You are on page 1of 41

1

MODEL 4
Chromatographic Separations

Mathematical Aspects:

(1) The Laplace transformation.


(2) Analysis of transient response by the moment technique.

Chromatography
In the separation process known as chromatography, a gas or a liquid containing low
concentrations of a mixture of chemical species, passes through a system (e.g. a packed bed)
containing a solid phase that interacts with the species in the liquid to provide separation. The
mechanism employed is often reversible adsorption on the solid surface. Due to the different
affinities of the various species in the mixture to the solid surface (i.e. different quantitative
liquid/solid or gas/solid partitioning), some species will be retained longer on the solid phase
than others, leading to a later elution. The process is illustrated in Figure 1.

Figure 1. Chromatographic separation: at t=0, a slug containing a mixture A+B is injected into
the main flow and the concentrations of A and B are measured in the effluent as a function of time
(elution curves). If the two compounds have different affinities for the solid phase (for example,
one adsorbs more strongly than the other on the solid surface), elution will occur at different times
and the compounds will be separated. In the example shown, B elutes later than A.

The objective of this model is to analyze the evolution of the concentration profiles of a
chemical species in a packed bed. Since the chemical species to be separated are usually present
in dilute concentrations, transport of each species can be analyzed separately. The analysis will
be based on an equation representing the mole balance of a chemical species at a point in the
bed. To establish the governing equation, we will use as dependent variable the average species
concentration over an averaging volume that is small with respect to the size of the column but
2

that is sufficiently large to smooth out concentration variations at the pore scale. The averaging
volume is illustrated in Figure 2.

Figure 2. Example of an averaging volume in a porous medium.

Let cp be the point (pore-level) concentration of a species in the fluid. The volume-averaged
concentration is defined by

1
c=
Vf ∫ c pdV (1)
Vf

The use of this average concentration allows us to formulate the problem describing the transport
of the species without having to calculate the rapidly varying cp. The analysis will be performed
under the following assumptions and conditions:
(1) Average fluid motion occurs along the main direction of motion (z) only. This means that,
even though the velocity field is three dimensional at the pore level, the average velocity vector
has only a nonzero z component. The average velocity will be characterized by means of the
superficial velocity, v, calculated as

Q
v= (2)
AB
3

where Q is the volumetric flow rate of fluid moving through the bed in the z direction, and AB is
the cross-sectional area of the column (AB=πd2/4 for a cylindrical column of diameter d).
(2) The only dependence of the volume-averaged concentration on position is with respect to
z (i.e. this concentration is uniform over a cross section of the bed) so that c=c(t,z).
(3) The bed porosity, defined as (Figure 2)

V
ε= f (3)
V

and the specific solid surface area

As
av = (4)
V

are uniform throughout the column.


(4) The mass exchange between solid and fluid may be represented by a convective mass
transfer coefficient (km). Let Js be the flux of species from the fluid to the solid phase (moles of
species per unit time and per unit of fluid/solid surface area). We can state

J s = k m (c − c e ) (5)

where ce is the concentration of species in the fluid that is directly in contact with the solid
surface. This concentration will be assumed to be at equilibrium with the concentration of
species adsorbed on the solid surface, ĉs (moles of species sorbed on the solid surface per unit
area), and the equilibrium relation (adsorption isotherm) is linear, so that

c e = Kĉ s (6)

where K is a constant (sometimes called partition coefficient). The assumption of equilibrium at


the solid/fluid interface will be valid if the adsorption/desorption process is fast with respect to
the mass transfer process (i.e. mass transfer from the fluid to the solid is the limiting rate
process). Under these conditions, a mole balance of species expressed as a point equation can be
written as follows,

∂c ∂c ∂ 2c
ε +v =D − a vJs (7)
∂t ∂z ∂z 2

accumulation convection dispersion transport to solid

where D is the effective dispersivity of the species in the column. Each term in this equation
represents a component of the mole balance expressed as moles per unit time and per unit bed
volume. The dispersion term includes molecular diffusion plus dispersion, a mass transport
mechanism arising from the fluctuations in the fluid velocity at the pore level.
Substitution of equations (5) and (6) into equation (7) leads to
4

∂c ∂c ∂ 2c
ε +v =D − a v k m (c − Kĉ s ) (8)
∂t ∂z ∂z 2

This equation has two dependent variables (c and ĉs ). To complete the formulation of the
problem, we need a mole balance of species in the solid phase:

∂ĉ s
= k m (c − Kĉ s ) (9)
∂t

This mole balance assumes that the only mechanism that affects the adsorbed concentration of
species is mass transfer from the fluid.
Equations (8) and (9) must be solved subject to appropriate initial and boundary conditions. In
a typical chromatographic process, the bed is initially free of solute,

c=0, t=0 (10)

ĉs = 0 , t=0 (11)

and the inlet concentration is a prescribed function of time:

c=ci(t), z=0 (12)

A second boundary condition for c is needed. The second condition is not straight forward
sometimes and has been historically a matter of debate in the literature. Under certain conditions,
it can be assumed that changes in concentration with position at the exit of the bed are small and
state mathematically

∂c
= 0 , z=L (13)
∂z

where L is the length of packing. Furthermore, if the bed is long enough, condition (13) can be
applied letting L→∞, which somewhat simplifies the solution.
The PDEs (8) and (9) constitute a system of two linear, coupled, homogeneous differential
equations. Note that the only non homogeneous condition is equation (12). To solve this
problem, we will apply the Laplace transformation method. In what follows, we will introduce
the Laplace transform as a tool for the solution of differential equations.

The Laplace transform


The Laplace transform is a particular case of an integral transform, which is an operator that,
when applied to a function, generates a different function with a new independent variable. For a
function f(t), the general integral transform takes the form

b
f (s) = ∫ K (s, t )f ( t )dt (14)
a
5

where f (s) is the transform of f(t), a and b are integration limits in the domain in which f is
defined, and K(s,t) is the kernel of the transformation.
The functional form of the kernel and the values of a and b define a particular transformation.
One of the most useful transforms is the Laplace transform, in which the kernel is an exponential
function:


f (s) = L {f ( t )} = ∫ e − st f ( t )dt (15)
0

where the operation L {f(t)} means "Laplace transform of f(t)."


The Laplace transform is a useful tool for the solution of linear differential equations. As we
will see later in this chapter, the usefulness of the Laplace transform is that it converts ordinary
differential equations into algebraic expressions, and partial differential equations into ordinary
differential equations. A flow chart of the application of the Laplace transform is shown in
Figure 3. The success of the application of the Laplace transform to solve a specific problem is
assured only if the three steps in these procedures can be performed analytically. Exception to
this are methods that can be used to invert a Laplace domain solution numerically, but these are
not used very much in practice, and methods that do not seek to find the final solution but only
certain properties of it, such as the method of moments that will be analyzed later in this chapter.

Figure 3. Typical steps followed in the application of the Laplace transformation to the
solution of ODEs or PDEs.

The following example gives an introduction to the use of Laplace transform to solve ODEs.
6

Evolution of concentration in two CSTRs in series (isothermal, first-order reaction)


Consider a reaction that is carried out in two CSTRs in series (Figure 4) under isothermal
conditions. Here, c refers to the concentration of reactant, which is consumed by a first-order
reaction.

Figure 4. Two CSTRs in series with the same volume V.

Mole balances of reactant in both reactors yield the problem (see Chapter I)

dc1 1
= (c i − c1 ) − kc1 (16)
dt t R

dc 2 1
= (c1 − c 2 ) − kc 2 (17)
dt tR

c1=c2=c0, t=0 (18)

where tR=V/Q is the residence time in each reactor, and we have assumed that both reactors have
the same initial concentration of reactant, c0. To solve this problem, we will apply the Laplace
transform to each ODE. To do this, we multiply each equation by e-st and then integrate t from 0
to ∞. For example, applying this procedure to equation (16) yields
7

∞ ∞
dc1 1 
∫ e −st dt
dt = ∫ e − st  (c i − c1 ) − kc1 dt
tR 
(19)
0 0

The left-hand side of this equation can be resolved applying integration by parts, as follows,

∞ ∞
dc1 ∞
∫ e −st dt
dt = e − st c1 + s ∫ e − st c1dt
0
(20)
0 0

We restrict the values of the Laplace variable to s>0 and recognize that the last integral in this
equation is the transform of c1, and use the initial condition (18) to get


dc1
∫ e −st dt
dt = −c 0 + s c1 (21)
0

Substituting this equation into equation (19) and operating with the right-hand side yields

c  1 
s c1 − c 0 = i −  + k  c1 (22)
tR  tR 

Similarly, application of the Laplace transform to equation (17) leads to

c  1 
s c2 − c 0 = 1 −  + k  c2 (23)
tR  tR 

Since ci(t) is known (and, consequently, ci can be determined), equation (22) can be solved to
find c1 :

c0 1 ci
c1 = + (24)
s + α tR s + α

where

1
α= +k (25)
tR

Also, from equation (23) we find

c0 1 c1
c2 = + (26)
s + α tR s + α
8

The next step is to find the inverse transform of the functions in equations (24) and (26). To do
this, we need to know the inlet concentration as a function of time, ci(t). Let us take the simple
case in which ci is a constant for t≥0. We then have (using equation 15)


 1 ∞ c
ci (s) = ∫ e − st c i dt = c i  − e − st  = i (27)
0
 s 0  s

Equation (24) now can be written as

c0 c 1
c1 = + i (28)
s + α t R s(s + α)

Because of the linear nature of the Laplace operator, and using the notation

L −1 {f (s)} = f ( t ) (29)

equation (28) implies (taking the inverse)

 1  ci  1 
c1 ( t ) = c 0L −1  + L −1   (30)
s + α  t R  s(s + α ) 

The definition of the Laplace transform (equation 15) can be used to show that

L {e − αt } =
1
(31)
s+α

To find the second inverse transform in equation (30), we make use of partial fraction
decomposition. We seek to find constants A and B such that

1 A B
= + (32)
s(s + α) s s + α

which implies

1=A(S+α)+Bs (33)

Since this equation must be valid for any s, we can equate coefficients of terms having the same
power of s:

s0 : αA=1 ⇒ A=1/α (34)

s1 : A+B=0 ⇒ B = −1/α (35)


9

Hence

1 1 1 1 
=  −  (36)
s(s + α ) α  s s + α 

Therefore, using the linear nature of the Laplace transform, we get

 1  1  −1 1   1 
L −1   = L   − L −1   (37)
 s(s + α )  α  s   s + α 

From equation (31), this equation leads to

 1  1
L −1   = (1 − e − αt ) (38)
 s (s + α )  α

Using equations (31) and (38), the solution for c1 (equation 30) is

ci  c 
c1 ( t ) = +  c 0 − i e − αt (39)
tRα  tRα 

To find c2, we first substitute equation (28) into equation (26):

c0 c 1 c 1
c2 = + 0 + i (40)
s + α t R (s + α) 2 t 2R s(s + α ) 2

Using the definition of Laplace transform (equation 15), it is possible to show that

 1 
L −1  = te − αt
2
(41)
 (s + α ) 

which will take care of the second term in equation (40). As for the third term, we once again use
partial fraction decomposition:

1 C Ds + E
= + (42)
s(s + α) 2 s (s + α) 2

(This is not the only possible partial fraction decomposition, but all should lead eventually to the
same result.) Finding the coefficients in a manner similar to the procedure used before yields

1 1 1 1 s 2 1
= − − (43)
s(s + α ) 2 α 2 s α 2 (s + α ) 2 α (s + α ) 2
10

Equation (40) now can be written as follows

c0 c 1 c  1 1 1 s 2 1 
c2 = + 0 + i  − − (44)
s + α t R (s + α) 2 t R  α s α (s + α)
2 2 2 2 α (s + α) 2 

Using equations (31) and (41) (or directly equation 15), it is possible to show that

 s 
L −1   = (1 − αt )e − αt (45)
 (s + α ) 2

Now we can find the inverse transform of equation (44), using equations (31), (38), (41) and
(45). The result is, after manipulations,


c 2 ( t ) = c 0 1 +
t  − αt
e
c
[
+ i 1 − (1 + αt )e − αt ] (46)
 tR  t 2R α 2

This example shows how a system of two ODEs can be transformed to a system of two algebraic
equations in the Laplace domain, whose solution can be found analytically. A key aspect of the
procedure is the knowledge of the inverse Laplace transform of the solution in the Laplace
domain. In practice, tables of Laplace transforms available in the literature are used to find the
inverse transform. In addition, there are certain properties of the Laplace transformation that aid
in the solution of problems such as the one considered in this section. Some of these properties
are presented below.

Properties of the Laplace transformation


All the properties presented in this section can be proven from the definition of Laplace
transform (equation 15). Some of these properties were derived and/or used in the previous
section. In what follows, a and b are constants, and n is an integer, n>0.

(1) Linearity:

L {af ( t ) + bg ( t )} = af (s) + bg (s) (47)

(2) Transform of derivatives:

 df 
L   = sf (s) − f (0) (48)
 dt 

dnf  n
L  = s f (s) − [s n −1f (0) + s n − 2 f ' (0) + Lf ( n −1) (0)]
n
(49)
 dt 

(3) Transform of integrals:


11

 t  1 a
L ∫ f (ξ)dξ = f (s) − ∫ f (ξ)dξ (50)
a  s 0

(4) Shifting properties:

L {e − at f ( t )} = f (s + a ) (51)

L {H ( t − a )f ( t − a )} = e − as f (s) , a>0 (52)

where H is the step function (sometimes called Heaviside function), defined by

0, t < a
H( t − a ) =  (53)
1, t ≥ a

(5) Other properties:

L {t n f ( t )} = (−1) n
dnf
(54)
ds n

L {f (at )} = f (s / a ) , a>0
1
(55)
a

(6) Convolution theorem:


The definition of the convolution product of two functions, f(t) and g(t) is

t
f * g ( t ) = ∫ f ( t − τ ) g ( τ ) dτ (56)
0

The convolution theorem states that

L {f * g ( t )} = f (s) g (s) (57)

This theorem is particularly useful to obtain the inverse Laplace transform of the product of two
functions in the Laplace domain of two functions whose individual inverse transforms are
known. For example, consider the inverse transform of the function

1
h (s) = (58)
(s + α ) 3

Here, we can use


12

1
f (s) = (59)
s+α

and

1
g (s) = (60)
(s + α ) 2

From equations (31) and (41), we have

f ( t ) = e − αt (61)

g ( t ) = te −αt (62)

Applying the convolution theorem, we can state

t t
t 2 − αt
h ( t ) = ∫ e − α ( t − τ) τe − ατ dτ = e − αt ∫ τdτ = e (63)
2
0 0

Before going back to the model for chromatographic separations, the next section will
illustrate how to use the Laplace transform method to solve a PDE.

Velocity profiles induced by an accelerating flat plate


Consider a deep pool of liquid in contact with a large flat plate. The liquid is initially at rest
and then, at t=0, the plate starts moving with a velocity v0. This will generate a velocity profile in
the liquid that will evolve with time (Figure 5).

Figure 5. Velocity profile generated by a suddenly-accelerated flat plate. The profile will
penetrate more into the liquid as time passes.

For a Newtonian fluid in laminar, incompressible flow conditions, we can postulate that
13

vy=vz=0, and that vx=vx(t,y). Since the fluid motion is caused exclusively by the plate motion,
there will be no pressure gradients in the x direction, and the x component of the Navier-Stokes
equations simplifies to

∂v x ∂2vx
=ν (64)
∂t ∂y 2

where ν is the kinematic viscosity (ν=µ/ρ). The initial and boundary conditions are

vx=0, t=0 (65)

vx=v0, y=0 (66)

vx=0, y→∞ (67)

For the case of constant v0, a simple change of variables will make this problem identical as
the leaching from an infinite solid treated at the end Chapter II. The solution was found in that
case using similarity transformation. Here, we will find the solution by Laplace transform
considering the more general case v0=v0(t) which, in general, cannot be solved by similarity
transformation. We start by taking the Laplace transform of equation (64). For the left-hand side
of the equation, use of equations (48) and (65) yields

 ∂v 
L  x  = s v x − v x ( 0) = s v x (68)
 ∂t 

For the right-hand side of the equation, we note that the Laplace operator is interchangeable with
the derivative with respect to x, since the operator involves integration with respect to time. This
means that

∂ 2vx  d2 vx
L  = (69)
 ∂y 2  dy 2

The transform v x is a function of y and s but, since there are no derivatives with respect to s in
the formulation, we may consider s to be a parameter of the function v x ( y) . This is the reason
for the use of the total derivative in equation (69). Upon substitution of equations (68) and (69),
equation (64) becomes

d2vx s
− vx = 0 (70)
dy 2 ν

To find the boundary conditions for this ODE, we take the Laplace transform of the boundary
conditions (66) and (67) to get
14

v x = v 0 , y=0 (71)

v x = 0 , y→∞ (72)

The solution of equation (70) is

 s  s
v x = A exp − y  + B exp y
  ν
 (73)
 ν   

Application of condition (72) leads to B=0, and condition (71) yields A= v 0 and thus

 s
v x = v 0 exp − y 
 (74)
 ν 

Let us first consider the case of constant v0. In this case,

v
v0 = 0 (75)
s

and

v  s
v x = 0 exp − y  (76)
s  ν 

To find vx(t,y) we need the inverse transform of this expression. From Appendix D, Rice and Do,
transform number 61, we have

  α  e − α s
L erfc  = (77)
  2 t  s

where erfc is the complementary error function, defined by

erfc(x)=1−erf(x) (78)

We see that the inverse transform of equation (76) yields

 y 
v x ( t , y) = v 0 erfc  (79)
 2 νt 

which is the solution for the velocity for the case of constant v0. Note how it is possible to
express this solution in terms of a similarity variable of the form y / t .
15

Let us consider now the more general case in which v0 is a known function of time. The
solution of the problem in the Laplace domain is equation (74). To find the solution, we will use
the convolution theorem. From Rice and Do, Appendix D, transform number 60, we have

 α  α 2  − α s
L  exp −  = e (80)
 2 πt 3  4 t 

Application of the convolution theorem, equation (57), to equation (76) with

f = v0 (81)

and

 s
g = exp − y 
 (82)
 ν 

leads to

y t v 0 ( t − τ)  y2 
v x ( t , y) = ∫
2 νπ 0 τ3 / 2
exp − dτ
 4ντ 
(83)

It can be shown that, letting v0(t)=v0 (constant) in this equation, also leads to equation (79).
However, this equation gives the general solution for any v0(t). Note that, even if the integral
cannot be evaluated analytically, the velocity profile can be found from this equation by
numerical integration.
The convolution theorem is a very powerful tool in the solution of problems by Laplace
transform, especially those problems that involve the calculation of the transient response of a
system subjected to a temporal stimulus. In the problem solved here, the temporal stimulus is the
velocity of the plate. In the analysis of chromatography, the temporal stimulus is the inlet
concentration. In this case, the stimulus has a short duration, close to a pulse of solute in the inlet
flow. Before we go back to the analysis of chromatography, we will introduce the Dirac delta
function, used to characterize short-lived pulses of temporal variables.

The Dirac delta function


It is common to encounter processes in practice in which events occur at different time scales.
Sometimes an event occurs over a time that is much faster than the characteristic time over
which process variables change. The same type of behavior might be associated with length
scales. The Dirac delta function can be used to represent mathematically such events.
For instance, consider a CSTR in which a first-order reaction takes place (A→Products).
There is a continuous flow of solvent through the reactor and at t=0 a specified number of moles
of reactant, nA, are loaded into the reactor. If the loading process is fast (more on this below), we
can say that the reactant concentration in the reactor suffers a discontinuity at t=0, going from 0
to
16

n
c A 0 = A , t=0 (84)
V

where V is the volume of reactive mixture on the reactor. From this moment on, no more
reactant is added. A mole balance of A in the reactor yields (equation I-9 with cAi=0)

dc A  1 
+  k + c A = 0 , t>0 (85)
dt  tR 

Using the initial condition (84), the solution is

 t 
c A = c A 0 exp − (1 + kt R )  (86)
 tR 

Suppose that the process will be carried out until 90% of the added reactant is either consumed
or leaves the reactor; that is, until cA=0.90cA0. This means that the process will be carried out
until a time tf such that (equation 86)

 t 
0.1 = exp − (1 + kt R ) f  ⇒ t f =
2 .3 t R
(87)
 tR  1 + kt R

The loading of the reactor with reactant at the beginning of the process is not instantaneous in
practice. Let t0 be the time necessary to loan the nA moles of reactant at the beginning of the
process. We can see that the loading process can be considered instantaneous for all practical
purposes if t0«tf. That is, equation (86) will be valid only if this constraint is satisfied.
Consider now the case in which the process is cyclic: at time tf, a load of 0.9nA moles of
reactant is added to the reactor (over a period of time t0«tf) to compensate for the decay in
reactant concentration. This means that at time tf the concentration of A in the reactor will be
again cA0 and the process will start over. The analysis of this case is simple: we can "shift" the
time scale so that t=0 coincides with the beginning of each cycle, and equation (86) then will
represent the concentration evolution within each cycle. Alternatively, we can formulate the
problem so that we incorporate each load mathematically into the mole balance. This would be
more convenient to treat problems with a more complex loading pattern. One way to do this is to
add a term in the mole balance (85) to incorporate addition of A:

dc A  1 
+  k + c A = L( t ) (88)
dt  tR 

Each term in this balance represents moles of A per unit time and unit reactor volume so that L(t)
must represents the moles of A added at time t per unit time and reactor volume. Since the
addition is instantaneous, L(t) must be infinite at the loading time t=tf (and all subsequent cyclic
loading times) since a finite number of moles nA are added in a very short time interval (∆t→0).
That is
17

∞, t = t f
L( t ) =  (89)
0, t ≠ t f

But L(t) must reflect the fact that nA moles are added so that, over the loading interval ∆t→0, we
must have that

t f + ∆t / 2
nA
∫ L(t )dt = V
(90)
t f − ∆t / 2

A mathematical representation for L(t) is accomplished by means of the Dirac delta function,
δ(t), which has the properties (see more formal definition below)

∞, t = 0
δ( t ) =  (91)
0, t ≠ 0

and

∞ ∆t / 2
∫ δ(t )dt = ∫ δ( t )dt = 1 (92)
−∞ − ∆t / 2

Using this definition, we can incorporate instantaneous loads into the mole balance as follows.
Let nAn be the number of moles of A loaded into the reactor at time tn, with n=1,2,… If these
loads are instantaneous then we can say that the nth value of L(t) is

n
L n ( t ) = An δ( t − t n ) (93)
V

Note that this satisfies equation (90) due to property (92). The sum of all loads will give the total
addition of A, so that the mole balance, equation (88), can be expressed as

dc A  1  n
+  k + c A = ∑ An δ( t − t n ) (94)
dt  tR  all n
V

In summary, multiple additions of reactant can be treated two ways: (1) use the solution (96) to
represent concentration evolution between loads, and then shift the time axes and apply the
appropriate initial condition to solve for each period between loads, or (2) solve directly equation
(94). The latter treatment is more convenient when using Laplace transform to find a solution, as
will be illustrated in the next section.
Equations (91) and (92) represent properties of the Dirac delta function, but they are far from
being a mathematically rigorous definition. A formal definition follows from the use of
sequences of functions that are continuous and differentiable at each point. Consider the
following set of functions:
18

n 1
ψ n (t ) = , n=1,2,… (95)
π 1+ n2t 2

For a given n, this function is continuous and differentiable in −∞<t<∞. The coefficient n/π has
been chosen so that


∫ ψ n dt = 1 , n=1,2,… (96)
−∞

The graphical representation of this sequence of functions is shown in Figure 6. Each function is
symmetric around t=0 and, as n increases, the value of ψn(0) increases but the value of the
function decreases beyond a narrowing interval around t=0. As n→∞, this function satisfies the
properties in equations (91) and (92). We can define the Dirac delta function as

δ( t ) = lim ψ n ( t ) (97)
n →∞

This representation has important mathematical implications, since δ(t), despite its singular
behavior, can be considered in some cases to behave as a continuous, differentiable function.
If we accept that any function that satisfies the properties (91) and (92) can be considered a
Dirac delta function, then the representation given above is not unique. In fact, other functions
that can be used in the definition (97) are

n
ψ n (t ) = exp(− n 2 t 2 ) , n=1,2,… (98)
π

1 sin 2 nt
ψ n (t ) = , n=1,2,… (99)
nπ t 2

Other properties of the Dirac delta function follow. Consider the following integral
containing an arbitrary function f(t):

∞ a + ∆t / 2
∫ δ(t − a )f ( t)dt = ∫ δ( t − a )f (t )dt (100)
−∞ a − ∆t / 2

where ∆t is an arbitrary interval around t=a. We can transform the integration limits this way
because δ(t-a) will be zero everywhere except at t=a. Using the mean value theorem, the right-
hand side of this equation can be expressed as

∞ a + ∆t / 2
∫ δ(t − a )f ( t)dt =f (ξ) ∫ δ(t − a )dt (101)
−∞ a − ∆t / 2
19

where a−∆t/2≤ξ≤a+∆t/2. The integral on the right-hand side of this equation is unity. Letting
∆t→0 yields the following property for the delta function:


∫ δ(t − a )f ( t )dt =f (a ) (102)
−∞

valid for any function f(t).

ψn 1
0.9 3
0.8
0.7
0.6
2
0.5
0.4
0.3
n=1
0.2
0.1
0
-4 -2 0 2 4
t

Figure 6. The first three functions of the sequence given by equation (95).

Next, notice that

t
0, t < a
∫ δ(τ − a )dτ =1, t > a (103)
−∞

which is the definition of the step function (equation 53). We conclude that

t
∫ δ(τ − a )dτ =H( t − a ) (104)
−∞

or
20

dH( t − a )
= δ( t − a ) (105)
dt

Finally, the Laplace transform of the delta function is


L {δ( t − a )} = ∫ e − st δ( t − a )dt = e − as , a≥0 (106)
0

Note that, for a=0,

L {δ( t )} = 1 (107)

Now we can solve equation (94). Taking the Laplace transform of equation (94) and using
equation (106) leads to

 1  n
s cA +  k +  cA = ∑ An exp(−st n ) (108)
 tR  all n
V

which leads to

n An exp(−st n )
cA = ∑ V s + k + 1/ t R
(109)
all n

Taking the inverse transform of this equation and taking advantage of the linearity of the
operation yields

n An −1  exp(−st n ) 
cA = ∑ V
L  
 s + k + 1/ t R 
(110)
all n

Consider the property (52). We can see that

L −1 {exp(−st n )f (s)} = f ( t − t n )H( t − t n ) (111)

In this case (equation 110), we have

1
f (s) = (112)
s + k + 1/ t R

which means that (see equation 31)

f ( t ) = exp[− (k + 1 / t R )t ] (113)
21

Equation (110) then becomes

cA = ∑
n An
V
[
H( t − t n ) exp − (k + 1 / t R )( t − t n ) ] (114)
all n

This is the solution of equation (94).

Passive dispersion in packed beds


In this section we consider a simplified version of the problem formulated as Model 4. Here,
we will deal with the transport of a chemical species in a packed bed, considering that the species
does not interact with the solid phase. The transport equation in this case is equation (7) with
Js=0:

∂c ∂c ∂ 2c
ε +v =D (115)
∂t ∂z ∂z 2

with initial and boundary conditions:

c=0, t=0 (116)

c=ci(t), z=0 (117)

∂c
= 0 , z→∞ (118)
∂z

In this problem, we will describe how a chemical species fed into the packed bed with an inlet
concentration ci(t) distributes in time in space due to dispersion and convection. As we will see,
dispersion will play a major role in the concentration profiles and, because of the fact that the
species does not react or interact with the solid phase, the process is called passive dispersion.
To solve the problem, we apply Laplace transform to equation (115) to get

dc d2c
ε(s c − c 0 ) + v =D (119)
dz dz 2

Since c0=0 (equation 116), this equation can be written as

d 2 c v d c εs
− − c=0 (120)
dz 2 D dz D

This is a second-order ODE with constant (with respect to z) coefficients. The characteristic
equation is

v εs
r2 − r− =0 (121)
D D
22

The roots of this equation are:

2
v 1 v 4ε s
r= ±   + (122)
2D 2  d  D

Note that one root is positive (r+) and the other negative (r-). The solution of equation (120) is

c = Ae zr+ + Be zr− (123)

We note that, to keep c finite as z→∞, we need to have A=0. This replaces boundary condition
(118). Also, taking the Laplace transform of boundary condition (117) yields

c = ci , z=0 (124)

Applying this boundary condition leads to the solution for the concentration in the Laplace
domain:

 v  4εsD  
c = ci exp  1 − 1 + z  (125)
 v 2  
 2D 

Once again, the feasibility of finding the final solution depends on whether the inverse of this
transform can be obtained analytically and this, in turn, depends on ci (s) . As a first case,
consider that a certain amount of the transported species is injected into the carrier fluid at t=0
(Figure 1). A number of moles M will be injected uniformly during a short time interval ∆t. A
mole balance of species at the injection point, considering that the volumetric flow rate of the
carrier fluid, Q, is not changed appreciably by the injection process, yields

∆t / 2
M= ∫ Qci dt (126)
− ∆t / 2

If ∆t→0, then ci may be expressed as proportional to δ(t) and to satisfy this last equation we see
that

M
ci = δ( t ) (127)
Q

The Laplace transform of the inlet concentration is then

M
ci (s) = (128)
Q

Equation (125) then becomes


23

M  v  4εsD  
c= exp  1 − 1 + z  (129)
 2 
Q  2 D  v  

To find the inverse Laplace transform of this expression, we will use equations (51) and (80).
First, we rearrange equation (129) as follows

M  vz   ε v 2 
c= exp  exp − z s+ (130)
Q  2D   D 4εD 

Now, let

v2
a= (131)
4εD

ε
α=z (132)
D

Equation (130) can be expressed as

c=
M  vz 
exp (
 exp − α s + a ) (133)
Q  2D 

Now, if we let

(
f (s + a ) = exp − α s + a ) (134)

We see, from equations (51) and (80) that

M  vz  − at α
c= exp e e −α 2 / 4t (135)
Q  2 D  2 πt 3

Substituting equations (131) and (132) and rearranging yields the solution

Mz ε  ε  vt  
2
c( t , z ) = exp −
 z −   (136)
2Q πDt 3  4Dt  ε 

Figures 7 and 8 show plots of this response. At a specific axial position (Figure 8), the response
has a maximum at a time that increases with z, which reflects the delay in the response due to the
fact that the fluid moves downstream at a certain velocity (i.e., it reflects the residence time of
the fluid in the bed before z). Also, as z increases, the spread of the curves increases. The spread
is due to dispersion (i.e., the fact that individual fluid elements move faster or slower than the
24

average velocity, see below). It is interesting to note that two different species could not be
separated in this process, since they would both move with the fluid's average velocity.

c
t1

t2

t3

0
0 z

Figure 7. Qualitative representation of concentration profiles from equation (136) at various


times. As time passes (t3>t2>t1), the "tail" of the profiles grows longer due to dispersion.

z=0
c z1

z2

z3

0
t

Figure 8. Qualitative representation of concentration as a function of time from equation (136) at


various axial positions (z3>z2>z1).
25

From a mathematical point of view, we can express the previous solution in terms of the
response to a pulse input by using

ci(t)=δ(t) (137)

(mathematically, this is equivalent to setting M/Q=1). The concentration that corresponds to this
input will be termed cδ. From equation (129) we have

 vz   ε v 2 
cδ = exp  exp − z s+ (138)
 2D   D 4εD 

and the inverse is

z ε  ε  vt  
2
c δ ( t, z) = exp −  z−   (139)
2 πDt 3  4Dt  ε  

Comparing equations (125) and (138), we can see that any response in the Laplace domain can
be expressed as

c = ci cδ (140)

For an arbitrary ci , the inverse of this function can be expressed using the convolution theorem
as follows,

t
c( t , z) = ∫ c i ( t − τ)c δ (τ, z)dτ (141)
0

This means that the solution corresponding to any input can be found after an integration step,
once the solution to a delta function input is known (equation 139).
An interesting particular case of equation (141) occurs when the input is a step function:

ci(t)=H(t) (142)

Let cH be the solution for this particular case. From equation (141), we have

t
c( t , z) = ∫ H ( t − τ)c δ (τ, z)dτ (143)
0

But H(t−τ)=1 for τ≤t. Therefore, we find


26

t
c H ( t , z) = ∫ c δ (τ, z)dτ (144)
0

or

∂c H
= cδ (145)
∂t

which means that the responses to step and pulse inputs are related in the same way as the input
functions (equation 105).

Passive transport in packed beds: plug flow limit


If dispersion is negligible in a packed bed or porous medium, the transport of a chemical
species is described by the following simplified form of equation (115)

∂c ∂c
ε +v =0 (146)
∂t ∂z

with initial and boundary conditions:

c=0, t=0 (147)

c=ci(t), z=0 (148)

Note that now a single boundary condition is enough to formulate the problem, since the
transport equation only contains first derivatives with respect to z.
Applying Laplace transform to equation (146) using the initial condition (147) yields

dc
sc + v =0 (149)
dz

Integration of this equation along with the Laplace transform of boundary condition (148) leads
to the solution in the Laplace domain:

εzs
c = ci exp( − ) (150)
v

From equation (106) we know that

εzs  εz 
exp(− ) = L δ( t − ) (151)
v  v 

Applying the convolution theorem to equation (150) and using equation (151) leads to
27

t
εz
c( t , z ) = ∫ c i ( t − τ)δ( τ − ) dτ (152)
0
v

which yields, using equation (102),

εz
c( t , z ) = c i ( t − ) (153)
v

Recall that v is the superficial velocity of fluid through the medium, v=Q/AB, where AB is the
cross-sectional area of packing and Q is the volumetric flow rate (equation 2). The surface area
that the fluid really occupies is Af=εAB, so that a more realistic average velocity of the fluid in
the pore space is Q/Af; i.e., vi=v/ε. This is called “interstitial velocity”. The ratio

tR=z/vi (154)

that appears in equation (153) is the residence time of the fluid between the entrance of the bed
(z=0) and the position z. Equation (153) then states that the concentration at a given position z
will be identical to the inlet concentration, but delayed a time tR, which agrees with the concept
of plug flow: particles of fluid (e.g. contaminant molecules) move uniformly with the fluid
velocity, without spreading.

The chromatography model


Let us go back to the original formulation of Model 4, given by equations (8) to (13). The
problem consists of two coupled PDEs with initial and boundary conditions:

∂c ∂c ∂ 2c
ε + v = D 2 − a v k m (c − kĉ s ) (155)
∂t ∂z ∂z

∂ĉ s
= k m (c − kĉs ) (156)
∂t

c = ĉ s = 0 , t=0 (157)

c=ci(t), z=0 (158)

c=finite, z→∞ (159)

We start by applying Laplace transform to the differential equations (155) and (156), which,
using the initial conditions (157), leads to

dc d2c
εs c + v = D 2 − a v k m ( c − kĉs ) (160)
dz dz
28

s ĉs = k m ( c − kĉs ) (161)

From this last equation we find:

km
ĉs = c (162)
s + kmk

Note that the solid and fluid concentrations are directly proportional in the Laplace domain. The
solid surface concentration can be found by applying the convolution theorem to equation (162)
and recalling that

{
L e − k m kt = } 1
s + kk m
(163)

The result is

t
ĉ s ( t, z ) = k m ∫ e − k m k ( t − τ ) c( τ, z )dτ (164)
0

Hence, a knowledge of c(t,z) will allow us to find ĉ s ( t, z ) from this equation. In what follows,
we will focus on the fluid concentration. Equation (162) can be substituted into equation (160) to
obtain, after manipulations, the following ODE

d2c v dc
2
− − f (s ) c = 0 (165)
dz D dz

where

s a k 
f (s ) =  ε + v m  (166)
D s + kk m 

The characteristic equation associated with the ODE (165) is

v
r2 − r −f =0 (167)
D

and the roots are

2
v 1 v
r= ±   + 4f (168)
2D 2  D 
29

Note that one of the roots is positive (r+) and the other is negative (r-). The solution of equation
(165) is, therefore,

c = Ae r+ z + Be r−z (169)

By virtue of boundary condition (159), we must have A=0. Then, applying the Laplace transform
of boundary condition (158) we find

B = ci (170)

Therefore, the solution in the Laplace domain is

  2 
  v 1 v 
c = ci exp z

−   + 4f   (171)
  2D 2  D  
  

Finding the inverse Laplace transform of this function could be a formidable task (if at all
possible) due to the complexity of the dependence of c on s. However, this solution in the
Laplace domain can be used to find specific quantitative aspects of the fluid concentration and,
eventually, an approximate solution. For this purpose, we will use the method of moments.

Analysis of transient responses: the method of moments


Moments are parameters used commonly in the description of probability distributions. In this
and subsequent section, we will apply these parameters to the description of transient response of
the concentration of a solute in the chromatography model. We will associate our definition of
moments to the concentration of solute c(t,z), but the definitions can be easily generalized to any
function of t. The nth moment of c(t,z) is defined as


m n (z) = ∫ t n c( t , z)dt , n=0,1,2... (172)
0

Note that the moments are functions of position (since time is integrated out).
The zeroth moment is the integral of the concentration over all times at a fixed axial position.
Let dM be the number of moles of the transported species that pass through a cross section of the
bed at position z in a time interval dt. We can state that

dM=Qcdt (173)

Therefore, the total number of moles of species that pass through the cross section is


M T = Q ∫ cdt = Qm 0 (174)
0
30

The zeroth moment is then directly proportional to MT. If the species is conserved, i.e., if the
species is not consumed or produced in chemical reactions and it is not irreversibly lost to the
solid phase, then, as a consequence of mass conservation, m0 will be a constant (independent of
z).
The nth normalized moment is defined as

m n (z)
µ 'n (z) = , n=1,2,... (175)
m 0 (z)

Consider again equation (173). Integration of this equation between t=0 and t yields the moles
that have passed through the cross section before time t:

t
M ( t ) = Q ∫ cdt (176)
0

In general, M will be also a function of z but, for the purposes of the present analysis, we will
consider z constant (i.e., we are analyzing what happens at a specific axial position.) Note that
M(∞)=MT. The function M=M(t) can be inverted to find t=t(M). This last function expresses the
total time required for a number of moles M to pass through the specified cross section. In a
conservative system, this would be the time that all the molecules in M spent in the section of the
bed between z=0 and z. We can define residence time as the average of this function; i.e., the
average time spent by molecules of the species in the section of the bed between z=0 and z:

MT
1
tR =
MT ∫ tdM (177)
0

Performing a change of variable in this integral using equation (173), we get:


Q 1 ∞
tR =
MT ∫ tcdt = m 0 ∫0
tcdt (178)
0

and, comparing with the definition of the first normalized moment (equation 175), we see that

t R = µ1' (179)

i.e., the first normalized moment is the residence time.


The nth central moment is defined as


1
µ n (z) = ∫
m0 0
( t − µ1' ) n c( t , z)dt , n=2,3,… (180)

For example, the second central moment is


31

∞ ∞ ∞ ∞
1 1  2 '2

m 0 ∫0 m 0  ∫0
µ 2 (z) = − µ = − µ 1∫ + µ 1 ∫
' 2 '
( t 1 ) c ( t , z ) dt  t cdt 2 tcdt cdt  (181)
0 0 

which leads to

2
µ 2 (z) = µ '2 − µ1' (182)

The second central moment is a measure of the spread of the concentration field around the
residence time. In statistics, the second central moment of a probability distribution is called the
variance of the distribution.
The moments represent specific characteristics of the function c(t,z), especially with regards
to its behavior in time. Next we will show how the moments can be obtained from the Laplace
domain solution, c (s, z) . Consider the definition of the Laplace transform of c(t,z),


c (s, z) = ∫ e −st c( t , z)dt (183)
0

Now, we will expand the exponential in a Taylor series around t=0:

s 2 2 s3 3
e −st = 1 − st + t − t + ... (184)
2! 3!

Substituting into equation (183), taking factors containing s outside the integrals, and using the
definition of moments, equation (182), leads to

s2 s3
c = m 0 − sm1 + m 2 − m 3 + ... (185)
2! 3!

Now consider a Taylor series expansion at constant z of the function c (s, z) around s=0:

dc s2 d2c s3 d3c
c = c (0, z) + s (0, z) + (0, z) + (0, z) + ... (186)
ds 2! ds 2 3! ds 3

Comparing equations (185) and (186) we find

dnc
m n = (−1) n (0, z) (187)
ds n

Therefore, the moments of c(t,z) can be obtained directly from the Laplace transform of c(t,z).
As an example, consider the case of passive dispersion in a packed bed in response to a pulse
input. The solution in Laplace domain is given by equation (129),
32

M  v  4εsD  
c= exp  1 − 1 + z  (129)
 2 
Q  2 D  v  

From equation (187) for n=0, we have

M
m 0 = c (0, z) = (188)
Q

as expected (equation 174 with MT=M). Also from equation (187), we have

dc
m1 = − (0, z) (189)
ds

From equation (129) we can show that

dc Mzε 1  v  4εsD  
=− exp  1 − 1 + z  (190)
4εsD  2 
ds Qv
1+  2 D  v  
v2

which yields

M zε
m1 = (191)
Q v

and, from equation (180), we get


µ 1' = (192)
v

which we recognize as the residence time of the fluid between z=0 and z.
The second moment is given by

d2c
m2 = (0, z) (193)
ds 2

which requires taking the derivative of equation (190) with respect to s. After doing that and
evaluating it at s=0, we find, after many manipulations,

z 2ε 2 2zε 2 D M
m2 = ( + ) (194)
v2 v3 Q
33

From equations (175) and (194), we find that

z 2ε 2 2zε 2 D
µ '2 = + (195)
v2 v3

and using equations (182) and (193) we get

2zε 2 D
µ2 = (196)
v3

We see that the second central moment is directly proportional to the dispersion coefficient. This
indicates that a high dispersion coefficient will lead to a wide spread of the response in time.

Moment analysis of Model 4


We found the solution in the Laplace domain (equation 171):

  2 
  v 1 v 
c = ci exp z

−   + 4f   (171)
  2D 2  D  
  

with

s a k 
f (s ) =  ε + v m  (166)
D s + kk m 

Consider a pulse input:

M M
ci (t) = δ( t ) , ci = (197)
Q Q

Evaluating the zeroth moment yields

M
m 0 = c s =0 = (198)
Q

Still, the zeroth moment is proportional to the injected number of moles; i.e., all injected moles
will eventually pass through any section of the bed and m0 is then a constant.
The first moment is (algebra will be omitted)

dc M zε  a v 
m1 = − = 1 +  (199)
ds s =0 Q v εk 

or
34

zε  a v 
µ 1' = 1 +  (200)
v εk 

The first term of this equation represents the hydrodynamic residence time, which is also the
residence time for the transported species in the passive dispersion case (equation 192). The
factor av/εk represents the fraction of the hydrodynamic residence time that is added to the solute
due to its temporary retention on the solid surface during adsorption. For this reason, av/εk is
called retardation factor. Equation (200) provides the basis for using chromatography as a
separation process: two components with different affinities for the solid (i.e., different k's), will
have different residence times and will elute from the column at different times. However, the
fact that two solutes have different µ 1' does not mean necessarily that they will be completely
separated: because of the spread, their elution curves may overlap. It is therefore important to
determine the second moment:

d2c
m2 = (201)
ds 2 s =0

Using equation (171), we can find m2 and then determine the second central moment. The result
is

2
2zε 2 D  a v  2a v z
µ2 =  1+  + (202)
v3  εk  vk m k 2 ε

Comparing with equation (196) we can see that the retardation factor adds to the spread by
means of dispersion (first term). We can also see that the mass transfer coefficient (in reality its
inverse, which can be interpreted as a mass transfer resistance) adds to the spread, even though it
does not affect the residence time. Note that for fast mass transfer (km→∞) only dispersion
contributes to the second central moment.
As an example of application of the previous analysis, consider the elution of pulses of m-
xylene, ethylbenzene and p-xylene from a chromatography column packed with a Zeolite-Y
(Morbidelli et al., Ind. Eng. Chem. Process Des. Dev. 1985, 24, 83-88). The elution curves (mole
fraction of the various species at the exit of the column as a function of time) are shown in
Figure 9.
The moments can be calculated from the experimental data using numerical integration of
equation (172). In this case, since the data are given in terms of mole fraction, the moments were
calculated using the reported mole fractions (x) instead of the molar concentration, c. So, the
moments (equation 172) were redefined as follows for this particular case,


m n = ∫ t n xdt (203)
0

Notice that the mole fraction is related to the species molar concentration by
35

mole 0.45
fraction
0.4 m-xylene
ethylbenzene
0.35
p-xylene
0.3

0.25

0.2

0.15

0.1

0.05

0
0 10 20 30 40 50 60
t(min)

Figure 9. Experimental elution curves from a chromatography column packed with Zeolite-Y.
The carrier solvent is isopropylbenzene (which does not adsorb on the solid.) The hydrodynamic
residence time in the column is Lε/v=0.9 min, where L is the total length of packing. Solid lines
are straight-line segments joining data points.

c = xc M (204)

where cM is the mixture's molar density. This means that the zeroth moment calculated from
equation (203) will be (see equation 198)

M M
m0 = = (205)
Qc M F

where F=QcM is the total molar flow rate of the mixture. This analysis assumes that the molar
density is a constant. The normalized and central moments will be all equal to those obtained in
the previous section.
The integrals were evaluated using the trapezoidal rule. To evaluate


I = ∫ g ( t )dt (206)
0
36

we used the formula

(g i + g i +1 )
I= ∑ 2
( t i +1 − t i ) (207)
all intervals

From the calculated values of m0, m1 and m2, the first normalized and second central moments
were evaluated. The results are shown in Table 1.

Compound m-xylene Ethylbenzene p-xylene


m0(min) 3.3 3.2 3.5
m1(min2) 57.2 70.8 106.5
m2(min3) 1031 1601 3340
µ1' (min) 17.3 22.1 30.5
µ '2 (min 2 ) 312 499 957
µ 2 (min 2 ) 11.6 12.9 26.4
R 18.2 23.6 33

Table 1. Moments of the elution curves in Figure 9, calculated by numerical integration using
the trapezoidal rule.

The fact that the zeroth moments are approximately the same for all three compounds means
that the injected pulses were the same (i.e., M was the same). From the calculated value of the
first normalized moment, µ1' , using equation (200) at z=L, we can calculate the retardation factor,
defined as

av
R= (208)
εk

knowing that Lε/v=0.9 min. The corresponding values are shown in Table 1. The results indicate
that the adsorption constant k is different for the three compounds. Knowledge of the structure of
the medium (av and ε) would allow us to determine k for each compound from the experimental
data provided. In addition, the use of equation (202) would allow us to find a relation between D
and km or, if km is large enough (which is a desired trait of a chromatographic process), we could
directly determine D for each compound directly.
The preceding example shows that a solution in the Laplace domain can be used to find the
moments of a response curve analytically. These relations, combined with experimental data, can
be used to obtain quantitative physicochemical information about the process.

Approximate solution: Hermite polynomials


A truly predictive solution of Model 4 would involve the inversion of the Laplace domain
solution given by equation (171) for a pulse input (equation 197). Since this inverse is not
available, we will find an approximate solution for c(t,z) based on the expansion of this function
in an infinite series. The leading term of the series will be a Gaussian distribution, which shares
some characteristics with the expected response to a pulse input (e.g. data in Figure 9).
37

The Gaussian distribution with mean t and variance σ2 is given by

1  1  t − t 2 
G(t) = exp −    (209)
σ 2π  2  σ  

where the distribution has been normalized, so that


∫ G( t )dt = 1 (210)
−∞

Figure 10 shows a qualitative representation of G(t). It is worthwhile to point out that the
Dirac delta function can be expresses as the limit of a Gaussian distribution as the variance
becomes small (equation 98).

G(t)

95.5% of the area is


contained in this region

t − 2σ t t + 2σ
0.5 0.6 0.7 0.8 0.9 1 1.1 1.2 1.3 1.4 t1.5
Figure 10. The Gaussian distribution and some of its properties.

One important feature of the Gaussian distribution is that it extends over the domain -∞<t<∞,
so that the definition of moments for this type of distribution must be extended to the whole
domain; i.e.


mn = ∫ t n G (t )dt (211)
−∞

Evaluating the moments of G(t) yields,


38

µ1' = t (212)

µ2 = σ2 (213)

As said before, the approximate solution to be proposed for Model 4 will be a series
expansion whose leading term is a Gaussian distribution. Even though the response is not
Gaussian (see the long tails and asymmetry of the data in Figure 9), higher order terms should
provide a correction to the basic Gaussian shape. The proposed series solution will be of the
form


c( t , z ) = e − x 2 ∑ a n H n ( x ) (214)
n =0

where

t − µ1'
x= (215)
2µ 2

with the condition that

H 0 (x ) = 1 (216)

so that the leading term is a Gaussian distribution. To find the coefficients of the series, an, we
will require that the basis functions, Hn(x), be polynomials and that they satisfy the orthogonality
relation


∫ e −x H n ( x )H m ( x )dx = 0 , for n≠m
2
(217)
−∞

A set of functions with these characteristics is the family of Hermite polynomials, which are
generated by Rodrigues' formula

d n e −x 2
H n ( x ) = (−1) n e x 2 (218)
dx n

The first few Hermite polynomials are (H0 is given by equation 216):

H1 ( x ) = 2 x (219)

H 2 (x ) = 4x 2 − 2 (220)

H 3 ( x ) = 8x 3 − 12x (221)
39

H 4 ( x ) = 16x 4 − 48x 2 + 12 (222)

They also obey the recurrence formula

H n +1 ( x ) = 2xH n ( x ) − 2nH n −1 ( x ) (223)

All these properties can be shown to hold from the generating formula, equation (218).
Furthermore, we can prove that they satisfy the orthogonality relation

∞ 0, n ≠ m
∫ e − x 2 H ( x )H ( x )dx =
n m  n (224)
−∞ 2 n! π , n = m

To find the coefficients of the series in equation (214), we multiply both sides of the equation
by Hm(x) and integrate t to get


∫ c( t, z)H m ( x )dx = a m 2 m m! π (225)
−∞

Now, at a given z, from equation (215) we see that

dt
dx = (226)
2µ 2

so that we can express equation (225) as (replacing m by n)


1 t − µ1'
an =
2 n n! 2πµ 2
∫ c( t , z ) H n (
2µ 2
)dt (227)
−∞

Because the functions Hn(x) are polynomials, we can see that the integral in the right-hand side
of this equation will be expressible in terms of moments of c (note that the lower limit of
integration can be taken to be zero since c(t,z)=0 for t<0.) For example, for n=0, we get

m0
a0 = (228)
2πµ 2

For n=1, we have, from equation (219),

t − µ1' t − µ1'
H1 ( )=2 (229)
2µ 2 2µ 2

Equation (227) then yields


40


1 t − µ1'
a1 =
2πµ 2
∫ c(
2µ 2
)dt (230)
−∞

Or


1 1
a1 =
2µ 2 π −∞
∫ c(t − µ1' )dt = 2µ 2 π
(m1 − µ1' m 0 ) =0 (231)

(since µ1' = m1 / m 0 ). Also, it can be shown that, after manipulations,

  ∞ 
2 
4 t − µ1
1 '
 − 2 cdt = 0
a2 = ∫
8 2πµ 2 −∞   2µ 2  
(232)
  

Next consider n=3. Using equations (227) and (221) it can be shown that, after manipulations,

m 0µ 3
a3 = (233)
24 πµ 22

i.e., a3 depends on the second and third central moments.


Keeping the first four terms in expansion (214), we find an approximation for the
concentration field:

   
3
  
 ( t − µ1' ) 2   1 µ3 8 t − µ1
'
 − 12 t − µ1
'
 
c( t , z) ≈ m 0 exp −  + (234)
 2µ 2   2πµ 2 24µ 22 π   2µ 2 

 2µ
 2

 
 

Figure 11 shows how this model fits the data in Figure 8. Note that the fits are satisfactory in the
sense that the model can reproduce the long tails and asymmetry of the response curves.
41

mole 0.45
fraction m-xylene
0.4
ethylbenzene
0.35 p-xylene
p-xylene fit
0.3
ethylbenzene fit
0.25 m-xylene fit

0.2

0.15

0.1

0.05

0
0 10 20 30 40 50 60
t(min)

Figure 11. Model predictions for the data in Figure 9 using equation (234).

You might also like