G. Vlachos, Recent developments on multiscale, hierarchical modelling of chemical reactors, Chem. Eng. J. 90 (2002) 3–23.
[215] D.G. Vlachos, A review of multiscale analysis: examples from systems biology, materials engineering, and other fluid-surface interacting systems,
Adv. Chem. Eng. 30 (2005) 1–61.
[216] M. Salciccioli, M. Stamatakis, S. Caratzoulas, D.G. Vlachos, A review of multiscale modeling of metal-catalyzed reactions: mechanism development
for complexity and emergent behavior, Chem. Eng. Sci. 66 (2011) 4319–4355.
[217] P.S. Chung, M.S. Jhon, L.T. Biegler, The holistic strategy in multi-scale modeling, in: G.B. Marin (Ed.), Advances in Chemical Engineering, Vol. 40,
Elsevier, Amsterdam, 2011, pp. 59–118.
[218] N. Hansen, F.J. Keil, Multiscale modeling of reaction and diffusion in zeolites: from the molecular level to the reactor, Soft Mater. 10 (2012) 179–201.