P.C.

Chau (UCSD, 1999)

A few comments on combination of variables Combination of variables, also named similarity transform, is one of the easiest methods of solving PDEs. The reading of the text is a lot easier if you have gone through the mathematical details first. Classical example and limitations of the method Circumstances when the combination of variables applies are even more restricted than the method of separation of variables. The additional constraint is now in the auxiliary conditionsthe problem must describe a semi-infinite domain, and the initial condition must be the same as the boundary condition at infinity. Textbook examples almost always begin with the problem ct = Dcxx with initial condition t = 0, c = 0, for all x (2) (1)

and boundary conditions x = 0, and x = , c = co , c = 0, t>0 t>0 (3) (4)

As written, we may think of this statement as equivalent physically to a gas absorption problem where a solute is diffusing into the liquid phase. What we model is the transient diffusion in the liquid, and x = 0 is the gas-liquid interface. The message a second time: We may have different values of initial and boundary conditions in different problems, but no matter what, the value of the initial condition at t = 0 must be the same as the boundary condition at x = . We now define the transform variable

x 2 Dt

(5)

and it's time for some calculus. We first note that

= 1 x t 3 / 2 = 1 x 1 = 1 t 22 D 2 2 Dt t 2 t
and

(6)

= 1 x 2 Dt

(7)

With Eqs. (6) and (7) at hand, we do the calculus for the terms in Eq. (1):

c c c = = 1t 2 t t
c c c 1 = = x x 2 Dt
and

(8)

(9)

2 c c , 2 = x x x

2 c c 1 2 c 1 = = 2 2 Dt x 2 4Dt x

(10)

We finally can substitute (8) and (10) into Eq. (1) to obtain

P.C. Chau (UCSD, 1999)

2c 1 c 1 t = D 2 4Dt 2
and after cancellation of terms and rearrangement, we have the final transformed equation

d2c dc + 2 = 0 d d 2

(11)

where we stop using the notation for partial derivatives since we now have a differential equation with only the independent variable . We turn to the boundary conditions next. Boundary condition (3) at x = 0 becomes = 0, c = co (12)

and most important of all, both the initial condition (2) and boundary condition (4) at x = become the same second boundary condition to (11) as = , c=0 (13)

Here you should observe immediately that integration of Eq. (11) once would lead to

dc 2 = a e d

(14)

where a is some integration constant.1 We find the constant a by integrating Eq. (14) with the two boundary conditions (12) and (13)
0

dc = a
co 0

2 e d = a 2

where the value of the integral on the RHS is well known in the sense that again, it is in your freshmen calculus text. So the constant a is

a = co 2
We go back to Eq. (14) one more time
c co

(15)

dc = c o 2

e d = c o erf ()
2

where we have used the definition of the error function. We finish the chore on the LHS, and we have the solution to Eq. (11) with its boundary conditions (12) and (13) in a clean looking

c() c o = erf () co
which we often rewrite as

(16)

c() = 1 erf () = erfc () co

(16a)

If you don't see (14), you should review your freshmen calculus. The textbook steps (also in Middleman p. 126) are to define first, say, u = dc/d and then integrate du + 2u = 0 which gives d 2 + ln(a), where a is some integration constant. ln(u) =
2

P.C. Chau (UCSD, 1999)

where erfc is the complementary error function. We are not done yet. Like most mass transfer problems, we want to know the flux, or in other words, the concentration gradient at some boundary. While we expect you to integrate Eq. (11), we understand that error function is new to many people and more importantly, few people really understand Leibnitz rule from introductory calculus.2 So we will evaluate the diffusive flux at the boundary x = 0,

D
With Leibnitz rule,

c x

x=0

= Dc o d [1 erf ()] dx

x=0

d erf () = 2 d dx dx

2 e d = 2

e d + e 2 1 x 2 Dt
2

In calculus, most problems would have and x being two independent variables and thus the partial derivative in the integral at the far right is zero. In our case, is a function of x. Still, we do not have to do any work! That's because we purposely want the flux at x = 0 which is equivalent to = 0. The value of the integral must be zero (since both the lower and upper integration limits are zero), and the exponential function in the second term is unity at = 0. Hence, the diffusive flux at the boundary x = 0 is simply

c x

= Dc o
x=0

1 = c o Dt

D t

(17)

and the cumulative amount transported over time (per unit area) is
t

D
0

c x

dt = c o
x=0

D t dt = 2c o

Dt

(18)

_________________________

Comments on Middleman, Chapter 4 Example 4.2.2, pp. 124-128 With the simplest example out of our way, we can move back to the first example in which Middleman applied the technique. First and foremost, if you have not tried to use the u = rC definition to transform the diffusion equation from spherical coordinates to a planar geometry, do it now! [In solving these problems, cylindrical coordinate is the "oddball," leading to solutions with Bessel functions.] Second, the algebra is a lot cleaner if we define (in fact most texts do), x = r Ro as the space coordinate. So let's do it. The (mathematical) problem statement in (4.2.60-63) is now ut = Duxx with t = 0, and at x = 0,
2

(19) 0x< (20)

u = 0,

u = uo ,

t>0

(21)

Leibnitz's rule is

d dx

b(x)

f(,x) d =
a(x) a

f d + f(b,x) db f(a,x) da dx dx x

where the integral limits can be constants or functions of x.

P.C. Chau (UCSD, 1999)

x = ,

u = 0,

t>0

(22)

With the transform variable as in our Eq. (5),

x , x=rR o 2 Dt

(23)

and following the identical intermediate steps from (6) to (10), we should obtain, again,

d2u du + 2 = 0 d d 2
with and = 0, u = uo

(24) (25) (26)

= , u = 0

And by now, it should be very obvious that we should also obtain the solution in Eq. (16a)

u() = 1 erf () = erfc () uo

(27)

which of course is Middleman's (4.2.96). If you are sharp, you could have arrived at this step as soon as you get to Eq. (23). We now substitute u = rC, uo = RoCsat, and the definition of back in to obtain

R r Ro c(r, t) = ro 1 erf ( ) 2 Dt c sat

(28)

which is (4.2.97). To find the flux, we know we need to take the derivative. The intermediate steps between our Eqs. (16) and (17) give us the cue on what to do with the error function. At r = Ro (equivalent to x = 0 when we first try these steps), the derivative of the error function must now be

d erf ( r R o ) dr 2 Dt

=
r = Ro

1 Dt

To find the concentration gradient of Eq. (28) at r = Ro, the intermediate step is hence

1 c c sat r

r = Ro

1 = R o 12 R o R o Dt

which leads finally to the result (4.2.98) in Middleman

c r

=
r = Ro

Dc sat Ro 1+ Ro Dt

(29)

We can appreciate the physics better if we rewrite the equation as

c r

=
r = Ro

D c + c sat R o sat

D t

(29a)

where on the RHS, the first term represents steady state diffusion while the second term is the transient contribution which gradually vanishes. Take note of the functional dependence on D and t; you will see this again in more "advanced" problems.

P.C. Chau (UCSD, 1999)

Section 4.2.1, pp. 132-135 Note that the initial condition and boundary conditions of (4.2.129-132) are the opposite of our Eqs. (1-4). So there is a sign difference in the solutions: Middleman's (4.2.137) vs. our Eq. (16). [We'll not repeat the exercise with Middleman's boundary conditions. You can try that yourself.] The first source of confusion comes in at the bottom of p. 131 when Middleman flipped the coordinate system. So the mass flux in his example is toward his negative y direction. Note that it is in (4.2.139) that Middleman made a sign change to look at flux in the reverse direction in his example. Then Eqs. (4.2.140-150) in Middleman are confusing and uncalled for. His final result (4.2.150) is equivalent to our Eq. (18). Is it wrong with what Middleman had done? No. What you can learn from Section 4.2.1 is that the choice of your coordinate system may depend on the solution method. But this section is definitely difficult and confusing to read for a beginner. You just have to read through it several times to soak it all in.

Comments on Middleman, Chapter 5 Section 5.1, p. 207 This is standard text book stuff with a slight change in notations. Nevertheless, the writing in the text creates much confusion. The key is to recognize that the mathematical problem statement is Eq. (5.1.15) with the three boundary conditions (5.1.11), (5.1.12) and (5.1.16). If we follow the steps as outlined in our first example, we should arrive safely at Eq. (5.1.25), a slightly more complex analogous form to our Eq. (17). The change of variables (to Y and u) on p. 207 distorts the physical picture, and is not necessary. The only justification is that it helps clean up intermediate algebraic steps. If we do so, we should recognize that the problem statement (5.1.18-21) has the same initial/boundary conditions as our first example, and Eq. (5.1.23) in Middleman is really our (16a). You should rederive and confirm the final results (5.1.25 & 26) in this section as an exercise.

Section 5.2, pp. 216-217 The situation where an approximate model takes the form of a semi-infinite medium pops up again in what is also called the Graetz-Nusselt problem. Let's fill in the missing mathematics details. The problem statement is ycz = Dcyy with auxiliary conditions z = 0, y = 0, and y = , c = 0, c = co , c = 0, for all y z>0 z>0 (31) (32) (33) (30)

We now define the transform variable

= y

9Dz

1/3

(34)

and the steps to follow parallel what we have from Eqs. (6) to the bitter(!) end. The useful calculus intermediate steps are

= 1 y 9D 3 z

1/3

z 4 / 3 = 1 z , and = 9Dz 3 y
5

1/3

(35)

P.C. Chau (UCSD, 1999)

c c c = = 1z 3 z z c c c = = y y 9Dz
and
1/3

(36)

(37)
2/3

2 c c 2 = 9Dz y

1/3

2 c = y 2 9Dz

(38)

Substitution of Eqs. (36) and (38) into (30), and after a few algebraic steps would eventually lead us to 2 d c dc (39) + 3 2 =0 d d 2 with and = 0, = , c = co c=0 (40) (41)

The general solution is

dc 3 = a e d
and the integration constant is evaluated from 3
0

(42)

dc = a
co 0

3 e d = a (4/3) , or

a=

co (4/3)

(43)

With a more general integration limit, we have our solution as analogous to Eq. (16) 4

1 c co = (4/3) co

e d
3

(44)

which you can clean it up with one more algebraic step. Of course, we are more interested in finding the concentration gradient at y = 0. When we take the derivative of Eq. (44) with respect to y and evaluate the gradient at y = 0, we have only the (Leibnitz's rule) term associated with the upper integral limitsame old steps that have led to Eq. (17). So we immediately can jump to our conclusion

c y

=
y=0

Dc o (4/3) 9Dz

1/3

(45)

which of course is Middleman's (5.2.13). Eq. (5.2.14) is just one simple integration step away. Take note of the hint given in the text: 4/3 (4/3) = (7/3).

The gamma function is defined as

(x) =

t x 1e t dt

but it permits direct integration when x is certain whole fractions. We can look them up in, say, Handbook of Mathematical Functions. Yes, we are skipping the details here. But take my word; the definition of (4/3) is correct. 4 Note that there is a slight difference between our Eq. (44) and Middleman's (5.2.11). That's because he chose to integrate (42) from any out to infinity, while we choose to integrate from = 0 out to any .
6

P.C. Chau (UCSD, 1999)

How do they figure out the transform variable? (Reading optional) Since this is not an applied math course, we always take whatever transform variable for granted. If you are curious how the variable is chosen, we'll try Eq. (30) as an example. It only takes a few calculus steps. Let's presume the transform variable has the form

= y a z

1/b

with unknowns a and b. The goal is to find values of a and b such that, after we introduce into the partial differential equation, the independent variables y and z disappear and only remains. If this is the case, the partial differential equation is transformed into an ordinary differential equation. Now, we have (we'll do something slightly different from before)

= = 1 y a1 / b z1 / b 1 = 1z b b z z = = a z y y
and
1/b

2/b

2 2 = y

a z

1/b

= a z y

2 2

Substitution of these intermediate steps in Eq. (30) leads to

a z

1/b

1 dc = D a z b z d

2/bd

c d 2

where we also have replaced y by the definition of . We want to get rid of the z terms. Collecting all the z terms onto the left-hand side gives z2/bz1/bz1 = z3/b1, which means that we must choose b = 3. Now the equation looks like

d 2 c + 2 a 2 / 3a 1 / 3 dc = 0 D b d d 2
The final step is to choose a such that the factor inside the square bracket is a nice multiple of 2 so that when we integrate the equation, we get a clean algebraic result. The choice here is a = /9D, and so this is how we get the definition of in Eq. (34). With a few more cancellation steps, you should find Eq. (39). We can take the same route to find in Eq. (5) for (1), but we'll skip that. Instead, we see why this method is also called similarity transform. The order of magnitude scaling of Eq. (1) is

c D c t x2
which suggests that the order of Dt and x2 are similar, and a change in Dt is similar in magnitude to a change in x2. If we take the square root of these two quantities, we can define a dimensionless variable x/Dt, which is the form used in the definition of . Of course, we need a bit more work to find the factor 1/2 in (5) so that the resulting equation (11) can integrate cleanly. With this simple scaling applied to Eq. (30), we can also see the logic behind the definition of in (34).