View Article Online / Journal Homepage / Table of Contents for this issue

PAPER www.rsc.org/pccp | Physical Chemistry Chemical Physics

Analysis of temperature programmed desorption (TPD) data for the

characterisation of catalysts containing a distribution of adsorption sitesw
Patrick J. Barrie
Received 12th November 2007, Accepted 8th January 2008
First published as an Advance Article on the web 7th February 2008
DOI: 10.1039/b717430f
Published on 07 February 2008. Downloaded by Syracuse University on 27/11/2013 17:31:55.

This paper discusses some methods of analysing TPD data for samples obeying first-order
desorption kinetics and proposes several improvements to existing practice. The methods apply in
the case when the Arrhenius parameters A and E for each site are independent of coverage, and
thus are normally suitable for the characterisation of porous solids. An improved implementation
of the condensation approximation method is proposed to gain an initial estimate of the
adsorption site distribution. Further, a variation of the method is proposed that can be used when
A is a function of E. The initial estimate of the distribution can be used to analyse data obtained
by an interrupted TPD experiment, in which heating is halted at a specified temperature. This
method provides a reliable method of determining the parameter A for a peak in the distribution.
Finally, regularisation procedures for obtaining physically sensible distributions from ‘‘noisy’’
TPD data are discussed. It is shown that a penalty function based on the square of the second
derivative of the distribution is normally most suitable for analysing TPD data, at least in the
case when the L-curve method is used to select the regularisation parameter.

1. Introduction
Temperature programmed desorption (TPD) experiments
have been used as a tool to characterise adsorption sites on
catalysts for many years.1–3 The experiment involves adsorp-
tion of a chosen molecule at a low temperature T0, and then
measuring the rate of its desorption as the temperature is
increased. The temperature T is normally increased at a linear
rate, and so:
T = T0 + bt
where b is a constant and t is time.
This paper is concerned with analysing experimental TPD
data when the overall rate of desorption obeys first-order
kinetics. In this case, the overall rate of desorption from a
particular site obeys the equation:
dy
rd ¼
¼ ky
dt
where y is the fractional coverage of the site and k is the rate
constant for desorption. The rate constant k is usually
assumed to obey the Arrhenius equation:



E
k ¼ Aexp ð3Þ
RT
where A is the pre-exponential factor and E is the activation
energy for desorption.
Department of Chemical Engineering, University of Cambridge,
Pembroke Street, Cambridge, UK CB2 3RA. Fax: (+44)1223
334796; Tel: (+44)1223 331864
w Electronic supplementary information (ESI) is available: Results of
additional simulations on datasets containing different levels of noise,
and differing numbers of data points. See DOI: 10.1039/b717430f
There are cases when the assumption about first-order kinetics
is not valid, particularly for desorption from porous catalyst
particles. The assumption may be poor when readsorption of
molecules is significant, when internal diffusion is a significant
mass transfer limitation, and when there is a concentration
gradient within the packed bed of sample.4–6 However, there
are many cases in which the experimental configuration is such
that the rate of desorption does indeed obey overall first-order
kinetics.7–9 For instance, readsorption may be negligible if high
(1) vacuum conditions are used. Even when internal diffusion is a
limiting factor, first-order kinetics may still be obeyed, but the
parameters A and E now have contributions from both internal
diffusion and surface desorption processes.
This paper describes methods for analysing data obtained on
samples that contain a range of different adsorption sites.
Methods for analysing TPD data when the sample contains a
single type of adsorption site are well known, and may be based
ð2Þ on peak shape, peak width or the temperature at which the rate
of desorption is a maximum.1–3,10–15. This paper describes
methods that can be used when the parameters A and E are
independent of coverage. This assumption is often valid if the
initial coverage is low, and can apply for desorption from
porous solids even at high initial coverages. It is recognised that
the assumption will not hold in other cases.1,2,12,16
Consider a sample containing a distribution of adsorption
sites, each having a particular binding strength. The total
number of adsorbed molecules will be given by:
Z1
Ntotal ðTÞ ¼ NðE; TÞdE ð4Þ
0
where N(E,T) denotes the number of molecules adsorbed at
temperature T on a site having an activation energy for
desorption E.

1688 | Phys. Chem. Chem. Phys., 2008, 10, 1688–1696 This journal is 
c the Owner Societies 2008
 View Article Online

Assuming that the molecules on each site obey first-order Combining eqn (5) and (7) shows that the overall rate of
desorption kinetics, then the overall rate of desorption will be: desorption from the sample at temperature T is given by:

 Z1
Z1
dNtotal ðTÞ
E rd ðTÞ ¼ f ðE; TÞ N0 ðEÞ dE ð12Þ
rd ðTÞ ¼
¼ AðEÞ exp NðE; TÞdE ð5Þ
dt RT 0
0
where
where the rate constant for each site is given by the Arrhenius


E
equation parameters A(E) and E. Note that this expression f ðE; TÞ ¼ AðEÞexp y0 ðE; TÞ ð13Þ
allows for the case when the pre-exponential factor is a RT
Published on 07 February 2008. Downloaded by Syracuse University on 27/11/2013 17:31:55.

function of the activation energy E. Eqn (12) is an example of a Fredholm integral equation of the
The number of molecules adsorbed on a site with activation first kind.20 It is well-known that inversion of such an equation
energy for desorption E obeys: to obtain N0(E) from knowledge of rd(T) is a numerically ill-

 posed problem.21 Very different distribution functions N0(E)
dNðE; TÞ
E

¼ AðEÞ exp NðE; TÞ ð6Þ can give rise to almost identical rd(T) functions.
dt RT In the TPD experiment, discrete data point values of rd(T)
are measured and it is desired to obtain N0(E), as this
Assuming that A(E) and E are independent of coverage for
distribution function characterises the different adsorption
each adsorption site, then integration gives:
sites that are present within the sample. As discrete data points
N(E,T) = y 0 (E,T)N0(E) (7) are measured, rather than a continuous function, in practice
the integral will need to be replaced by a summation. How-
where ever, the ill-posed nature of eqn (12) means that a small
8 9 amount of noise in experimental rd(T) values can have a major
>
< ZT
 > = impact on any N0(E) function derived from it.
AðEÞ
E
y0 ðE; TÞ ¼ exp
exp dT ð8Þ This paper discusses some of the methods for obtaining the
>
: b RT >
;
T0 N0(E) distribution function from experimental TPD measure-
ments and proposes some improvements to existing practice.
In eqn (7), N0(E) is the activation energy distribution for the The paper has the following sections:
adsorbed molecules at the initial temperature T0. 1. Introduction
The function y 0 (E,T) defined by eqn (8) corresponds to the 2. Estimating N0(E) when A is a known constant
fraction of molecules still adsorbed on sites with activation 3. Estimating N0(E) when A is a known function of E
energy E at temperature T compared to those adsorbed on 4. Methods of determining A
those sites at temperature T0, i.e. 5. Optimising the N0(E) distribution function
6. Discussion and conclusions
NðE; TÞ This paper assumes that the activation energy for de-
y0 ðE; TÞ ¼ ð9Þ
N0 ðEÞ sorption from a site is independent of its coverage. Some
comments on what to do when this is not the case are made
If the adsorption sites had been fully occupied at temperature at the end of the paper.
T0, then y 0 (E,T) would be the actual fractional coverage of the
sites. However, in the general case, the fractional coverage of a
site with activation energy for desorption E at temperature 2. Estimating N0(E) when A is a known constant
T obeys:
The condensation approximation is one simple method of
N0 ðEÞ 0 circumventing the problem of the ill-posed nature of eqn
yðE; TÞ ¼ y ðE; TÞ ð10Þ (12), and is particularly useful when the pre-exponential factor
Ns ðEÞ
A is a constant (independent of E). The method has been
where Ns(E) is the number of molecules that would be widely used when analysing adsorption isotherms to obtain
adsorbed at saturation conditions on sites with energy E. approximate adsorption energy distributions.22,23 It gives
The integral in eqn (8) giving y 0 (E,T) does not have stable results and is simple. The condensation approximation
an analytical solution, but it can readily be evaluated numeri- assumes that, at a particular temperature T, all molecules
cally to any desired accuracy (e.g. using Simpson’s rule). There occupying sites with binding energies less than a critical energy
are also a variety of analytical expressions that approximate E* will have desorbed, while all molecules occupying sites with
the value of the integral in the case when T0 is small.17–19 For binding energy greater than E* will still be adsorbed. Thus the
example, the following expression gives good accuracy over condensation approximation treats y 0 (E,T) in eqn (13) as a
the full range of values of E and T while still being reasonably step function:
simple mathematically:
y 0 (E,T) E 0 for E o E*(T); y 0 (E,T) E 1 for E 4 E*(T)(14)
"
#

ART 2 1
E Fig. 1 shows that the actual function y 0 (E,T) is sigmoidal in
y0 ðE; TÞ  exp pffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi exp ð11Þ shape at a specified T (—). The figure also shows the step
bE 1 þ 2RT=E RT

This journal is 
c the Owner Societies 2008 Phys. Chem. Chem. Phys., 2008, 10, 1688–1696 | 1689
 View Article Online
Published on 07 February 2008. Downloaded by Syracuse University on 27/11/2013 17:31:55.

Fig. 1 Graph showing y 0 (E,T) as a function of E at a fixed T. The Fig. 2 Graph showing f(E,T) as a function of E at a fixed T. It is
dotted line shows the step function used by the condensation approx- proposed that E*(T) is chosen such that Area 1 = Area 2.
imation to estimate y 0 (E,T).
be 1/e, though this work has not been able to prove this should
function used to approximate it (--). The two functions differ be the case analytically. While this is a non-linear equation, it
only in the region between E*
2RT and E* + 4RT. is straightforward to calculate E* if the parameter A is known
Different methods of calculating E*(T) have been proposed (e.g. using the Newton–Raphson method).
in the literature. One suggestion is to choose E*(T) such that It will be necessary to obtain an expression for dE*/dT to
molecules on sites with that activation energy have a max- use the condensation approximation. To do this, it is con-
imum in their rate of desorption at that specified T.24 When A venient to define the following function F:
is a constant, it is straightforward to show that this corre-
 

sponds to E* and T being related by the equation: E AT E
F ¼ ln
0:368
ln þ
0:368 ð19Þ


 RT b RT
E E AT
exp ¼ ð15Þ The triple product rule for partial derivatives implies:
RT RT b

 @F 
Another suggestion is to choose E* such that Area 1 and Area @E
@T 
¼  @F  E ð20Þ
2 in Fig. 1 are the same. Seebauer showed that the value of E* @T F @E  T
giving equal areas in Fig. 1 is the one satisfying:25

 
  Using this method, algebra leads to the following expression
E 1 E 1 AT for the derivative for the case that A is a constant:

exp
¼ ð16Þ
RT 2 RT 2 b
2 
dE  x þ 1:632x
0:368
¼R ð21Þ
In this work, an alternative suggestion for the ‘‘best’’ E*(T) dT x þ 0:632
function is proposed. From eqn (12), it can be seen that the
function f(E,T) describes the rate at which molecules are where x* = E*/RT. In the event that x* is large, then dE*/dT
desorbing from sites with activation energy E once tempera- E E*/T.
ture T has been reached. This work proposes that the best When the condensation approximation applies, the number
choice of E* is the one in which: of adsorbed molecules at temperature T is:

EZ ðTÞ
Z1
Z1
Ntotal ðTÞ ¼ N0 ðEÞdE ð22Þ
f ðE; TÞdE ¼ f ðE; TÞdE ð17Þ
E  ðTÞ
0 E  ðTÞ
Differentiation gives:
The form of the function f(E,T) is shown in Fig. 2; this
criterion corresponds to Area 1 and Area 2 in this figure being dNtotal ðTÞ
¼
N0 ðE  ðTÞÞ ð23Þ
equal. dE 
A numerical investigation was performed into the selection Hence:
of E* using this criterion. Values of AT/b spanning more than
dNtotal ðTÞ dt dT
20 orders of magnitude were studied. In all cases, when A is a N0 ðE  ðTÞÞ ¼
ð24Þ
constant, the value of E* that gave equal areas in Fig. 2 dt dT dE 
corresponded to that given by the equation: The first term on the right-hand side is simply the rate of

 
  desorption as measured by the TPD experiment, and so:
E E AT

0:368 exp
0:368 ¼ ð18Þ
RT RT b 1 rd ðTÞ
N0 ðE  ðTÞÞ ¼ ð25Þ
b dE  =dT
It can be seen that this criterion is intermediate between the
alternatives proposed in the literature (eqn (15) and (16)). It is Eqn (25) means that the distribution function N0(E) can easily
worth noting that the parameter 0.368 in eqn (18) happens to be extracted from TPD data if the condensation

1690 | Phys. Chem. Chem. Phys., 2008, 10, 1688–1696 This journal is 
c the Owner Societies 2008

approximation is valid. For each temperature at which a TPD
data point is obtained, E* can be calculated from eqn (18) if A
is known, while dE*/dT can be calculated from eqn (21).
Substitution of these values into eqn (25) then gives the desired
N0(E) distribution function.
In order to test the performance of the condensation
approximation method on samples with constant values of
A, ‘‘experimental’’ TPD data were generated using eqn (12) for
samples of known adsorption characteristics. Using the known
values of A, the data were then analysed using the condensa-
Published on 07 February 2008. Downloaded by Syracuse University on 27/11/2013 17:31:55.
tion approximation to predict the N0(E) distribution function
and the TPD profile. Some of the results of the simulations
and the actual distributions are shown in Fig. 3.
It can be seen that there is fair agreement between the
distributions derived from the condensation approximation
and the actual distributions. In particular, the maxima in the
derived N0(E) distribution functions are at the correct activa-
tion energy. Note that this is only the case when eqn (18) is
used to calculate E*(T)—use of eqn (15) and (16) give too low
and slightly too high values of E*, respectively. However, the
condensation approximation gives rise to a systematic
error—the distribution predicted by the condensation approx-
imation is broader than the actual distribution, and it can be
seen that this error is more significant at low values of A/b
than at high values of A/b.
Fig. 3 Simulations using the condensation approximation method.
(a) Curve: actual distribution function for a sample with a Gaussian
distribution of sites (E = 100 kJ mol
1, s = 5 kJ mol
1, A = 1013
s
1). Data points: N0(E) distribution derived from analysing the TPD
data shown in (b) using eqn (25). (b) J: ‘‘Experimental’’ TPD data
generated for a sample with the N0(E) distribution function described
in (a) with b = 0.1 K s
1 and points recorded every 5 K. ’:
Simulation of the TPD profile using the N0(E) distribution derived
from eqn (25). (c) Curve: actual distribution function for a sample with
a Gaussian distribution of sites (E = 60 kJ mol
1, s = 3 kJ mol
1, A
= 105 s
1). Data points: N0(E) distribution derived from analysing the
TPD data shown in (d) using eqn (25). (d) J: ‘‘Experimental’’ TPD
data generated for a sample with the N0(E) distribution function
described in (c) with b = 0.1 K s
1 and points recorded every 5 K.
’: Simulation of the TPD profile using the N0(E) distribution derived
from eqn (25).
This journal is 
c the Owner Societies 2008
View Article Online
In this paper, the condensation approximation is treated as
a useful initial estimate of the distribution that can be used as a
starting point for iterative fitting schemes.24,26
3. Estimating N0(E) when A is a known function of
E
The condensation approximation is a simple robust method to
calculate an approximate N0(E) distribution function when the
Arrhenius pre-exponential factor A is a constant. However,
the parameter A may vary for the different adsorption sites
present. Indeed, it has been found that it often obeys the so-
called ‘‘compensation effect’’ equation, in which sites with
higher activation energies have higher pre-exponential
factors:27–29
A = a0exp(a1E) (26)
Such an equation has been observed for many rate processes
on solid surfaces including temperature programmed deso-
rption.27,30,31 The origin of the compensation effect remains a
topic of debate in the literature. For instance, it has been
pointed out that, when E depends on coverage because of
adsorbate–adsorbate interactions, then analysis of TPD data
assuming that E is independent of coverage leads to a ‘‘false’’
compensation effect in which the A obtained appears to vary
with E when it is in fact constant.32,33 Other interpretations for
the compensation effect have also been postulated.34
This section proposes an alternative method to the conden-
sation approximation that is robust and straightforward to
implement to calculate an approximate N0(E) distribution
function when A is a known function of E.
Eqn (12) states:
Z1
rd ðTÞ ¼ f ðE; TÞN0 ðEÞdE ð12Þ
0
If the distribution function is treated as a constant, N0(E*), at
a particular temperature, then:
Z1
rd ðTÞ  N0 ðE  Þ f ðE; TÞdE ð27Þ
0
As discussed in the previous section on the condensation
approximation, the ‘‘best’’ value of E* can be calculated at a
particular temperature using eqn (18) (though now recognising
explicitly that A is a function of E):

 
 
E E AðEÞT

0:368 exp
0:368 ¼ ð18Þ
RT RT b
This means that an approximate distribution function can be
obtained from experimental TPD data by:
rd ðTÞ
N0 ðE  Þ ¼ R1 ð28Þ
f ðE; TÞdE
0
For each temperature at which a TPD data point is obtained,
E* can be calculated from eqn (18) if A(E) is known. The
integral in the denominator of eqn (28) can be evaluated
numerically; in practice, the integration limits can be taken
Phys. Chem. Chem. Phys., 2008, 10, 1688–1696 | 1691

to be E*
5RT and E* + 10RT without loss of accuracy (as
implied by Fig. 2). Substitution of these values into eqn (28)
then gives the desired N0(E) distribution function.
As in the previous section, in order to test the method,
‘‘experimental’’ TPD data were generated using eqn (12) for
samples of known adsorption characteristics. In the example
shown in Fig. 4, the test sample has two adsorption sites, each
with a Gaussian distribution, and with a pre-exponential
factor obeying a compensation effect relationship such that
the values of A for the two sites differ by an order of
Published on 07 February 2008. Downloaded by Syracuse University on 27/11/2013 17:31:55.
magnitude. The results were processed in three different ways.
Fig. 4 TPD simulations and derived distribution functions for a
sample with two adsorption sites each having a Gaussian distribution
(E1 = 100 kJ mol
1, s1 = 5 kJ mol
1, E2 = 140 kJ mol
1, s2 = 5 kJ
mol
1). The pre-exponential factor is given by the compensation
relation A = a0exp(a1E) where a0 = 108 s
1 and a1 = 0.05 mol
kJ
1. (a) Curve: actual distribution. Data points: N0(E) distribution
derived from analysing the TPD data shown in (b) using eqn (25)
assuming A is a constant and equal to 6.25 x 1010 s
1. (b) J:
‘‘Experimental’’ TPD data generated with b = 0.1 K s
1 and points
recorded every 5 K. ’: Simulation of the TPD profile using the N0(E)
distribution derived in (a). (c) Curve: actual distribution. Data points:
N0(E) distribution derived from analysing the TPD data shown in (d)
using eqn (28) assuming A is a constant and equal to 6.25 x 1010 s
1.
(d) J: ‘‘Experimental’’ TPD data generated with b = 0.1 K s
1 and
points recorded every 5 K. ’: Simulation of the TPD profile using the
N0(E) distribution derived in (c). (e) Curve: actual distribution. Data
points: N0(E) distribution derived from analysing the TPD data shown
in (f) using eqn (28) assuming A obeys the compensation effect (eqn
(26)) with a0 = 108 s
1 and a1 = 0.05 mol kJ
1. (f) J: ‘‘Experimental’’
TPD data generated with b = 0.1 K s
1 and points recorded every
5 K. ’: Simulation of the TPD profile using the N0(E) distribution
derived in (e).
1692 | Phys. Chem. Chem. Phys., 2008, 10, 1688–1696
View Article Online
Firstly, the data were analysed assuming A was constant using
the condensation approximation, eqn (25) (Fig. 4(a) and (b)).
Secondly, the data were analysed assuming A was constant
and using eqn (28) (Fig. 4(c) and (d)). Thirdly, the data were
analysed assuming the actual compensation effect relationship
and eqn (28) (Fig. 4(e) and (f)).
It can be seen that, when analysing data assuming that A is a
constant, eqn (28) gives almost identical distributions to the
condensation approximation method (eqn (25)). In this parti-
cular case, the assumption of constant A is not valid, and so
there is an error in the derived N0(E) distributions because of it
(Fig. 4(a) and 4(c)). When the correct value of A is used for the
major peaks, then eqn (28) does give peaks at the correct
energies in the N0(E) distribution function (Fig. 4(e)). As
previously, the distribution functions obtained tend to be
broader than the actual distribution function.
As is the case when using the condensation approximation,
eqn (28) is a robust method of estimating N0(E) that gives
sensible distributions even in the presence of significant noise
in the data. It is straightforward to employ eqn (28) even in
cases when A is not a constant, provided that A is a known
function of E. Ways of determining values of A for different
adsorption sites are discussed in the next section. The dis-
tributions derived from this method are convenient starting
points for the iterative schemes described in section 5.
4. Methods of determining A
Sometimes the parameter A can be calculated theoretically by
statistical thermodynamics. In other cases, A may be esti-
mated. For instance, because A corresponds to the rate con-
stant for desorption at a very high temperature, it is often
taken to be 1013 s
1 for desorption of simple molecules from
single crystal surfaces; this rate corresponds to the frequency
of a molecular vibration. However, it remains difficult to
estimate A for desorption of large molecules from single
crystal surfaces, or to estimate the overall value of A if the
sample is porous and the rate of desorption is affected by
internal diffusion. In these cases, the parameter A needs to
obtained experimentally.
The most common method for finding A experimentally is
to perform the TPD experiment at more than one heating
rate.1,2,12,25 The temperature at which a peak maximum occurs
can then be measured. Near the maximum of a particular
peak, A can be treated as constant, and based on eqn (15) the
temperature of a peak maximum Tp will obey:
!


b AR E 1
ln ¼ ln
ð29Þ
Tp2 E R Tp
This implies that a graph of ln(b/T2p) against 1/Tp will be
linear; the gradient enables the value of E corresponding to the
peak maximum to be found, and the intercept enables A for
sites at that value of E to be obtained. However, it may be
difficult to measure the temperature of the peak maximum
accurately for broad distributions in the presence of experi-
mental noise, and the method requires a long extrapolation of
the data to find the intercept. For good results, the heating rate
b needs to be varied by at least an order of magnitude, and this
This journal is 
c the Owner Societies 2008
 View Article Online

can pose problems in certain cases. This method provides a

useful estimate of A but it is not always very accurate.
An improved method for obtaining A from data obtained at
more than one temperature is to fit the entire TPD profiles to
find the value of A that gives best overall agreement between
experimental and simulated data.35 This is not trivial as it
involves optimisation of both the N0(E) distribution and the
value of A; ways of optimising the N0(E) distribution are
discussed below.
Another method that has been proposed to obtain experi-
Published on 07 February 2008. Downloaded by Syracuse University on 27/11/2013 17:31:55.

mental values of A is to modify the TPD experiment in the

following way. Adsorption takes place at the normal low
temperature, and the sample is then heated to a specified
temperature, Tsoak, for a specified time period. After this
period, the sample is cooled and then the normal linear heating
rate TPD measurement is performed. Analysing data for
different Tsoak values enables A to be obtained for different
values of E.30,36,37
In this paper we propose a new method for obtaining the
parameter A based on an interrupted TPD experiment. This
technique has occasionally been used by researchers,38 though
to our knowledge it has not previously been proposed as a
method for obtaining the parameter A. The sample is heated at
the chosen heating rate b, until a specified temperature Tint is
reached. The sample is then held at this temperature for a
chosen duration, before the temperature is increased further.
For the initial experiments, the temperature chosen should
correspond to a peak maximum in the TPD experiment
performed without interruption. Fig. 5 shows an example
temperature profile and the resulting rate of desorption as a
function of time.
During the interrupted TPD experiment, the change in the
rate of desorption as a function of time during the isothermal
period provides a method for determining the parameter A at
energies close to E*(Tint). For first-order desorption kinetics,
we can modify eqn (12) to give the following equation for the
rate of desorption during the isothermal period:
Z1 


E
rint ¼ f ðE; Tint ÞN0 ðEÞexp
tint AðEÞexp dE
RTint
0 how many data points are used in the fit to obtain the best
ð30Þ
Here the time tint is the time from the point at which heating is
halted.
If all desorbing molecules were coming from a single site, as
assumed by the condensation approximation, then it would be
expected that a graph of ln(rint) against tint would be linear
with a gradient that depends on the value of A. However, the
rate of desorption during the interrupted heating period
depends on the populations of sites with energies both below
and above E*(Tint), and so the graph is not linear. Therefore,
an iterative scheme is needed to find the best estimate of the
parameter A at energy E*. First the value A is guessed. This
enables the N0(E) distribution function to be estimated by eqn
(25) or eqn (28). This function can then be used in eqn (30) to
predict the desorption rate during the interrupted heating
period, with the integration performed using the trapezium
rule. Comparison of the actual and predicted behaviour then
Fig. 5 (a) The temperature profile and generated TPD data for an
interrupted heating experiment. The adsorption site distribution was
Gaussian with E = 100 kJ mol
1, s = 5 kJ mol
1, A = 1013 s
1, b =
0.1 K s
1. The heating was interrupted at 344 K for 400 s. (b) Least
squares fit of the desorption rate data during the interrupted heating
period. The experimental data are solid points, and the curve is the
best fit, corresponding to A = 2.8  1013 s
1 in this case. Note that if
only the first 100 s of data are analysed, then the resulting best value of
A is 0.85  1013 s
1.
enables a better value of parameter A to be chosen and the
process is repeated until the optimal fit is obtained as shown in
Fig. 5(b).
The interrupted TPD experiment is a good method for
calculating A for adsorption sites that correspond to peak
maxima in TPD profiles. There is a decision to be made over
value of A. Based on the simulations performed, such as those
shown in Fig. 5(b), we have found that when the data contains
little noise then it is best to fit only the first few data points
recorded under isothermal conditions. For instance, in the
absence of noise, the most accurate value of A comes from
fitting data corresponding to up to a two-fold reduction in rate
of desorption. At higher values of tint, there is a small error
associated with using the approximate N0(E) distribution
generated by eqn (25) or eqn (28) rather than the actual
distribution. However, if the TPD data contains significant
noise, then it is important to make sure that sufficient data
points are used to observe a reliable trend, even if this means
using values of tint that give the small error associated with
using an approximate N0(E) distribution.
The fact that the N0(E) distribution generated by eqn (25) or
eqn (28) is only an approximation also creates a systematic
error in the value of A obtained by this method if the
interruption to the heating takes place at a temperature other

This journal is 
c the Owner Societies 2008 Phys. Chem. Chem. Phys., 2008, 10, 1688–1696 | 1693
 View Article Online

than the normal TPD peak maximum. If Tint is smaller than
the peak maximum, then the calculated value of A using the
N0(E) distributions predicted by eqn (25) or eqn (28) is lower
than the actual value, while if Tint is greater than the peak
maximum, then the calculated value of A by this method is
higher than the actual value. For adsorption sites that do not
correspond to maxima in the normal TPD profile, a more
accurate N0(E) distribution function is required if the A value
is to be determined accurately by the interrupted TPD method
described here. The method for obtaining an accurate N0(E)
Published on 07 February 2008. Downloaded by Syracuse University on 27/11/2013 17:31:55.
minimisation should be performed with the constraint that all
N0(Ei) values must be greater than or equal to zero.
However, the difficulty with this procedure is that the
inversion problem is mathematically ill-posed, with very simi-
lar TPD results being possible from vastly different energy
distributions, as is illustrated by the following example. Fig.
6(a) shows TPD results for a sample containing two types of
adsorption site each with a Gaussian distribution. A signifi-
cant amount of random noise has been added to the TPD data
in order to mimic real experimental data. The best fit to the

distribution function is described in the next section.
5. Optimising the N0(E) distribution function
The methods described in the above sections have given us
reliable measurements of A for adsorption sites corresponding
to the TPD peak maxima, and an approximate N0(E) distribu-
tion. It is now desired to perform a least squares fitting routine
to obtain an accurate N0(E) distribution and good agreement
between experimental and numerically simulated TPD results.
As stated earlier, we desire to obtain the energy distribution
function N0(E) from experimental rates of desorption rd(T)
that obey:
Z1
rd ðTÞ ¼ f ðE; TÞN0 ðEÞdE ð12Þ
0
experimental data, obtained by minimising the residual sum of
the squares, is shown as the solid line in the figure. Fig. 6(b)
shows the energy distribution function that produced the best
fit to the ‘‘noisy’’ experimental data. This is certainly not
similar to that which was used to generate the original TPD
data. In particular, its excessive ‘‘spikiness’’ is unrealistic of a
real catalyst. Experience indicates changes in the physical
properties of a heterogeneous surface happen more gradually
than is shown here.
A common method to obtain physically sensible distribu-
tions from an analysis of ill-posed equations such as eqn (12) is
to use a regularisation procedure as initially suggested by
Tikhonov.42,43 Rather than simply optimising the fit to experi-
mental data, a penalty function is included in the optimisation

where f(E,T) is given by eqn (13).

The actual experimental values of rd(T) will be n discrete
data points, rather than a continuous function. Simulation of
the experimental data points can therefore only sensibly
assume m discrete N0(Ei) values (with m r n), and the values
of Ei need to be chosen appropriately. If the N0(Ei) values are
known, then the integral in eqn (12) can be approximated by
the trapezium rule. For instance, in the special case that the
values of Ei are equally spaced, then:
X
m
rd ðTÞ  f ðEi ; TÞN0 ðEi Þ DEi ð31Þ
i¼1

The function f(Ei,T) can be evaluated numerically for all

combinations of Ei values (m of these) and T values (n of
these) under consideration. We can then readily calculate
predicted rates of desorption, rpred(T), for a starting set of
values for N0(Ei). We can then vary the N0(Ei) values to obtain
the best fit to the experimental data. Mathematically, this is a
linear least squares fitting procedure which minimises the
residual sum of the squares:
X
n
rss ¼ ½rd ðTj Þ
rpred ðTj Þ2 ð32Þ
j¼1
Fig. 6 (a) Data points: TPD profile generated for a sample containing
Such a minimisation has been performed by a variety of two types of adsorption site, one with E = 85 kJ mol
1 and a standard
deviation of 3 kJ mol
1; the other site with E = 110 kJ mol
1 and a
different research groups to fit TPD data.24,26,30,31,35,36,39
41
standard deviation of 10 kJ mol
1; the relative area of the two sites is
A sensible choice of Ei values are those predicted by the
1 : 2. Both sites have A = 1010 s
1. The heating rate is b = 0.1 K s
1.
condensation approximation method (eqn (18)), in which case Random experimental noise has been added. The solid line displays
there will be the same number of Ei values as T values (i.e. the best simulation of the experimental profile using the N0(E)
m = n). Using this basis ensures that there are no more points in distribution shown in part (b). (b) Solid line: actual distribution. Data
a peak in the resulting distribution function than are justified by points linked by dashed line: distribution obtained by least squares
the experimental data. In order to be physically realistic, the fitting of the TPD data by minimising eqn (32).

1694 | Phys. Chem. Chem. Phys., 2008, 10, 1688–1696 This journal is 
c the Owner Societies 2008
 View Article Online

procedure in order to generate physically sensible distribu-

tions. Thus we now desire to minimise:
z = rss + lp (33)
where p is a penalty function that is a measurement of the
unnaturalness of the distribution, and l is a regularisation
parameter that needs to be chosen. Of the groups that have
analysed TPD data by using eqn (32), only a few have used
regularisation techniques.24,26,35,40 However, there remains the
decision on what is the best choice of penalty function for
Published on 07 February 2008. Downloaded by Syracuse University on 27/11/2013 17:31:55.

analysing TPD data.

Two penalty functions are commonly considered in regular- Fig. 7 L-Curve plot for determining the optimum choice of regular-
isation problems.43 The first is based on the area under the isation parameter l.
energy distribution curve: On the other hand, when the penalty function is based on
Z1 the second derivative as in eqn (35), then the value of l chosen
p¼ N02 ðEÞdE ð34Þ by L-curve method is a good one. The peak positions and
widths for the two adsorption sites are accurately predicted,
0
though the area of the narrower peak is slightly lower than
The second possibility is to use a penalty function based on the should be the case.
second derivative of the distribution function: It should be noted that, during the course of this investiga-
Z1  2 2 tion, data analysis was performed on TPD data generated for
d N0 ðEÞ a large number of energy distributions and containing widely
p¼ dE ð35Þ
dE 2 differing levels of noise. Some of these analyses are presented
0
in the ESI to this paper.w It is observed that the performance
Both these penalty functions have been investigated in this work. of the two penalty functions is often similar for cases when a
There remains the important decision on how the parameter single peak is present, but differs when more than one peak is
l is chosen. If the value of l is too small, then the fit is likely to
be spiky with spurious peaks (as in Fig. 6(a)). However, if the
value of l is too large, then the resulting distribution will
become too smooth and real features of the distribution may
be lost. A sensible choice of l is the highest value at which the
fit to the experimental data remains good. On this basis, it is
sometimes possible to choose l ‘‘by eye’’, but this remains a
subjective judgement. Therefore mathematical techniques for
choosing l objectively have been developed. One such techni-
que is the L-curve method.21,44 This is based on plotting a
graph of the penalty function against the residual sum of the
squares (i.e. p against rss) for the optimised N0(E) distributions
obtained at different values of l. At low values of l, a small
increase in l can greatly reduce the penalty function without
significantly affecting the residual sum of squares. However, at
high values of l, a small increase in l can greatly reduce the
quality of fit (as measured by rss) whilst only slightly affecting
the penalty function. The L-curve method proposes that the
optimum value of l corresponds to the point of maximum
curvature on the p against rss plot, as shown in Fig. 7.
Whilst the L-curve method has been used in this paper, it is
worth noting that other mathematical methods for choosing l
have also been proposed. These include generalised cross-
validation (GCV),45,46 singular value decomposition,47 and
the discrepancy principle method.48,49
The results of simulating the noisy TPD data in Fig. 6(a) using
regularisation are shown in Fig. 8. When eqn (34) is chosen as
the penalty function, it can be seen that there are multiple peaks
Fig. 8 Analysis of the TPD data in Fig. 6(a) using the regularisation
in the optimised distribution function. The value of l chosen by technique with the parameter l chosen using the L-curve method. (a)
L-curve method for this penalty function is too low, resulting in Using the regularisation penalty function based on the square of the
an under-regularised distribution. There is a danger that a user area under the distribution function (eqn (34)). (b) Using the regular-
might infer that there are more than two types of adsorption site isation penalty function based on the square of the second differential
present from the analysis of the results. of the distribution function (eqn (35)).

This journal is 
c the Owner Societies 2008 Phys. Chem. Chem. Phys., 2008, 10, 1688–1696 | 1695

present, particularly if the peaks have different linewidths. The
general rule discovered is that the penalty function based on
the square of the second derivative is preferable for analysing
TPD data, at least when the L-curve method is being used to
select the regularisation parameter l.
6. Discussion and conclusions
This paper has reviewed methods of analysing TPD data for
samples obeying first-order kinetics when the parameters A
Published on 07 February 2008. Downloaded by Syracuse University on 27/11/2013 17:31:55.
and E for each adsorption site are independent of coverage.
Four significant new proposals are described. Firstly, an
improved equation for E*(T) is proposed when using the
condensation approximation to gain an initial estimate of
the N0(E) distribution. Secondly, an equation that gives an
estimate of the N0(E) distribution when A is a function of E is
described. Thirdly, the interrupted TPD experiment is demon-
strated to give a reliable measurement of the parameter A for
the peak in the distribution. Fourthly, regularisation using a
penalty function based on the square of the second derivative
of the distribution is shown to be the preferred technique for
processing data containing significant experimental noise.
Some difficulties remain. For instance, when experimental
TPD results show a single broad peak, it is still not generally
possible to establish whether the parameter A is a constant or
a function of E. Different N0(E) distributions can be generated
for these two cases that fit the experimental data equally well;
this is the case even when interrupted TPD experiments and
varying heating rate experiments are performed. When more
than one peak can be identified, it is possible by either
interrupted TPD or varying heating rate experiments to
determine the parameter A for each peak.
A further difficulty is the methods in this paper have
assumed that A and E are independent of coverage. This
may apply at low loadings, and can apply at high loadings
as well for adsorption in porous solids. However, this assump-
tion should be tested experimentally where possible, for in-
stance by repeating the experiment at a different initial
loading. However, this is not always straightforward, particu-
larly for porous catalyst samples. When not at saturation
loading, molecules will normally be adsorbed at the strongest
binding sites, and so the N0(E) distribution will differ from
Ns(E) as described by eqn (10). Further there may be a
difficulty in establishing equilibrium throughout the whole
sample at the starting temperature—unless suitable precau-
tions are taken, molecules might be adsorbed just at the outer
edges of the particle, rather than being adsorbed throughout it.
Notwithstanding these difficulties, it is clear that TPD
is a useful technique for the characterisation of adsorption sites
within catalysts, and that useful quantitative infor-
mation can be derived from it using the methods described in
this paper.
References
1 D. A. King, Surf. Sci., 1975, 47, 384.
2 J. L. Falconer and J. A. Schwarz, Catal. Rev. Sci. Eng., 1983, 25, 141.
3 J. W. Niemantsverdriet, Spectroscopy in Catalysis: an Introduction,
Wiley-VCH, Weinheim, 2000.
4 R. J. Gorte, Catal. Today, 1996, 28, 405.
1696 | Phys. Chem. Chem. Phys., 2008, 10, 1688–1696
View Article Online
5 K. M. Kanervo, T. J. Keskitalo, R. I. Slioor and A. O. I. Krause, J.
Catal., 2006, 238, 382.
6 X. Xia, J. Strunk, S. Litvinov and M. Muhler, J. Phys. Chem. C,
2007, 111, 6000.
7 R. A. Demmin and R. J. Gorte, J. Catal., 1984, 90, 32.
8 E. Tronconi and P. Forzatti, J. Catal., 1985, 93, 197.
9 E. Tronconi and P. Forzatti, Chem. Eng. Sci., 1987, 42, 2779.
10 C. M. Chan, R. Aris and W. H. Weinberg, Appl. Surf. Sci., 1978, 1,
360.
11 R. Spinicci, J. Therm. Anal., 1984, 29, 1061.
12 A. M. de Jong and J. W. Niemantsverdriet, Surf. Sci., 1990, 233,
355.
13 S. Zhou, G. Schulz-Ekloff and D. Popovic, J. Therm. Anal., 1994,
41, 767.
14 E. Kautto, J. Kuhalainen and M. Manninen, Phys. Scr., 1997, 55,
628.
15 R. Spinicci, Thermochim. Acta, 1997, 87, 296.
16 E. G. Seebauer, A. C. G. Kong and L. D. Schmidt, Surf. Sci., 1988,
193, 417.
17 V. M. Gorbachev, J. Therm. Anal., 1975, 8, 349.
18 R. K. Agrawal and M. V. Sivasubramanian, AIChE J., 1987, 33,
1212.
19 C. Junmeng, H. Fang, Y. Weiming and Y. Fusheng, Chem. Eng. J.
(Amsterdam, Neth.), 2006, 124, 15.
20 R. Kress, Linear Integral Equations, Springer-Verlag, Berlin, 1989.
21 P. C. Hansen, Inverse Problems, 1992, 8, 849.
22 W. Rudzinski and D. M. Everett, Adsorption of Gases on Hetero-
geneous Surfaces, Academic Press, London, 1992.
23 G. F. Cerofolini and N. Re, J. Colloid Interface Sci., 1995, 174,
428.
24 V. I. Bogillo and V. P. Shkilev, J. Therm. Anal. Calorim., 1999, 55,
483.
25 E. G. Seebauer, Surf. Sci., 1994, 316, 391.
26 P. Kowalczyk, K. Kaneko, A. P. Terzyk, H. Tanaka, H. Kanoh
and P. A. Gauden, J. Colloid Interface Sci., 2005, 291, 334.
27 V. P. Zhdanov, Surf. Sci. Rep., 1991, 12, 183.
28 G. C. Bond, Catal. Today, 1999, 49, 41.
29 H. Lynggaard, A. Andreasen, C. Stegelmann and P. Stoltze, Prog.
Surf. Sci., 2004, 77, 71.
30 C. Costa, J. M. Lopes, F. Lemos and F. Ramôa Ribeiro, J. Mol.
Catal. A: Chem., 1999, 144, 221.
31 C. Costa, I. P. Dzikh, J. M. Lopes, F. Lemos and F. Ramôa
Ribeiro, J. Mol. Catal. A: Chem., 2000, 154, 193.
32 J. B. Miller, H. R. Siddiqui, S. M. Gates, J. N. Russell, Jr, J. T.
Yates, Jr, J. C. Tully and M. J. Cardillo, J. Chem. Phys., 1987, 87,
6725.
33 D. L. S. Nieskens, A. P. van Bavel and J. W. Niemantsverdriet,
Surf. Sci., 2003, 546, 159.
34 W. Rudzinski, T. Borowiecki, T. Panczyk and A. Dominko, Adv.
Colloid Interface Sci., 2000, 84, 1.
35 K. Koch, B. Hunger, O. Klepel and M. Heuchel, J. Catal., 1997,
172, 187.
36 M. C. Ma, T. C. Brown and B. S. Haynes, Surf. Sci., 1993, 297,
312.
37 M. Arai, M. Fukushima and Y. Nishiyama, Appl. Surf. Sci., 1996,
99, 145.
38 S. O. Vásquez, J. Chem. Soc., Faraday Trans., 1992, 88, 2051.
39 H. G. Karge, V. Dondur and J. Weitkamp, J. Phys. Chem., 1991,
95, 283.
40 B. Hunger, M. Heuchel, L. A. Clark and R. Q. Snurr, J. Phys.
Chem. B, 2002, 106, 3882.
41 S. A. Scott, J. S. Dennis, J. F. Davidson and A. N. Hayhurst,
Chem. Eng. Sci., 2006, 61, 2339.
42 A. N. Tikhonov, Dokl. Akad. Nauk SSSR, 1943, 39, 195.
43 A. N. Tikhonov and V. A. Arsenin, Solutions of Ill-posed Problems,
Winston, Washington, 1977.
44 P. C. Hansen, SIAM Rev., 1992, 34, 561.
45 G. Wahba, SIAM J. Numer. Anal., 1977, 14, 651.
46 J. D. Wilson, J. Mater. Sci., 1992, 27, 3911.
47 M. von Szombathely, P. Bräuer and M. Jaroniec, J. Comput.
Chem., 1992, 13, 17.
48 V. A. Morozov, Methods for Solving Incorrectly Posed Problems;
Springer-Verlag, New York, 1984.
49 K. G. Hollingsworth and M. L. Johns, J. Colloid Interface Sci.,
2003, 258, 383.
This journal is 
c the Owner Societies 2008