This action might not be possible to undo. Are you sure you want to continue?

Welcome to Scribd! Start your free trial and access books, documents and more.Find out more

**Contents lists available at ScienceDirect
**

Chemical Engineering Journal

j our nal homepage: www. el sevi er . com/ l ocat e/ cej

Experimental errors in kinetic tests and its inﬂuence on the precision of

estimated parameters. Part I—Analysis of ﬁrst-order reactions

André L. Alberton

a

, Marcio Schwaab

b

, Martin Schmal

a

, José Carlos Pinto

a,∗

a

Programa de Engenharia Química/COPPE, Universidade Federal do Rio de Janeiro, Cidade Universitária, CP: 68502, Rio de Janeiro, RJ 21941-972, Brazil

b

Departamento de Engenharia Química, Universidade Federal de Santa Maria, Av. Roraima, 1000, Cidade Universitária, Santa Maria, RS 97105-900, Brazil

a r t i c l e i n f o

Article history:

Received 7 May 2009

Received in revised form 31 July 2009

Accepted 16 August 2009

Keywords:

Experimental error

Kinetic analysis

First-order reaction

Precise parameter estimation

a b s t r a c t

The proper characterization of the experimental errors is essential for the correct evaluation of esti-

mated model parameters, model ﬁt and model predictions based on kinetic rate expressions. However, it

is commonto ignore the inﬂuence of experimental errors during kinetic studies due to difﬁculties to char-

acterize howexperimental errors depend on the reaction conditions. The behavior of experimental error

depends on the speciﬁc features of the experimental system; however, in many cases the main sources of

experimental errors are the unavoidable oscillations of the input variables. This work analyzes how the

experimental errors affect kinetic studies based on catalytic tests when oscillations of the input variables

are the main sources of uncertainties. The ﬁrst part of this work assumes that the reaction rate can be

describedaccuratelyas aﬁrst-order reactioninaPFR. Analytical expressions arederivedfor thevarianceof

the reactant conversionindistinct scenarios andare usedtoanalyze the quality of the obtainedparameter

estimates. It is shown here that the conversion variances can be described as functions of the mea-

sured conversion values, normally presenting a point of maximum for conversion values in the range of

0.6<X<1.0 when observed experimental ﬂuctuations are controlled by the ﬂuctuations of the input vari-

ables. Constant conversion variances should be expected only when ﬂuctuations are controlled by ana-

lytical conversion measurements. As a consequence, optimum parameter estimation may be performed

either with differential or integral methods, depending on the behavior of the conversion variances.

© 2009 Elsevier B.V. All rights reserved.

1. Introduction

Kinetic expressions are useful tools for the evaluation, compari-

sonandpredictionof catalyst activity, allowingfor theoptimization

of catalyst performance both in laboratory and industrial scales.

The analysis of models and kinetic parameters may also help to

clarify catalytic routes and understand reaction paths, as usually

performed in kinetic studies [1,2].

In order to analyze catalyst performance and obtain kinetic

parameters, differential methods have normally been employed.

The main reason for the widespread usage of this kind of kinetic

analysis is the simple calculation of reaction rates (which does not

require numerical integration of balance equations), based on the

assumptionof negligible variationof concentration(or partial pres-

sure) of the reactants. In these systems, the conversion is kept at

low values (<20%), so that it is possible to assume that the kinetic

rate is constant along the reactor length [2,3].

On the other hand, the use of differential methods may not be

appropriate for kinetic studies. First, the conditions required to

∗

Corresponding author. Tel.: +55 21 25628337; fax: +55 21 25628300.

E-mail address: pinto@peq.coppe.ufrj.br (J.C. Pinto).

keep the conversion at lowvalues may impose limits on the exper-

imental ranges of operation variables, such as volumetric ﬂow,

partial pressure of reactants and reaction temperature in order to

keep the conversion at low values. Second, obtained conversion

values may be very similar to each other, making the statistical dis-

crimination of available data very difﬁcult, when the experimental

errors are taken into consideration. These restricted conditions are

not imposed for the use of integral methods, since they can be used

at any conversion level.

Conversely, statistical experimental designs have been pro-

posed to allow for adequate variation of operation variables and

process responses, allowing for more precise estimation of model

parameters and identiﬁcation of experimental effects. In this con-

text the use of differential methods in order to avoid the numerical

integration of balance equations may not be justiﬁed, as sophisti-

cated numerical techniques and computer software are available

nowadays for the immediate integration of mass and energy bal-

ances. Therefore, it is not necessary to assume that operation

conditions are constant along the reactor volume and/or during

the experimental run in order to analyze obtained kinetic data.

The estimation of model parameters is usually performed

through minimization of a metric function that represents the

distance between the experimental and the calculated response

1385-8947/$ – see front matter © 2009 Elsevier B.V. All rights reserved.

doi:10.1016/j.cej.2009.08.012

A.L. Alberton et al. / Chemical Engineering Journal 155 (2009) 816–823 817

Nomenclature

B sensitivity matrix

b

ij

ij-element of sensitivity matrix

C constant value deﬁned in Eq. (12)

S objective function

T temperature

v volumetric ﬂow rate

V

covariance matrix of parameter estimates

V

y

experimental covariance matrix

W catalyst mass

X conversion

x inlet variables

y outlet or response variables

˛ packing quality of the catalyst bed

model parameter

2

XA

variance of analytical measurement of variance

2

W

variance of catalyst mass

2

X

variance of conversion

2

y

variance of outlet variable

2

Â

variance of parameter estimate

2

T

variance of temperature

2

v

variance of volumetric ﬂow rate

values. Most kinetic studies do not take into account the variance

of the measured values and, in this case, statistical tests cannot be

applied properly to evaluate the model adequacy and the quality

of model parameters and of the model ﬁt [4].

A more fundamental procedure is based on the assumption that

theexperimental errors followthenormal distribution, themodel is

perfect and experiments are well done. In this case, it can be shown

that the parameter values can be obtained through the minimiza-

tion of the likelihood function, deﬁned as [5,6]:

S =

i

(y

e

i

−y

c

i

(x, ))

2

2

yi

(1)

where the index i indicates the experimental point; y

e

i

represents

the values of experimental values of process responses; y

c

i

repre-

sents the calculated values of process responses, calculated as a

functionof inlet designvariables (x) andmodel parameters (); and

2

yi

is the variance of process response measurements. (Rigorously,

the determination of the probability distribution of a particular

measurement can be obtained only through extensive replications.

However, most times this is not feasible, due to time and cost con-

straints. The normal distributionis oftenassumed because it allows

for close representation of many measurement distributions and

because the mathematical treatment of the data is simple, as only

the mean value and the variance are needed. The proposed analysis

can be extended to non-normal problems with the help of more

involving numerical schemes [5,6]; however, this is beyond the

scope of the present manuscript.)

In Eq. (1), each difference between experimental and model

response values is weighted by the respective experimental error.

So, experimental points that are subject to large experimental

uncertainties do not inﬂuence the objective function signiﬁcantly.

As a consequence, the distance between model responses and

experimental data tends to be smaller when the experimental

variances are small. Therefore, the correct characterization of

experimental error is a fundamental issue during estimation of

model parameters and analysis of experimental data. This may

be very important because the precision of experimental mea-

surements is transferred to the precision of kinetic parameters,

which in turn are transferred to precision of output model pre-

dictions.

Usually, the experimental error is completely ignored in kinetic

problems. For this reason, estimation of model parameters and

model building are carried out with the help of the well-known

least squares function [7–10]. The underlying hypothesis that jus-

tiﬁes the use of the least squares function is that the experimental

error is the same for all measurements. In some cases, the vari-

ance in Eq. (1) is considered equal or an arbitrary fraction of the

measured variable [11,12].

The resistance that many offer for proper characterization of

experimental errors is understandable. As a matter of fact, as

already said the correct characterization of experimental errors

usually demands a large number of replicates, which may be many

times unfeasible, due to existence of time and cost constraints.

Therefore, techniques that allowfor the simultaneous evaluationof

experimental errors and minimization of the experimental efforts

are welcome [4,5,13–16].

The parameter uncertainties can be represented in terms of the

covariance matrix of parameter uncertainties (V

), deﬁned as [4,5]:

V

= (B

T

· V

−1

y

· B)

−1

(2)

whereV

y

is thecovariancematrixof experimental process response

measurements and B is the sensitivity matrix, deﬁned as:

B = [b

ij

] =

_

∂y

j

∂Â

i

_

(3)

Inorder to obtainprecise parameter values, a speciﬁed metric of

the matrix V

must be minimized, which can be used for the design

of the experimental conditions. For a ﬁrst-order reaction (A→B,

r

A

=−·P

A

) where the output process response is the conversion

(X) of reactant A, the precision of parameter after NE experiments

can be obtained as:

1

2

Â

=

NE

i=1

(∂X

i

/∂Â)

2

2

X

i

(4)

For one to use Eq. (4) in order to characterize the precision of

the kinetic parameter and of the calculated model responses, it is

necessary ﬁrst to characterize the experimental error of experi-

mental data. However,

2

X

i

may depend on speciﬁc features of the

analyzed catalyst system, which may involve the precision of reac-

tant feed rates, the control of reactor conditions, the precision and

accuracyof off-line laboratoryanalyses, etc. The technical literature

has largely overlooked this issue.

It seems clear that experimental errors are not necessarily con-

stant throughout the experimental range. For example, Pereira et

al. [17] studied the gibbsite dissolution in the presence of NaOH

and concluded (based on an extensive experimental plan), that

the conversion variances depended on the measured conversion

values. The authors ﬁtted an empirical model to available data

to describe the conversion variance as a function of the conver-

sion.

Larentis et al. [18] investigated the sources of error in reaction

systems involving gaseous species and on-line chromatographic

analyses. The proposed error analysis allowed for the identiﬁca-

tion of the predominant reactions that were occurring in different

experimental ranges. Besides, it was shown that the main sources

of error were probably associated with the catalyst activity, which

can be affected by the catalyst itself, the preparation of the catalyst

bed or the operational conditions of the reactor.

Gutierrez and Danielson [19] analyzed the error structure of

dynamic enzymatic assays when the initial enzyme and substrate

concentrations are subject to uncertainties. They showed that the

variance of the substrate conversion presents a heteroscedastic

818 A.L. Alberton et al. / Chemical Engineering Journal 155 (2009) 816–823

behavior inthese cases, reaching a maximumvalue around60–80%,

which could be described with the help of a hybrid empirical-

phenomenological model.

Based on previous remarks, it may be concluded that the proper

characterization of experimental errors can be extremely useful for

those who are working in the ﬁeld of chemical kinetics and catal-

ysis. Besides, it can also be said that experimental errors are not

necessarily constant in the experimental region for better under-

standing of the reaction system, as shown experimentally and

explored theoretically. For these reasons, model equations are pro-

posedinthe present worktorepresent howexperimental variances

depend on reaction conditions in typical catalyst tests. The main

pursued objective is the development of a theoretical framework

that may allow for simultaneous characterization of variances and

minimization of the number of replicate runs. In this manner, more

involving and rigorous statistical analyses of available data may

become possible, even when the number of replicates is small.

It is assumed here that a ﬁrst-order reaction is carried out in

a plug ﬂow reactor at isothermal conditions. It is also assumed

that diffusion constraints are absent. These constraints should not

be overemphasized because catalyst tests are normally performed

isothermically in plug ﬂow reactors and because ﬁrst-order reac-

tion rate expressions are frequently used for model building and

kinetic analysis. Initially, ﬁrst-order kinetics is used frequently as

a ﬁrst model building stage. Second, ﬁrst-order kinetics may be

regarded as a local linearization of more involving kinetic mech-

anisms. Third, the main objective of the present paper is showing

that experimental errors may depend quite considerably on the

operation conditions, which has been largely overlooked in the

ﬁeld of kinetics and catalysis. In order to illustrate this point, com-

plex kinetic schemes are not necessary, as shown in the present

manuscript. In spite of the previous remarks, it is important to say

that the analysis of more complex reaction systems can also be

performed with the help of the proposed differential analysis.

It is shown here that the conversion variances can be described

as functions of the measured conversion values, presenting a

point of maximum conversion variance in the conversion range

of 0.6<X<1.0 when observed experimental ﬂuctuations are con-

trolled by the ﬂuctuations of the input variables. As presented in

the following sections, constant conversion variances should be

expected only when ﬂuctuations are controlled by errors in the

analytical measurement of the outlet variables, as the conversion.

As a consequence, optimum parameter estimation may be per-

formed either with differential or integral methods, depending on

the behavior of the conversion variances.

2. Conversion variance for a ﬁrst-order reaction

Let us assume that a catalytic system is designed to operate

isothermically. At the desired reaction temperature (T

D

), volu-

metric ﬂow rate (v

D

) and catalyst bed mass (W

D

), the expected

conversion is X

D

. However, in this hypothetical system, the experi-

mental design variables are subject to ﬂuctuations, so that the real

temperature (T), volumetric ﬂow rate (v) and mass of catalyst (W)

are different from the desired design values, leading to a different

conversion (X).

Expressing the conversion as a function of the design variables

in terms of a Taylor series truncated after the ﬁrst term:

X ≈ X

D

+

_

_

∂X

∂T

_¸

¸

¸

¸

T=T

D

_

(T −T

D

) +

_

_

∂X

∂v

_¸

¸

¸

¸

v=v

D

_

(v −v

D

)

+

_

_

∂X

∂W

_¸

¸

¸

¸

W=W

D

_

(W −W

D

) (5)

or

X ≈

_

_

∂X

∂T

_¸

¸

¸

¸

T=T

D

_

T +

_

_

∂X

∂v

_¸

¸

¸

¸

v=v

D

_

v

+

_

_

∂X

∂W

_¸

¸

¸

¸

W=W

D

_

W (6)

Squaring and averaging both sides of Eq. (6), in order to obtain

the conversion variance (≈ is replaced by =, observing that the

expected value of z

2

is the variance

2

z

, where z is any variable):

2

X

=

__

_

∂X

∂T

_¸

¸

¸

T=T

D

_

T +

_

_

∂X

∂v

_¸

¸

¸

v=v

D

_

v +

_

_

∂X

∂W

_¸

¸

¸

W=W

D

_

W

_

2

(7)

Assuming that random ﬂuctuations of the design variables are

not correlated to each other:

2

X

=

_

_

∂X

∂T

_¸

¸

¸

¸

T=T

D

_

2

2

T

+

_

_

∂X

∂v

_¸

¸

¸

¸

v=v

D

_

2

2

v

+

_

_

∂X

∂W

_¸

¸

¸

¸

W=W

D

_

2

2

W

(8)

Eq. (8) is frequently used in statistics to perform robust pro-

cesses or product design [20].

For a ﬁrst-order reaction, the conversion can be written as:

X = 1 −exp

_

−Â

0

· e

−(Â

1

/T)

W

v

_

(9)

Therefore,

∂X

∂T

= −exp

_

−Â

0

· e

−(Â

1

/T)

·

W

v

_

. ¸¸ .

1−X

Â

0

· e

−(Â

1

/T)

·

W

v

. ¸¸ .

ln(1−X)

Â

1

T

2

(10a)

∂X

∂v

= −exp

_

−Â

0

· e

−(Â

1

/T)

·

W

v

_

. ¸¸ .

(1−X)

−Â

0

· e

−(Â

1

/T)

·

W

v

. ¸¸ .

ln(1−X)

1

v

(10b)

∂X

∂W

= −exp

_

−Â

0

· e

−(Â

1

/T)

·

W

v

_

. ¸¸ .

(1−X)

−Â

0

· e

−(Â

1

/T)

·

W

v

. ¸¸ .

ln(1−X)

1

W

(10c)

Inserting Eqs. (10) in Eq. (8) and rearranging:

2

X

= [(1 −X) ln(1 −X)]

2

·

_

Â

2

1

T

4

2

T

+

2

v

v

2

+

2

W

W

2

_

(11a)

If the evaluation of X is also subject to a measurement error with

variance

2

XA

, then Eq. (11a) should be written as:

2

X

= [(1 −X) ln(1 −X)]

2

·

_

Â

2

1

T

4

2

T

+

2

v

v

2

+

2

W

W

2

_

+

2

XA

(11b)

Now it is necessary to evaluate the terms involving the design

variables and its variances. If the analyzed experimental range is

not very large and/or the relative error measurements are con-

stant, the terms inside the brackets may be regarded as constant

and represented by C. In this case, Eq. (11) can be written as:

2

X

= C · [(1 −X) ln(1 −X)]

2

(12a)

2

X

= C · [(1 −X) ln(1 −X)]

2

+

2

XA

(12b)

Eq. (12) presents interesting characteristics, as shown in Fig. 1,

for an arbitrary C value. In Fig. 1A, the conversion variance presents

amaximumvalueat Xequal to0.63, withinﬂectionpoints at Xequal

A.L. Alberton et al. / Chemical Engineering Journal 155 (2009) 816–823 819

Fig. 1. Variance of the conversion (A) and relative conversion variance (B) as a

function of conversion for Eq. (11).

to 0.32 and 0.93. When conversion is equal to zero, conversion vari-

ance is also equal to zero because the reactant concentration does

not change and the reactor output is not subject to process ﬂuc-

tuations. As the conversion increases, the conversion variance also

increases, since variations of the reaction conditions perturb the

measured conversion values and the inﬂuence of process ﬂuctua-

tions is larger when reaction rates are high and the magnitude of

conversion values increases. However, the continuous increase of

conversion leads to appearance of a point of maximum conversion

variance. After this point of maximum, the variance of conversion

decreases, as reactionrates at highconversionvalues become lower

and perturbations of the operating conditions cause less signiﬁcant

effects on the observed reaction rates. If the relative conversion

variances are considered (that is, the conversion variance divided

by squared conversion values), as presented in Fig. 1B, one can

observe that the relative conversion variance decreases almost lin-

early with conversion, given the higher reaction rates observed at

lower conversions.

It must be notedthat Gutierrez andDanielson[19,21] developed

an equation for the variance of substrate conversion in dynamic

enzymatic assays that is similar to Eq. (12). This clearly shows that

Eq. (12) may ﬁnd application in distinct experimental systems.

Eq. (12) can be very useful because it describes the conversion

variance in terms of the measured conversion values. Therefore,

if replicates are available at a certain experimental condition, it is

possible to determine the constant C:

2

X

= (

2

X0

−

2

XA

) ·

[(1 −X) ln(1 −X)]

2

[(1 −X

0

) ln(1 −X

0

)]

2

+

2

XA

(13)

where

2

X0

represents the available conversionvariance and X

0

rep-

resents the average conversion obtained from available replicates.

Obviously, the characterization of the variance in a single point

may lead to inaccurate representation of conversion variances in

the whole experimental range. Nevertheless, it seems much more

reasonable to use Eq. (13) for more fundamental data analysis than

to ignore the effect of ﬂuctuations of the design variables on the

conversion variances.

It must be pointed that the parameter C will not be constant if

the experimental region of design variables is large and variances

of the design variables are not proportional to measured values.

Therefore, one must consider that Eq. (13) can constitute a useful

approximation of the real variance behavior.

3. Optimum estimation of the kinetic parameter

Temperature effects are neglected in the following paragraphs.

Therefore, the well-known Arrhenius equation is not used to repre-

sent the temperature dependence of the kinetic rate constant. This

assumption will be relaxed in the following sections. In this case,

using Eq. (4) to obtain 1/

2

Â

, it is possible to write:

1

2

Â

=

(∂X/∂Â)

2

2

X

(14)

The partial derivative of conversion in respect to the kinetic

constant is:

∂X

∂Â

= −exp

_

−Â ·

W

v

_

. ¸¸ .

(1−X)

⎡

⎢

⎢

⎣

−Â ·

W

v

. ¸¸ .

ln(1−X)

1

Â

⎤

⎥

⎥

⎦

(15)

This can also be written in terms of the conversion, as:

Â

2

2

Â

=

(1 −X)

2

ln

2

(1 −X)

2

X

(16)

The most precise estimation of parameter can be achieved

when the variance of the model parameter has a minimum value.

This is equivalent to maximizing the right-hand side of Eq. (16).

In order to do that, three cases are considered below: (i) con-

stant conversion variance (least squares function), meaning that

uncertainties are controlled by off-line laboratory analyses; (ii)

conversion variance described by Eq. ((12)–(13)), meaning that

uncertainties are controlled by the constant precision of the input

variables; and(iii) conversionvariance describedbyEq. (11), mean-

ing that uncertainties are controlled by the varying precision of the

input variables.

3.1. Constant conversion variance

Based on Eq. (11), the conversion variance is constant when

uncertainties are controlled by off-line laboratory analyses. In this

case, maximizing Eq. (16) is equivalent tomaximizing the following

equation:

max

1

2

Â

= max (1 −X)

2

ln

2

(1 −X) (17)

which has a maximumplaced at the value for X equal to 0.63. Based

on this result, it may be concluded that, when the experimental

error is constant (and the analytical errors are predominant), it is

preferable to use integral methods for model building and precise

and accurate parameter estimation. This may be regarded as a very

important conclusion, given the widespread usage of differential

kinetic analysis in the ﬁeld of catalysis.

However, according to Eq. (12), this is the point where the

conversion variance is maximum (when it is assumed that the

conversion variance is not constant). This constitutes a very impor-

tant result, because this means that the experimental optimum

design(whentheconversionvarianceis constant) indicates that the

optimal experiment is placed at the region where the conversion

variance attains its highest value (when the conversion variance is

not constant). Therefore, if one assumes that conversion variances

areconstant andtheassumptionis not correct, experimental design

820 A.L. Alberton et al. / Chemical Engineering Journal 155 (2009) 816–823

will erroneously recommend the experimental region where mea-

surement errors attain the largest values. This clearly illustrates

the importance of taking into account the proper interpretation of

experimental errors in order to avoid very bad data analysis and

experimental design.

3.2. Conversion variance described by Eq. (12)

If the conversion variance is described by Eq. (12), meaning that

uncertainties are controlled by the constant precision of the input

variables, the error of parameter estimate can be determined as:

Â

2

2

Â

=

(1 −X)

2

ln

2

(1 −X)

(1 −X)

2

ln

2

(1 −X) · C +

2

XA

=

1

C +(

2

XA

/(1 −X)

2

ln

2

(1 −X))

(18)

If the analytical conversion variance

2

XA

is neglected, then the

parameter variance does not depend explicitly on the experimen-

tal conditions (although it depends on the obtained conversion,

because it is a function of , which is obtained as a function of the

observed conversion), which is very surprising! However, if

2

XA

is

considered, then Eq. (12) must be maximized in order to minimize

the parameter variance. Once more, this means that the optimum

experimental conditions are the ones that leadtoconversionvalues

around 0–63%, when the reaction follows the ﬁrst-order kinetics.

As observed in the previous case, this clearly shows that the kinetic

tests performed through differential methods may not be adequate

for kinetic analysis, as the parameter variance approaches its max-

imum value when conversion approaches zero! Therefore, it may

be concluded once more that it is preferable to use integral meth-

ods for model building and precise parameter estimation. Again,

this may also be regarded as a very important conclusion, given

the widespread usage of differential kinetic analysis in the ﬁeld of

catalysis.

3.3. Conversion variance described by Eq. (11)

If the conversion variance is described by Eq. (11), meaning that

uncertainties are controlled by the varying precision of the input

variables, the parameter variance can be written as:

Â

2

2

Â

=

1

_

Â

1

2

T

4

2

T

+

2

v

v

2

+

2

W

W

2

_

+

2

XA

(1−X)

2

ln

2

(1−X)

(19)

It is necessary to perform the maximization of Eq. (19) in order

toobtainthebest experimental conditions for estimatingthemodel

parameter. If thevariances of input variables areconstant, theright-

hand side of Eq. (19) presents a monotonic behavior with respect

to volumetric ﬂow rate and mass of catalyst. The dependence of

Eq. (19) with respect to temperature is more complex because the

model parameter certainly depends on the reaction temperature.

In order to show the main trends, it is preferable to evaluate the

variance behavior when the ﬂuctuations are controlled by a single

input variable.

3.3.1. The ﬂow rate effect is predominant

Considering that the effects of the temperature and catalyst

mass in the conversion variance are negligible and only the effect

of the ﬂowrate is important, fromEq. (11), the conversion variance

can be written as:

2

X

= [(1 −X) ln(1 −X)]

2

·

_

2

v

v

2

_

+

2

XA

(20)

Fig. 2.

2

X

· W as a function of conversion for Eq. (22).

The volumetric ﬂow rate can be expressed in terms of conversion

as:

v = −

Â · W

ln(1 −X)

(21)

Inserting Eq. (21) into Eq. (20):

2

X

· W

2

=

2

v

Â

2

· (1 −X)

2

ln

4

(1 −X) +

2

XA

· W

2

(22)

If the analytical error can be neglected, the conversion variance

canbeevaluatedbytheexpression[(1−X)

2

ln

4

(1−X)], presentedin

Fig. 2 as a function of X. Considering the variance of the volumetric

ﬂow constant, the right-hand side of Eq. (22) presents a maximum

at X equal to 0.86. Again, for lowand high values of conversion, the

conversion variance tends to be lower than for intermediate values

of conversion.

In order to determine the optimal experimental run for pre-

cise parameter estimation, the conversion variance deﬁned in Eq.

(22) must be inserted into Eq. (16) (or one can simplify Eq. (19) by

assuming that the effect of the ﬂow rate is predominant) resulting

in:

Â

2

2

Â

=

v

2

2

v

(23)

If thevarianceof thevolumetric ﬂowrateis constant, theparam-

eter is estimated with more precision when the volumetric ﬂow

is maximum. Eq. (23) indicates that is estimated with more pre-

cision when the conversion is lower. Again, this may be regarded as

a very important result, as it indicates that the correct evaluation

of the experimental error changes completely the determination of

the best region for precise parameter estimation. Therefore, when

the variance of the feed ﬂow rate controls the precision of output

responses, the use of differential methods in the kinetic analysis

may be considered, since at high volumetric ﬂow the conversion is

usually lowand the reaction rate can be regarded as constant along

the reactor length.

3.3.2. The temperature effect is predominant

This may be the case of highly endo/exothermic reactions or

systems wherethetemperaturecontrol is difﬁcult. Consideringthat

only the deviations inthe temperature are important, fromEq. (11),

the conversion variance can be written as:

2

X

= [(1 −X) ln(1 −X)]

2

_

Â

2

1

T

4

2

T

_

(24)

The conversion variance cannot be expressed as a function of

conversion only, as it also depends on the ﬂowrate and mass of cat-

alyst. If reactor temperature is assumedtobe constant, the behavior

of the variance conversion is similar to the one illustrated in Fig. 1.

In this case, the conversion variance is higher for lower tempera-

A.L. Alberton et al. / Chemical Engineering Journal 155 (2009) 816–823 821

Fig. 3. Variance of the conversion a function of conversion: (A) for three temperatures and several spatial times; and (B) for three spatial times and several temperatures.

tures, as clearly seen in Eq. (24). In order to illustrate this effect, let

us assume that:

Â = 100 · exp

_

−

3000

T

_

(25)

Based on Eq. (25), the conversion value and the respective con-

version variance were calculated for distinct temperatures and

residence times. The obtained results are presented in Fig. 3. When

temperature is varied, the behavior of the conversion variance may

be very complex and depends on the kinetic parameters and the

spatial times. Despite that, in all cases intermediate conversion

value leads to larger conversion variances.

In order to determine the optimal conditions for precise param-

eter estimation, Eq. (19) can be written as:

Â

2

2

Â

=

1

__

Â

2

1

T

4

_

2

T

_ (26)

Maximization of Eq. (26) can be achieved at high temperatures,

meaning that when the error is controlled by temperature effects,

experiments should be performed at the maximumallowable tem-

perature values in order to minimize the variance of the kinetic

parameter. One must observe that high temperature conditions

also lead to lower values of the conversion variance, as shown

in Fig. 3. This is very interesting, since when the conversion vari-

ance is not assumed constant and the temperature effect in the

experimental error is predominant, thedesignedexperiments must

be performed at experimental conditions where conversion vari-

ance has lower values. One must observe that when conversion

variance was considered constant (Section 3.1) the experimental

condition that minimizes the parameter estimate variance was the

same where conversion variance reaches a maximum value. Once

more, this shows the importance of considering the real behav-

ior of the conversion variance, instead of using the much simpler

assumption of constant conversion variance during data analy-

sis.

As it becomes clear that the precision of estimated parameters

depends not only on the experimental precision, but also on the

selected experimental conditions. When the error is controlled by

temperature effects, it can be concluded that experiments should

be performed at the maximum allowable temperature values.

3.3.3. The catalyst mass effect is predominant

Considering now that the ﬂuctuations in the catalyst mass are

predominant, Eq. (19) can be expressed as:

1

2

Â

=

1

2

W

W

2

Â

2

(27)

The variance of the catalyst mass is probably constant and

related to the precision of the balance used to weigh the mass.

According to Eq. (27), the optimal parameter estimation can be

achieved when the mass of catalyst is equal to the maximum

allowed value. Therefore, the parameter variance is smaller when

the catalyst mass is larger. As one might already expect, the weigh-

ing of lowcatalyst masses may introduce large experimental errors

in the analysis. As conversion tends to increase when the mass of

catalyst increases, it can be said indirectly that the integral method

is preferable when the conversion variance is controlled by ﬂuc-

tuations of the catalyst mass, since at high conversion values the

differential method is not applicable.

3.4. The packing effect

It must be pointed out that packing of the catalyst bed does not

depend directly on the design variables and should not be con-

sidered separately. The random packing of the catalyst bed may

promote the occurrence of stagnant zones and short-cuts, resulting

in different contact times between the ﬂuid phase and the cata-

lyst particles. For PFR reactors, this effect can be modeled in terms

of two parallel reactors: one containing the catalyst and the other

formed by void space, as illustrated in Fig. 4 [3]. According to the

proposed model, ˛ (which is quality of the packing of the catalyst

bed) indicates the fraction of ﬂuid that short-cuts the catalyst bed.

˛ values can be estimated from Residence Time Distribution (RTD)

experiments, using a trace step input [3].

The value of ˛ may change from bed to bed, consequently lead-

ing to unavoidable ﬂuctuation of the conversion. For a ﬁrst-order

Fig. 4. Schematic representation of the real catalyst bed and the proposed bed

model.

822 A.L. Alberton et al. / Chemical Engineering Journal 155 (2009) 816–823

Fig. 5. Conversion variance as a function of conversion for several values of ˛ in Eq.

(32).

reaction, the conversion can be written as:

X = (1 −˛)(1 −e

−(Â·W/v(1−˛))

) (28)

As performed previously, the conversion variance can be

expressed as a function of characteristic ﬂuctuations of ˛. So, the

conversion variance can be written as:

2

X

=

_

∂X

∂˛

_

2

2

˛

(29)

As

∂X

∂˛

= −1 +exp

_

−k

0

e

−(E/RT)

·

W

v

0

(1 −˛)

_

+exp

_

−k

0

e

−(E/RT)

·

W

v

0

(1 −˛)

_

k

0

e

−(E/RT)

·

W

v

0

(1 −˛)

(30)

Then

∂X

∂˛

= 1 −

_

1 −

X

1 −˛

_

·

_

1 +ln

_

1 −

X

1 −˛

__

(31)

Therefore

2

X

=

_

X

1 −˛

−

_

1 −

X

1 −˛

_

ln

_

1 −

X

1 −˛

__

2

2

˛

(32)

This equation presents a maximum at:

X =

1 +e

2

+˛(1 −e

2

)

e

2

(33)

For good packing of the catalyst bed, ˛ values must be close to

zero. In this case, the maximum conversion variance would occur

at X equal to 0.86. Fig. 5 shows the conversion variance as a func-

tion of conversion for several values of ˛. It can be observed that

the conversion where the conversion variance attains it maximum

value is shifted towards smaller values, as ˛ increases. Besides, as

˛ describes a short-cut fraction, conversion is limited to 1−˛.

The best region for precise parameter estimation can be derived

as shown previously. In this case, the parameter variance can be

written as:

Â

2

2

Â

=

(1 −X)

2

ln(1 −X)

2

_

X

1−˛

−

_

1 −

X

1−˛

_

ln

_

1 −

X

1−˛

__

2

1

2

˛

(34)

For a given temperature, the maximization of Eq. (34) leads to X

equal to0, as showninFig. 6for different values of ˛. As it is probable

that the quality of the catalyst bed may control the precision of the

catalyst tests frequently, this may be regarded as a very important

result. It indicates that reaction conditions should be selected in

order to give small conversion values, allowing the increase in the

precision of parameter estimates. It also can be observed that these

Fig. 6. Potential of estimation as a function of the conversion for several void frac-

tions of the reactor.

conditions enable the use differential methods and may lead to

more precise parameter estimation when errors are controlled by

packing effects.

3.5. Comparative analysis of the different scenarios

As shown in the previous paragraphs, the optimum experi-

mental region for the estimation of model parameters and model

building depends on the behavior of conversion variances. When

the conversion variance is constant (and is controlled by analytical

techniques), optimum determination of model parameters can be

carried out with integral methods, as the lowest parameter vari-

ances are obtained for conversions values from 0 to 63%. Similar

results can be obtained when the ﬂuctuations induced by random

variation of the design variables do not change considerably in the

designregion. Despite that, the maximumconversionvariances are

observed in the same experimental region, but are compensated by

similar increase of conversion sensitivity to parameter estimation.

When the conversion variances are controlled by the design

variables, then the scenario is not clear. For instance, experiments

can be analyzed through differential methods if conversion vari-

ances are controlled by ﬂowrate or packing ﬂuctuations, or should

be analyzed with integral methods if conversion variances are con-

trolled by temperature of catalyst mass ﬂuctuations.

4. Experimental validation of Eq. (13)

Based on experimental data reported in the literature [17,19], it

can be assumed that Eq.(12) can be used to represent conversion

variances in reaction systems described by kinetic rate expres-

sions that are more complex than a ﬁrst-order reaction. In order

to evaluate the applicability of the present analysis in a real prob-

lem, experimental data reported in the literature were taken as an

example. Pereira et al. [17] studied the dissolution of gibbsite in a

solution of NaOH in a batch reactor. One can readily observe that

similar analysis and equations can be obtained for batch reactors.

The authors carried out replicates at different conditions and used

an empirical expression to ﬁt the conversion variance as a function

of the measured conversion. The empirical equation was:

2

X

= exp(A +B · X +C · X

2

) (35)

where A, B and C are empirical ﬁtting parameters.

Fig. 7 presents the experimental data reported by Pereira et al.

[17] and a best ﬁt obtained for Eq. (13). The conversion variance

was described as a function of the X/X

eq

ratio, where X

eq

is the

equilibriumconversion, equal to 94% at the analyzed experimental

conditions. One must observe that maximumconversion variances

occur at intermediate conversion values, as predicted by the pro-

A.L. Alberton et al. / Chemical Engineering Journal 155 (2009) 816–823 823

Fig. 7. Eq. (13) ﬁtted to the data reported by Pereira et al. [17].

posed analysis. According to the authors, the kinetics of gibbsite

dissolutioncouldbe describedby a reversible expressionwrittenin

terms of the external area of gibbsite particles and hydroxide anion

concentration, which is not a simple ﬁrst-order reaction. Even so,

the adjustment of Eq. (13) to the available experimental data can

be considered satisfactory, as it must be pointed out that the con-

ﬁdence intervals of the true variances obtained fromtriplicates are

very large. The use of Eq. (13), even when obtained from a single

experimental point, as described in Eq. (14), is certainly much more

appropriate for sound evaluation of the experimental uncertain-

ties than the usual assumption of constant conversion variance, as

implicitly assumed when the procedure is used for the estimation

of model parameters and data analysis. It is worthwhile mention-

ing that Pereira et al. [17] conclude that their kinetic evaluation

depends on the assumed behavior of the experimental errors, thus

reinforcing the analysis performed here.

5. Conclusions

This paper assessed the effects of experimental errors on the

kinetic analysis of experimental data obtained during catalytic

tests. Particular attention was given to conditions where the con-

version variances were controlled by a single experimental design

variable. Simple and useful expressions were derived to express

how observed conversion variances depend on the experimental

conditions, assuming ﬁrst-order kinetics in a PFR.

It was shown that distinct optimum experimental designs for

precise parameter estimation are obtained when distinct assump-

tions are made for the behavior of the experimental errors.

This clearly indicates that proper characterization of experimen-

tal uncertainties may be of fundamental importance for rigorous

statistical analysis of kinetic data. Particularly, it was shown

that reaction temperatures, catalyst masses and spatial velocities

should be made equal to the maximum allowed values. Besides,

it was shown that conversion values should almost always lie in

the range (0.50, 0.80), indicating that kinetic studies should almost

always be performedthroughintegral methods. Despite that, it was

also shown that differential methods can be used when the errors

are controlled by the quality of the catalyst packing or the oscilla-

tions of feed ﬂow rates. Finally, it is important to emphasize that

the use of integral methods is not restricted to experimental con-

ditions that lead to small conversions and large relative conversion

variances, as it happens when differential methods are used.

Acknowledgment

The authors thank CNPq (Conselho Nacional de Desenvolvi-

mento Cientíﬁco e Tecnológico) for providing scholarships and for

supporting this work.

References

[1] M. Boudart, Kinetics of Chemical Processes, Prentice-Hall, Englewood Cliffs, NJ,

1968.

[2] G.F. Froment, K.B. Bischoff, Chemical Reactor Analysis and Design, Wiley, New

York, 1979.

[3] H.S. Fogler, Elements of Chemical ReactionEngineering, Prentice-Hall, NJ, 1999.

[4] Z.R. Lazic, Design of Experiments in Chemical Engineering: A Practical Guide,

Wiley–VCH, Weinheim, 2004.

[5] Y. Bard, Nonlinear Parameter Estimation, Academic Press, New York, 1974.

[6] M. Schwaab, J.C. Pinto, Análise de Dados Experimentais I: Fundamentos de

Estatística e Estimac¸ ão de Parâmetros, E-Papers, Rio de Janeiro, 2007.

[7] N.M. Bertero, C.R. Apesteguía, A.J. Marchi, Catalytic and kinetic study of the

liquid-phase hydrogenationof acetophenone over Cu/SiO

2

catalyst, Appl. Catal.

A 349 (2008) 100–109.

[8] M.I. Cabrera, R.J. Grau, Methyl oleate isomerization and hydrogenation over

Ni/␣-Al

2

O

3

: a kinetic study recognizing differences in the molecular size of

hydrogen and organic species, J. Mol. Catal. A: Chem. 287 (2008) 24–32.

[9] N.I. Dowling, P.D. Clark, Kinetic modeling of the reaction between hydrogen

and sulfur and opposing H

2

S decomposition at high temperatures, Ind. Eng.

Chem. Res. 38 (1999) 1369–1375.

[10] V. Vanrysselberghe, G.F. Froment, Kinetic modeling of hydrodesulfurization of

oil fractions: light cycle oil, Ind. Eng. Chem. Res. 37 (1998) 4231–4240.

[11] M. Daneshpayeh, A. Khodadadi, N. Mostouﬁ, Y. Mortazavi, R. Sotudeh-

Gharebagh, A. Talebizadeh, Kinetic modeling of oxidative coupling of methane

over Mn/Na

2

WO

4

/SiO

2

catalyst, Fuel Process. Technol. 90 (2009) 403–410.

[12] M. Joshib, A. Kremlinga, A. Seidel-Morgenstern, Model basedstatistical analysis

of adsorption equilibrium data, Chem. Eng. Sci. 61 (2006) 7805–7818.

[13] M. Schwaab, F.M. Silva, C.A. Queipo, A.G. Barreto Jr., M. Nele, J.C. Pinto, A new

approach for sequential experimental design for model discrimination, Chem.

Eng. Sci. 61 (2006) 5791–5806.

[14] M. Schwaab, J.L. Monteiro, J.C. Pinto, Sequential experimental design for model

discrimination taking into account the posterior covariance matrix of differ-

ences between model predictions, Chem. Eng. Sci. 63 (2008) 2408–2419.

[15] P.D.H. Hill, A review of experimental design procedures for regression model

discriminating, Technometrics 20 (1978) 15–21.

[16] G. Franceschini, S. Macchietto, Model-based design of experiments for param-

eter precision: state of the art, Chem. Eng. Sci. 63 (2008) 4846–4872.

[17] J.A.M. Pereira, M. Schwaab, E. Dell’Oro, J.C. Pinto, J.L.F. Monteiro, C.A. Henriques,

The kinetics of gibbsite dissolution in NaOH, Hydrometallurgy 96 (2009) 6–13.

[18] A.L. Larentis, A.M.P. Bentes, N.S. Resende, V.M.M. Salim, J.C. Pinto, Analysis of

experimental errors in catalytic tests for production of synthesis gas, Appl.

Catal. A 242 (2003) 365–379.

[19] O.A. Gutierrez, U.H. Danielson, Sensitivity analysis and error structure of

progress curves, Anal. Biochem. 358 (2006) 1–10.

[20] G.E.P. Box, J.S. Hunter, W.G. Hunter, Statistics for Experimenters: Design, Inno-

vation, and Discovery, 2nd edition, John Wiley & Sons, New Jersey, 2005.

[21] O.A. Gutierrez, M. Chavez, E. Lissi, A theoretical approach to some analyt-

ical properties of heterogeneous enzymatic assays, Anal. Chem. 76 (2004)

2664–2668.

(1).A. However. deﬁned as [5. As a matter of fact. Therefore. it was shown that the main sources of error were probably associated with the catalyst activity. In this case. however. Gutierrez and Danielson [19] analyzed the error structure of dynamic enzymatic assays when the initial enzyme and substrate concentrations are subject to uncertainties. The technical literature has largely overlooked this issue. as already said the correct characterization of experimental errors usually demands a large number of replicates. due to time and cost constraints. the experimental error is completely ignored in kinetic problems. 2 yi In order to obtain precise parameter values. which can be used for the design of the experimental conditions. As a consequence. Pereira et al. The authors ﬁtted an empirical model to available data to describe the conversion variance as a function of the conversion. the distance between model responses and experimental data tends to be smaller when the experimental variances are small. A more fundamental procedure is based on the assumption that the experimental errors follow the normal distribution. X may depend on speciﬁc features of the analyzed catalyst system. Most kinetic studies do not take into account the variance of the measured values and. / Chemical Engineering Journal 155 (2009) 816–823 817 Nomenclature B bij C S T sensitivity matrix ij-element of sensitivity matrix constant value deﬁned in Eq. due to existence of time and cost constraints. yi represents the calculated values of process responses. (12) objective function temperature volumetric ﬂow rate covariance matrix of parameter estimates experimental covariance matrix catalyst mass conversion inlet variables outlet or response variables packing quality of the catalyst bed model parameter variance of analytical measurement of variance variance of catalyst mass variance of conversion variance of outlet variable variance of parameter estimate variance of temperature variance of volumetric ﬂow rate v V Vy W X x y ˛ 2 XA 2 W 2 X 2 y 2 Â 2 T 2 which in turn are transferred to precision of output model predictions. yi represents c the values of experimental values of process responses. yi the determination of the probability distribution of a particular measurement can be obtained only through extensive replications.5]: V = (BT · V−1 · B) y −1 (2) v where Vy is the covariance matrix of experimental process response measurements and B is the sensitivity matrix. this is beyond the scope of the present manuscript. This may be very important because the precision of experimental measurements is transferred to the precision of kinetic parameters. Larentis et al. Usually.5. However. For this reason. rA = − ·PA ) where the output process response is the conversion (X) of reactant A. (Rigorously. In some cases. Besides.6]: S= i e (yi c − yi (x.L. statistical tests cannot be applied properly to evaluate the model adequacy and the quality of model parameters and of the model ﬁt [4]. experimental points that are subject to large experimental uncertainties do not inﬂuence the objective function signiﬁcantly. Alberton et al. which can be affected by the catalyst itself. most times this is not feasible. a speciﬁed metric of the matrix V must be minimized. (1) is considered equal or an arbitrary fraction of the measured variable [11. in this case. the preparation of the catalyst bed or the operational conditions of the reactor. It seems clear that experimental errors are not necessarily constant throughout the experimental range. which may involve the precision of reactant feed rates. it can be shown that the parameter values can be obtained through the minimization of the likelihood function. They showed that the variance of the substrate conversion presents a heteroscedastic i . The proposed analysis can be extended to non-normal problems with the help of more involving numerical schemes [5.6]. which may be many times unfeasible. deﬁned as [4. as only the mean value and the variance are needed. that the conversion variances depended on the measured conversion values. The normal distribution is often assumed because it allows for close representation of many measurement distributions and because the mathematical treatment of the data is simple. estimation of model parameters and model building are carried out with the help of the well-known least squares function [7–10]. The parameter uncertainties can be represented in terms of the covariance matrix of parameter uncertainties (V ). the precision of parameter after NE experiments can be obtained as: 1 2 Â NE = i=1 (∂Xi /∂Â) 2 Xi 2 (4) )) 2 (1) e where the index i indicates the experimental point. The proposed error analysis allowed for the identiﬁcation of the predominant reactions that were occurring in different experimental ranges. [17] studied the gibbsite dissolution in the presence of NaOH and concluded (based on an extensive experimental plan). techniques that allow for the simultaneous evaluation of experimental errors and minimization of the experimental efforts are welcome [4. For a ﬁrst-order reaction (A → B. the precision and accuracy of off-line laboratory analyses. (4) in order to characterize the precision of the kinetic parameter and of the calculated model responses. the model is perfect and experiments are well done. the control of reactor conditions. The underlying hypothesis that justiﬁes the use of the least squares function is that the experimental error is the same for all measurements. the correct characterization of experimental error is a fundamental issue during estimation of model parameters and analysis of experimental data.) In Eq. deﬁned as: B = [bij ] = ∂yj ∂Âi (3) values.13–16]. For one to use Eq. the variance in Eq. Therefore. calculated as a function of inlet design variables (x) and model parameters ( ). For example. So. it is necessary ﬁrst to characterize the experimental error of experi2 mental data. The resistance that many offer for proper characterization of experimental errors is understandable. [18] investigated the sources of error in reaction systems involving gaseous species and on-line chromatographic analyses. etc. each difference between experimental and model response values is weighted by the respective experimental error.12]. and 2 is the variance of process response measurements.

the expected conversion is XD . optimum parameter estimation may be performed either with differential or integral methods. Third. then Eq. depending on the behavior of the conversion variances. it may be concluded that the proper characterization of experimental errors can be extremely useful for those who are working in the ﬁeld of chemical kinetics and catalysis. (6). (12) presents interesting characteristics. leading to a different conversion (X). Expressing the conversion as a function of the design variables in terms of a Taylor series truncated after the ﬁrst term: X ≈ XD + ∂X ∂T ∂X ∂W (T − TD ) + T =TD or X ≈ ∂X ∂T ∂X ∂W T+ T =TD ∂X ∂v v v=vD + W W =WD (6) Squaring and averaging both sides of Eq. even when the number of replicates is small. ﬁrst-order kinetics is used frequently as a ﬁrst model building stage. In spite of the previous remarks. (8) is frequently used in statistics to perform robust processes or product design [20]. At the desired reaction temperature (TD ). the conversion can be written as: X = 1 − exp −Â0 · e−(Â1 /T ) Therefore. / Chemical Engineering Journal 155 (2009) 816–823 behavior in these cases. as the conversion.L. as shown experimentally and explored theoretically. Besides. which could be described with the help of a hybrid empiricalphenomenological model.63. It is shown here that the conversion variances can be described as functions of the measured conversion values. W ∂X = −exp −Â0 · e−(Â1 /T ) · v ∂T 1−X W v W Â1 v T2 W 1 (9) Â0 · e−(Â1 /T ) · ln(1−X) (10a) W ∂X = −exp −Â0 · e−(Â1 /T ) · v ∂v (1−X) −Â0 · e−(Â1 /T ) · ln(1−X) v v W 1 v W (10b) ∂X W = −exp −Â0 · e−(Â1 /T ) · v ∂W (1−X) −Â0 · e−(Â1 /T ) · ln(1−X) (10c) Inserting Eqs. it is important to say that the analysis of more complex reaction systems can also be performed with the help of the proposed differential analysis. ﬁrst-order kinetics may be regarded as a local linearization of more involving kinetic mechanisms. 1A. The main pursued objective is the development of a theoretical framework that may allow for simultaneous characterization of variances and minimization of the number of replicate runs.818 A. the main objective of the present paper is showing that experimental errors may depend quite considerably on the operation conditions. for an arbitrary C value. presenting a point of maximum conversion variance in the conversion range of 0.6 < X < 1. (11) can be written as: 2 X 2 X ∂X ∂v (v − vD ) v=vD = C · [(1 − X) ln(1 − X)]2 = C · [(1 − X) ln(1 − X)]2 + 2 XA (12a) (12b) + (W − WD ) W =WD (5) Eq. so that the real temperature (T). However. which has been largely overlooked in the ﬁeld of kinetics and catalysis. model equations are proposed in the present work to represent how experimental variances depend on reaction conditions in typical catalyst tests. as shown in the present manuscript. complex kinetic schemes are not necessary. the conversion variance presents a maximum value at X equal to 0. In Fig. As presented in the following sections. in this hypothetical system. Eq. reaching a maximum value around 60–80%.0 when observed experimental ﬂuctuations are controlled by the ﬂuctuations of the input variables. Initially. In this case. It is also assumed that diffusion constraints are absent. as shown in Fig. volumetric ﬂow rate (v) and mass of catalyst (W) are different from the desired design values. constant conversion variances should be expected only when ﬂuctuations are controlled by errors in the analytical measurement of the outlet variables. These constraints should not be overemphasized because catalyst tests are normally performed isothermically in plug ﬂow reactors and because ﬁrst-order reaction rate expressions are frequently used for model building and kinetic analysis. the experimental design variables are subject to ﬂuctuations. (11a) should be written as: 2 X = [(1 − X) ln(1 − X)]2 · 2 Â1 T4 2 T + v 2 v2 + W2 2 W + 2 XA (11b) Now it is necessary to evaluate the terms involving the design variables and its variances. For these reasons. the terms inside the brackets may be regarded as constant and represented by C. In this manner. For a ﬁrst-order reaction. Conversion variance for a ﬁrst-order reaction Let us assume that a catalytic system is designed to operate isothermically. Second. If the analyzed experimental range is not very large and/or the relative error measurements are constant. (8) and rearranging: 2 X = [(1 − X) ln(1 − X)]2 · 2 Â1 T4 2 T + v 2 v2 + 2 W W2 (11a) If the evaluation of X is also subject to a measurement error with 2 variance XA . As a consequence. more involving and rigorous statistical analyses of available data may become possible. observing that the 2 expected value of z2 is the variance z . 1. it can also be said that experimental errors are not necessarily constant in the experimental region for better understanding of the reaction system. It is assumed here that a ﬁrst-order reaction is carried out in a plug ﬂow reactor at isothermal conditions. where z is any variable): 2 2 X = ∂X ∂T T T =TD + ∂X ∂v v v=vD + ∂X ∂W W W =WD (7) Assuming that random ﬂuctuations of the design variables are not correlated to each other: 2 2 X 2 2 T = ∂X ∂T ∂X ∂W + 2 T =TD ∂X ∂v v v=vD 2 + W =WD 2 W (8) Eq. 2. in order to obtain the conversion variance (≈ is replaced by =. volumetric ﬂow rate (vD ) and catalyst bed mass (WD ). In order to illustrate this point. with inﬂection points at X equal . Alberton et al. Based on previous remarks. (10) in Eq.

1. This is equivalent to maximizing the right-hand side of Eq. one can observe that the relative conversion variance decreases almost linearly with conversion. meaning that uncertainties are controlled by the constant precision of the input variables. In this case. it is possible to determine the constant C: 2 X The most precise estimation of parameter can be achieved when the variance of the model parameter has a minimum value. (11). meaning that uncertainties are controlled by off-line laboratory analyses. as presented in Fig. the conversion variance divided by squared conversion values). as reaction rates at high conversion values become lower and perturbations of the operating conditions cause less signiﬁcant effects on the observed reaction rates. (12) can be very useful because it describes the conversion variance in terms of the measured conversion values. When conversion is equal to zero. when the experimental error is constant (and the analytical errors are predominant). it is possible to write: 1 2 Â = (∂X/∂Â) 2 X 2 (14) The partial derivative of conversion in respect to the kinetic constant is: ⎡ ⎤ ∂X W = −exp −Â · v ∂Â (1−X) ⎢ ⎥ ⎢−Â · W 1 ⎥ ⎣ v Â⎦ ln(1−X) (15) This can also be written in terms of the conversion. In order to do that. However. Therefore. Therefore. given the higher reaction rates observed at lower conversions. Obviously.L. 3. After this point of maximum. 3. the conversion variance also increases. 2 using Eq. (12). given the widespread usage of differential kinetic analysis in the ﬁeld of catalysis. if one assumes that conversion variances are constant and the assumption is not correct. (11).32 and 0. / Chemical Engineering Journal 155 (2009) 816–823 819 It must be pointed that the parameter C will not be constant if the experimental region of design variables is large and variances of the design variables are not proportional to measured values. as: Fig. Optimum estimation of the kinetic parameter Temperature effects are neglected in the following paragraphs. This may be regarded as a very important conclusion. the characterization of the variance in a single point may lead to inaccurate representation of conversion variances in the whole experimental range. meaning that uncertainties are controlled by the varying precision of the input variables. it may be concluded that. the well-known Arrhenius equation is not used to represent the temperature dependence of the kinetic rate constant. if replicates are available at a certain experimental condition. This clearly shows that Eq. Based on this result. Therefore. (16) is equivalent to maximizing the following equation: max 1 2 Â = max (1 − X)2 ln2 (1 − X) (17) =( 2 X0 − 2 XA ) · [(1 − X) ln(1 − X)]2 [(1 − X0 ) ln(1 − X0 )]2 + 2 XA (13) 2 where X0 represents the available conversion variance and X0 represents the average conversion obtained from available replicates. and (iii) conversion variance described by Eq. (12). Variance of the conversion (A) and relative conversion variance (B) as a function of conversion for Eq. Eq. it is preferable to use integral methods for model building and precise and accurate parameter estimation. this is the point where the conversion variance is maximum (when it is assumed that the conversion variance is not constant). 1. Therefore. (ii) conversion variance described by Eq. This constitutes a very important result.A.63. the continuous increase of conversion leads to appearance of a point of maximum conversion variance. If the relative conversion variances are considered (that is. (13) can constitute a useful approximation of the real variance behavior. it seems much more reasonable to use Eq. ((12)–(13)). (11). Alberton et al. (12) may ﬁnd application in distinct experimental systems. As the conversion increases. 1B. one must consider that Eq. However. the conversion variance is constant when uncertainties are controlled by off-line laboratory analyses. Constant conversion variance Based on Eq. according to Eq. because this means that the experimental optimum design (when the conversion variance is constant) indicates that the optimal experiment is placed at the region where the conversion variance attains its highest value (when the conversion variance is not constant). the variance of conversion decreases. It must be noted that Gutierrez and Danielson [19. Nevertheless.21] developed an equation for the variance of substrate conversion in dynamic enzymatic assays that is similar to Eq. Â2 2 Â = (1 − X)2 ln2 (1 − X) 2 X (16) to 0. This assumption will be relaxed in the following sections. three cases are considered below: (i) constant conversion variance (least squares function). experimental design . (13) for more fundamental data analysis than to ignore the effect of ﬂuctuations of the design variables on the conversion variances.93. maximizing Eq. conversion variance is also equal to zero because the reactant concentration does not change and the reactor output is not subject to process ﬂuctuations. which has a maximum placed at the value for X equal to 0. (4) to obtain 1/ Â . since variations of the reaction conditions perturb the measured conversion values and the inﬂuence of process ﬂuctuations is larger when reaction rates are high and the magnitude of conversion values increases. In this case. (16).

(16) (or one can simplify Eq. (23) indicates that is estimated with more precision when the conversion is lower. (21) into Eq.2. In order to determine the optimal experimental run for precise parameter estimation. Conversion variance described by Eq. from Eq. If the variances of input variables are constant. this may be regarded as a very important result.3. it may be concluded once more that it is preferable to use integral methods for model building and precise parameter estimation. Considering that only the deviations in the temperature are important. In this case. 3. the use of differential methods in the kinetic analysis may be considered. which is obtained as a function of the 2 observed conversion). this means that the optimum experimental conditions are the ones that lead to conversion values around 0–63%. (19) by assuming that the effect of the ﬂow rate is predominant) resulting in: Â2 2 Â 3. (19) presents a monotonic behavior with respect to volumetric ﬂow rate and mass of catalyst. the conversion variance can be evaluated by the expression [(1 − X)2 ln4 (1 − X)]. since at high volumetric ﬂow the conversion is usually low and the reaction rate can be regarded as constant along the reactor length. when the reaction follows the ﬁrst-order kinetics. Eq. As observed in the previous case. the righthand side of Eq. when the variance of the feed ﬂow rate controls the precision of output responses.1. 2 X · W as a function of conversion for Eq. then Eq. The temperature effect is predominant This may be the case of highly endo/exothermic reactions or systems where the temperature control is difﬁcult. which is very surprising! However. given the widespread usage of differential kinetic analysis in the ﬁeld of catalysis. 2 as a function of X. 3. then the parameter variance does not depend explicitly on the experimental conditions (although it depends on the obtained conversion.820 A. for low and high values of conversion. meaning that uncertainties are controlled by the varying precision of the input variables. the parameter variance can be written as: Â2 2 Â = v2 v 2 (23) = 1 Â1 2 T4 2 T + v2 v 2 + 2 W W2 + (1−X) ln (1−X) 2 XA 2 2 (19) It is necessary to perform the maximization of Eq. In order to show the main trends. the behavior of the variance conversion is similar to the one illustrated in Fig. if XA is considered. (22) presents a maximum at X equal to 0. it is preferable to evaluate the variance behavior when the ﬂuctuations are controlled by a single input variable. this clearly shows that the kinetic tests performed through differential methods may not be adequate for kinetic analysis. (11). Again. The ﬂow rate effect is predominant Considering that the effects of the temperature and catalyst mass in the conversion variance are negligible and only the effect of the ﬂow rate is important. (12) must be maximized in order to minimize the parameter variance. as it indicates that the correct evaluation of the experimental error changes completely the determination of the best region for precise parameter estimation. as the parameter variance approaches its maximum value when conversion approaches zero! Therefore. (11).L. the conversion variance can be written as: 2 X = [(1 − X) ln(1 − X)]2 2 Â1 T4 2 T (24) = [(1 − X) ln(1 − X)]2 · v 2 v2 + 2 XA (20) The conversion variance cannot be expressed as a function of conversion only. v=− Â·W ln(1 − X) (21) Inserting Eq. (11). 3. / Chemical Engineering Journal 155 (2009) 816–823 will erroneously recommend the experimental region where measurement errors attain the largest values. the conversion variance deﬁned in Eq. If reactor temperature is assumed to be constant. Considering the variance of the volumetric ﬂow constant. because it is a function of . (22). the error of parameter estimate can be determined as: Â2 2 Â = (1 − X)2 ln2 (1 − X) (1 − X)2 ln2 (1 − X) · C + 2 XA = 1 C +( 2 /(1 − X)2 XA ln2 (1 − X)) (18) Fig. Again. (19) with respect to temperature is more complex because the model parameter certainly depends on the reaction temperature. (11) If the conversion variance is described by Eq.3. as it also depends on the ﬂow rate and mass of catalyst. this may also be regarded as a very important conclusion. (19) in order to obtain the best experimental conditions for estimating the model parameter. The volumetric ﬂow rate can be expressed in terms of conversion as: 2 If the analytical conversion variance XA is neglected. The dependence of Eq. (12) If the conversion variance is described by Eq. (22) must be inserted into Eq. Therefore. Conversion variance described by Eq. (20): 2 X · W2 = v 2 Â2 · (1 − X)2 ln4 (1 − X) + 2 XA · W2 (22) If the analytical error can be neglected. presented in Fig. the conversion variance can be written as: 2 X If the variance of the volumetric ﬂow rate is constant. Once more. from Eq. 2. 1. the conversion variance is higher for lower tempera- . Again. meaning that uncertainties are controlled by the constant precision of the input variables. This clearly illustrates the importance of taking into account the proper interpretation of experimental errors in order to avoid very bad data analysis and experimental design.3. (12). the conversion variance tends to be lower than for intermediate values of conversion.86. Alberton et al.2. the parameter is estimated with more precision when the volumetric ﬂow is maximum. the right-hand side of Eq.

4. One must observe that when conversion variance was considered constant (Section 3. Despite that.1) the experimental condition that minimizes the parameter estimate variance was the same where conversion variance reaches a maximum value. For PFR reactors. The catalyst mass effect is predominant Considering now that the ﬂuctuations in the catalyst mass are predominant. the optimal parameter estimation can be achieved when the mass of catalyst is equal to the maximum allowed value. 4 [3]. and (B) for three spatial times and several temperatures. as shown in Fig. the designed experiments must be performed at experimental conditions where conversion variance has lower values. ˛ (which is quality of the packing of the catalyst bed) indicates the fraction of ﬂuid that short-cuts the catalyst bed. . Once more. (19) can be written as: Â2 2 Â The variance of the catalyst mass is probably constant and related to the precision of the balance used to weigh the mass. the parameter variance is smaller when the catalyst mass is larger. (25). (19) can be expressed as: 1 2 Â = 1 W2 2 Â2 W (27) Fig. experiments should be performed at the maximum allowable temperature values in order to minimize the variance of the kinetic parameter. In order to illustrate this effect. / Chemical Engineering Journal 155 (2009) 816–823 821 Fig. Therefore. instead of using the much simpler assumption of constant conversion variance during data analysis. Eq. 3. it can be concluded that experiments should be performed at the maximum allowable temperature values. As one might already expect. As conversion tends to increase when the mass of catalyst increases. (27). since at high conversion values the differential method is not applicable. this shows the importance of considering the real behavior of the conversion variance. The value of ˛ may change from bed to bed. the behavior of the conversion variance may be very complex and depends on the kinetic parameters and the spatial times. meaning that when the error is controlled by temperature effects.L.A. as clearly seen in Eq. This is very interesting. the weighing of low catalyst masses may introduce large experimental errors in the analysis. resulting in different contact times between the ﬂuid phase and the catalyst particles. 3. The packing effect It must be pointed out that packing of the catalyst bed does not depend directly on the design variables and should not be considered separately. According to Eq. in all cases intermediate conversion value leads to larger conversion variances.3. (24). The random packing of the catalyst bed may promote the occurrence of stagnant zones and short-cuts. 3. As it becomes clear that the precision of estimated parameters depends not only on the experimental precision. Variance of the conversion a function of conversion: (A) for three temperatures and several spatial times. When temperature is varied. (26) can be achieved at high temperatures. According to the proposed model. For a ﬁrst-order = 1 1 T4 Â2 (26) 2 T Maximization of Eq. In order to determine the optimal conditions for precise parameter estimation. ˛ values can be estimated from Residence Time Distribution (RTD) experiments. but also on the selected experimental conditions. this effect can be modeled in terms of two parallel reactors: one containing the catalyst and the other formed by void space. the conversion value and the respective conversion variance were calculated for distinct temperatures and residence times. 3. since when the conversion variance is not assumed constant and the temperature effect in the experimental error is predominant. consequently leading to unavoidable ﬂuctuation of the conversion. Schematic representation of the real catalyst bed and the proposed bed model.3. let us assume that: Â = 100 · exp − 3000 T (25) Based on Eq. as illustrated in Fig. Eq. using a trace step input [3]. 4. 3. One must observe that high temperature conditions also lead to lower values of the conversion variance. The obtained results are presented in Fig. it can be said indirectly that the integral method is preferable when the conversion variance is controlled by ﬂuctuations of the catalyst mass. When the error is controlled by temperature effects. tures. Alberton et al.

The best region for precise parameter estimation can be derived as shown previously.86. (13) As performed previously. Despite that. So. where Xeq is the equilibrium conversion. 3. the maximum conversion variance would occur at X equal to 0. the parameter variance can be written as: Â2 2 Â = (1 − X)2 ln(1 − X)2 X 1−˛ 1 X 1−˛ 2 2 ˛ − 1− X 1−˛ (34) ln 1 − = exp(A + B · X + C · X 2 ) (35) For a given temperature. In this case. (13). Alberton et al. 7 presents the experimental data reported by Pereira et al. 4. One can readily observe that similar analysis and equations can be obtained for batch reactors. this may be regarded as a very important result. In this case. optimum determination of model parameters can be carried out with integral methods. The authors carried out replicates at different conditions and used an empirical expression to ﬁt the conversion variance as a function of the measured conversion. experimental data reported in the literature were taken as an example. then the scenario is not clear. conversion is limited to 1 − ˛. ˛ values must be close to zero. Fig. the maximum conversion variances are observed in the same experimental region. Comparative analysis of the different scenarios As shown in the previous paragraphs. as ˛ increases. It can be observed that the conversion where the conversion variance attains it maximum value is shifted towards smaller values. Fig. Potential of estimation as a function of the conversion for several void fractions of the reactor. the conversion variance can be expressed as a function of characteristic ﬂuctuations of ˛. 5. For instance.(12) can be used to represent conversion variances in reaction systems described by kinetic rate expressions that are more complex than a ﬁrst-order reaction. as predicted by the pro- . The empirical equation was: 2 X For good packing of the catalyst bed. When the conversion variance is constant (and is controlled by analytical techniques). 6. (34) leads to X equal to 0. the conversion can be written as: X = (1 − ˛)(1 − e−(Â·W/v(1−˛)) ) (28) conditions enable the use differential methods and may lead to more precise parameter estimation when errors are controlled by packing effects. Besides. In order to evaluate the applicability of the present analysis in a real problem. Fig.19]. As it is probable that the quality of the catalyst bed may control the precision of the catalyst tests frequently. / Chemical Engineering Journal 155 (2009) 816–823 Fig. the optimum experimental region for the estimation of model parameters and model building depends on the behavior of conversion variances. One must observe that maximum conversion variances occur at intermediate conversion values. Conversion variance as a function of conversion for several values of ˛ in Eq. 5 shows the conversion variance as a function of conversion for several values of ˛. B and C are empirical ﬁtting parameters. the conversion variance can be written as: 2 X = ∂X ∂˛ 2 2 ˛ (29) As ∂X W = −1 + exp −k0 e−(E/RT ) · v0 (1 − ˛) ∂˛ +exp −k0 e−(E/RT ) · Then ∂X X =1− 1− 1−˛ ∂˛ Therefore 2 X W v0 (1 − ˛) k0 e−(E/RT ) · W v0 (1 − ˛) (30) · 1 + ln 1 − X 1−˛ X 1−˛ 2 (31) = X X − 1− 1−˛ 1−˛ ln 1 − 2 ˛ (32) This equation presents a maximum at: X= 1 + e2 + ˛(1 − e2 ) e2 (33) Based on experimental data reported in the literature [17. as ˛ describes a short-cut fraction. It indicates that reaction conditions should be selected in order to give small conversion values. (32). [17] studied the dissolution of gibbsite in a solution of NaOH in a batch reactor.5. as the lowest parameter variances are obtained for conversions values from 0 to 63%.822 A. The conversion variance was described as a function of the X/Xeq ratio. it can be assumed that Eq. as shown in Fig. It also can be observed that these where A. [17] and a best ﬁt obtained for Eq. but are compensated by similar increase of conversion sensitivity to parameter estimation. reaction. When the conversion variances are controlled by the design variables. 6 for different values of ˛. Experimental validation of Eq. equal to 94% at the analyzed experimental conditions.L. Similar results can be obtained when the ﬂuctuations induced by random variation of the design variables do not change considerably in the design region. or should be analyzed with integral methods if conversion variances are controlled by temperature of catalyst mass ﬂuctuations. the maximization of Eq. experiments can be analyzed through differential methods if conversion variances are controlled by ﬂow rate or packing ﬂuctuations. allowing the increase in the precision of parameter estimates. Pereira et al.

Resende. [20] G. Pinto. Particularly. Chem. Eng. [19] O. 1979. [18] A. References Fig.C.M. Kinetic modeling of oxidative coupling of methane over Mn/Na2 WO4 /SiO2 catalyst. Biochem. [21] O.S. assuming ﬁrst-order kinetics in a PFR. Catalytic and kinetic study of the liquid-phase hydrogenation of acetophenone over Cu/SiO2 catalyst. Hill. Catal. [14] M. Clark. Hunter. Eng. (13) to the available experimental data can be considered satisfactory. Sci. C. Chem. Ind. Appl. M. Silva. 63 (2008) 4846–4872. Res. SotudehGharebagh. J. 2007. Kremlinga. The use of Eq. Conclusions This paper assessed the effects of experimental errors on the kinetic analysis of experimental data obtained during catalytic tests. Wiley–VCH. E-Papers.A. Macchietto.G.J.M. Eq. Grau. [1] M.L. N. S. Alberton et al. is certainly much more appropriate for sound evaluation of the experimental uncertainties than the usual assumption of constant conversion variance. Bischoff. [11] M. Acknowledgment The authors thank CNPq (Conselho Nacional de Desenvolvimento Cientíﬁco e Tecnológico) for providing scholarships and for supporting this work.E. Rio de Janeiro. Joshib. Monteiro. A theoretical approach to some analytical properties of heterogeneous enzymatic assays. Sensitivity analysis and error structure of progress curves.. It was shown that distinct optimum experimental designs for precise parameter estimation are obtained when distinct assumptions are made for the behavior of the experimental errors. A. Sci.M. [5] Y. J.L. Appl. U. Prentice-Hall.A. [17]. ¸ [7] N. C. Catal. Pinto. [4] Z. Seidel-Morgenstern. Dowling. J. Model based statistical analysis of adsorption equilibrium data. Schwaab. even when obtained from a single experimental point. Model-based design of experiments for parameter precision: state of the art. Englewood Cliffs.J. J.F. Boudart. Ind. Catal. as it must be pointed out that the conﬁdence intervals of the true variances obtained from triplicates are very large. Even so. (13). New York. Apesteguía. J. Chavez. Y. Bertero.M. Innovation. Prentice-Hall. According to the authors. Besides. Anal. [13] M. Elements of Chemical Reaction Engineering. Schwaab. [12] M. Technometrics 20 (1978) 15–21. posed analysis. Simple and useful expressions were derived to express how observed conversion variances depend on the experimental conditions. Chem. P. J. Kinetic modeling of hydrodesulfurization of oil fractions: light cycle oil. A: Chem. A 349 (2008) 100–109. it is important to emphasize that the use of integral methods is not restricted to experimental conditions that lead to small conversions and large relative conversion variances. F. Sci. N. W.B. Pinto. The kinetics of gibbsite dissolution in NaOH. [9] N. [10] V. E. A. Larentis.80). it was also shown that differential methods can be used when the errors are controlled by the quality of the catalyst packing or the oscillations of feed ﬂow rates. This clearly indicates that proper characterization of experimental uncertainties may be of fundamental importance for rigorous statistical analysis of kinetic data.S. 37 (1998) 4231–4240. Eng. Technol. Hunter. 2004. [6] M.P.L. New York. it was shown that conversion values should almost always lie in . Nonlinear Parameter Estimation.M. Box. catalyst masses and spatial velocities should be made equal to the maximum allowed values.P. A. Eng. Methyl oleate isomerization and hydrogenation over Ni/ -Al2 O3 : a kinetic study recognizing differences in the molecular size of hydrogen and organic species. [3] H. Academic Press. Particular attention was given to conditions where the conversion variances were controlled by a single experimental design variable. Nele. 1968. Lissi. Anal. It is worthwhile mentioning that Pereira et al. A. Eng.M.A. Mol. Statistics for Experimenters: Design.C. 90 (2009) 403–410. Chem.H. Froment.G. Marchi. the adjustment of Eq. and Discovery. V. Design of Experiments in Chemical Engineering: A Practical Guide. 61 (2006) 7805–7818.H. Barreto Jr. John Wiley & Sons. G. 0. New Jersey.A. C. Chemical Reactor Analysis and Design. R. Chem. Dell’Oro. Pinto.D. 1974. Franceschini. [2] G. Froment. Fogler. Despite that. Bentes. thus reinforcing the analysis performed here. / Chemical Engineering Journal 155 (2009) 816–823 823 the range (0. NJ. as implicitly assumed when the procedure is used for the estimation of model parameters and data analysis.F. Gutierrez. 2005. Finally. Weinheim.R. Mortazavi.50. 61 (2006) 5791–5806. Pinto. [17] conclude that their kinetic evaluation depends on the assumed behavior of the experimental errors. Hydrometallurgy 96 (2009) 6–13. it was shown that reaction temperatures. Henriques.A. as it happens when differential methods are used. [8] M. Monteiro. J.F. A. indicating that kinetic studies should almost always be performed through integral methods. Analysis of experimental errors in catalytic tests for production of synthesis gas. Cabrera. which is not a simple ﬁrst-order reaction. 76 (2004) 2664–2668. E. Vanrysselberghe. Eng. (14). Chem. Schwaab. 358 (2006) 1–10. 5.S. Fuel Process. K. A. A new approach for sequential experimental design for model discrimination. Kinetic modeling of the reaction between hydrogen and sulfur and opposing H2 S decomposition at high temperatures. A. Res. [16] G.I. Daneshpayeh. 1999. J. 63 (2008) 2408–2419. Lazic. [15] P. Análise de Dados Experimentais I: Fundamentos de Estatística e Estimacão de Parâmetros. Sequential experimental design for model discrimination taking into account the posterior covariance matrix of differences between model predictions. A 242 (2003) 365–379. Chem.C.D. Sci. M. M.A.I. Khodadadi.C. Talebizadeh. Queipo. NJ. 287 (2008) 24–32. Danielson. as described in Eq. Schwaab. Bard. J. 2nd edition. 7. 38 (1999) 1369–1375. Pereira. (13) ﬁtted to the data reported by Pereira et al.C. [17] J.L. Mostouﬁ. Kinetics of Chemical Processes. the kinetics of gibbsite dissolution could be described by a reversible expression written in terms of the external area of gibbsite particles and hydroxide anion concentration. Salim.R. R. Gutierrez. Wiley. A review of experimental design procedures for regression model discriminating.

- Cellular Automata
- Conditional Modeling
- Breaking wave impact
- Tomography Rmp
- Level Set Methods - An Overview and Some Recent Results
- Lecture1.Ps
- Photon-Wavepackets as Flying Quantum Bits
- Computational Models for Turbulent Reacting Flows
- H. Lim et al- Chaos, Transport, and Mesh Convergence for Fluid Mixing
- Discrete Fourier transform
- TEMUCO
- Decoherence The Measurement Problem, And Interpretations of Quantum Mechanics 0312059
- Quantum Teleportation Report
- Space and time in a quantized world
- Klaus Colanero - Decoherence and Definite Outcomes
- The Discrete and the Continuous
- Some Worlds of Quantum Theory, by Jeremy Butterfield (WWW.OLOSCIENCE.COM)
- ForewordHiley_PrefaceLicata
- Quantum Propensities
- Quantum Parallelism and Scientific Realism
- Ichikawa Dylan EE266 FDTD
- Computational Fluid Dynamics for Naval Engineering Problems
- GENIUS MAKER - What is New in Genius Maker Software 3
- Contact Problem-Penalty method
- esr spectra
- 02 Finite Differences Per Gadin
- comnm_02
- 02 Finite Differences
- Description

Are you sure?

This action might not be possible to undo. Are you sure you want to continue?

We've moved you to where you read on your other device.

Get the full title to continue

Get the full title to continue listening from where you left off, or restart the preview.

scribd