Solution of Material Balances for Flow

Sheets Modeled with Elementary Modules:

The Constrained Case
Flow sheet material balance calculations are greatly simplified if all
process input streams, reaction stoichiometry, and conversions, as well as
all stream and component split fractions, are specified. Yet in many ap-
plications, especially those arising in process design, the available material MOHINDER K. SOOD
balance specifications can take on many alternate forms. Each of these
additional specifications can readily be expressed as a linear or nonlinear and
constraint equation involving species or total flow rates.
In this paper, an approach is developed for incorporating such specifi-
G. V. REKLAlTlS
cations within the computational framework for solving unconstrained ma- School of Chemical Engineering
terial balance problems. The linear as well as the nonlinear constraints are Purdue University
accommodated by generating a parametric solution to the underlying un- West Lafoyette, Indiana 47907
constrained problem. The parameters are determined by solving simultane-
ously first the linear and then the nonlinear constraint equations. Only as
many equations are solved iteratively as there are nonlinear constraints.

SCOPE
The necessity of incorporating specifications other than vergence is attained, Convergence, if it occurs, is by defi-
the natural input parameters of the standard modules em- nition a limiting process which often requires a large
ployed in modeling a flow sheet has, of course, been rec- number of iterations. Yet the underlying unconstrained
ognized for quite some time in the computer aided process problem, which constitutes the bulk of the equations, is a
design and simulation literature. In the sequential modular linear problem and thus should not require iterative solu-
approaches, such constraints are typically accommodated tion.
by means of additional iteration loops around the module The goal of this paper is to present a computational
whose normal specifications are undefined provided the strategy which exploits the natural decomposition between
constraints involve stream quantities associated with that material balance and constraint equations. The methodol-
module. If the constraints involve streams not incident to ogy developed is based on the generation and use of a
the underspecified module, then outer loops of iterations general parametric solution to the underlying uncon-
involving the entire flow sheet may be required. In the strained problem. The construction of the parametric
simultaneous approaches, constraints are incorporated by solution is accomplished by solving the unconstrained
directly adjoining these equations to the remaining flow problem a finite number of times equal to one plus the
sheet equations. If these equations and the conjtraints number of imposed problem constraints. The parameters
themselves are linear, then the constraints introduce no of the parametric solution are then determined by sepa-
particular computational burden. However, if the con- rate solution of the constraint equations. The validity of
straints are nonlinear, then the only alternative is to linear- this type of approach is strongly dependent on the struc-
ize them and to solve the entire linearized problem itera- tural properties of the model equations. Linear algebraic
tively. Both approaches require repetitive solution of the proofs are used in the paper to confirm the validity in the
entire material balance problem until satisfactory con- case of flowsheet material balance problems.

CONCLUSIONS AND SIGNIFICANCE

A methodology for accommodating constraint specifica- constrained material balance problem. These solutions are
tions in computerized flow sheet material balance calcula- then wed to construct a general parametric solution to
tions has been presented. As shown, a properly specified the unconstrained problem. The parameters of the gen-
constrained material balance problem with P constraints is eral parametric solution are calculated by expressing the
equivalent to a set of N-P constant coefficient linear equa- constraints in terms of these parameters and solving a set
tions, a set of P equations of special structure each in- of equations no larger than P. If the constraints are linear,
volving one of the P unknown natural parameters, and a this can be done directly without iteration. If only some
set of P linear or nonlinear constraints. A simple rule for of the constraints are linear, they can be partially solved
selecting a set of values of the P unknown natural param- to reduce the number of undetermined parameters. In the
eters is given, and it is shown that this rule will lead to a presence of nonlinear constrai:its, the numerical solution
set of linearly independent solutions of the underlying un- of a set of nonlinear equations no larger than the number
of nonlinear constraints is required. The use of the para-
0 The American Institute of Chem-
0001-1541-79-1149-0220-$01.15. metric solution form also facilitates identification of im-
ical Engineers, 1959. proper problem specifications.

Page 220 March, 1979 AlChE Journal (Vol. 25, No. 2)

 The technique presented is general in that it is inde- problem solutions, although intuitive, is very much prob-
pendent of the strategy used to solve the unconstrained lem structure dependent. Cases in which the property does
problem, a sequential modular or simultaneous approach not hold are easily constructed, Hence, the property must
could be employed. The methodology could, in principle, be separately verified for applications different from those
be employed with more detailed constrained flow sheet discussed in the paper. A material balancing program
simulation calculations. However, the property that a implementing this approach has been developed, and the
linearly independent set of natural simulation parameter sample results obtained with several large test problems
vectors leads to a linearly independent set of unconstrained do indicate that the methodology is practical and robust.

In a predecessor paper (Sood et al., 1979), a calculation linear, then the only alternative is to linearize them and
approach was developed for solving the material balances to solve the entire linearized problem iteratively until sat-
associated with process flow sheets expressed in terms of isfactory convergence is attained. The methodology de-
the elementary mixer, splitter, separator, and stoichio- veloped in this paper avoids the need for iterating on the
metric reactor modules. This approach was based on the complete problem by decomposing the calculation into
key assumption that all process input streams, reaction three separate stages: solution of the underlying uncon-
stoichiometry, and conversions, as well as all stream and strained problem, simultaneous solution of the linear con-
component split fractions, are specified. Yet in many ap- straints, and iterative solution of the nonlinear constraints.
plications, especially those arising in process design, the In principle, this strategy for accommodating constraints
available material balance specifications can take on many can be applied both in the sequential as well as the simul-
alternate forms. Typical examples of these might be speci- taneous approaches, although in the latter case it is only
fication of the composition of an undesirable species in a necessary to do so with nonlinear constraints. The only
recycle stream, product composition and flow rate require- requirement is that solutions of the unconstrained prob-
ments, ratios of reactant flows to a reactor, ratios of stream lem be carried out with reasonable efficiency.
compositions in a separator, or required solids to solution
ratios in settling operations. Each of these additional speci-
fications can readily be expressed as a linear or nonlinear THE CONSTRAINED MATERIAL BALANCE PROBLEM
equation involving species flow rates. In the previous work, the following elementary module
In this paper, we will develop an approach for incor- parameters were defined to comprise the set of natural
porating such additional specifications within the computa- simulation specifications: the species or stream split frac-
tional framework developed for unconstrained material bal- tions for each component separator or stream splitter, the
ance problems. In particular, we will show that these stoichiometry and key reactant conversion for all reactors,
linear or nonlinear constraint equations can be accom- and the species flow rates in all external process input
modated by generating a parametric solution to the under- streams. Following from this definition, a constraint speci-
lying unconstrained problem. The parameters in the gen- fication (sometimes called control specification) is under-
eral parametric solution can be determined by solving first stood to be any condition imposed on the species flows
the linear and then the remaining nonlinear constraints which is not one of the natural specifications. Further-
simultaneously. In this approach, only as many equations more, a constrained material balance problem (CMBP)
need to be solved iteratively as there are nonlinear con- is a material balance problem involving a flow sheet mod-
straints. The key idea underlying the strategy is derived eled in terms of elementary modules in which at least one
from the unpublished work of Kneile ( 1975). The strategy constraint specification is imposed.
itself has similarities to the work of Ravicz and Norman A CMBP problem may thus be viewed as an uncon-
(1964), but these similarities are only at the conceptual strained problem in which one or more of the natural
level. simulation specifications are unknown and are replaced by
The necessity of incorporating specifications other than constraints. Since it was previously shown that an un-
the natural input parameters of the standard modules em- constrained problem always has the correct number of
ployed in modeling a flow sheet has, of course, been recog- specifications to yield a unique solution (Sood et al.,
nized for quite some time in the computer aided process 1979), it follows that a correctly specified CMBP must
design and simulation literature (Ravicz and Norman, always have exactly as many independent constraints as
1964; Henley and Rosen, 1969; Hutchison, 1974). In the there are missing natural specifications. Of course, the fact
sequential modular approaches, such constraints are typi- that the specifications are correct in number does not
cally accommodated by means of additional iteration loops guarantee that they are proper; that is, the resulting set
around the module whose normal specifications are un- of equations will have maximum row rank. Hence, as in
defined (Seader et al., 1974) provided the constraints the unconstrained case, the assumption underlying the
involve stream quantities associated with that module. If subsequent development is that the combined set of speci-
the constraints involve streams not incident to the under- fications are, in fact, proper.
specified module, then outer loops of iterations involving Although constraint specifications can take many forms,
the entire flow sheet may be required. In the simultaneous in virtually all cases of practical significance they can be
approaches, constraints are incorporated by directly ad- reduced to one of four types of specifications: species or
joining these equations to the remaining flow sheet equa- total flow rates of internal or output streams, stream com-
tions. If the remaining flow sheet equations are linear positions, ratios of species flow rates, and ratios of stream
and the constraints themselves are linear, then the con- compositions.
straints introduce no particular computational burden The first three of these are simply linear equations in
(Hutchison, 1974). However, if the constraints are non- the species flow rates. The last is the chief source of non-

AIChE Journal (Vol. 25, No. 2) March, 1979 Page 221

linear constraint equations in ffow sheet material balance The verification of the proposition is quite straightfor-
problems. Any relation of the form ward. For illustrative purposes, we will demonstrate state-
ments 1 and 3.
yij = mykl The species balance equations for a reactor expressed
in terms of the conversion X k of key reactant k are
involving mole (or weight) fraction yij of species i in
stream j gives rise to the following equivalent equation in
the species flows N i ( j ) : NsoUt= N i n + (a,,- Qar
x,> Nkin s = 1, . . ., s
s#k
Ni(l) Nk'"
S
=m
S NkoUt= Nkin ( 1 - x k )
Eliminating the conversion from the first set of equations
s=l s=1 by using the second equation, we have
If the streams j and I are the same, then this equation re-
duces to one of the third type. Otherwise, it remains a NsoUt= Nsinf -( Nkin- NkoUt)
Qsr
S = 1, . , .,S
quadratic equation. Typical examples of constraint equa- - rkr
s#k
tions of this type are the equilibrium relations of a flash
unit. Even more common are the composition constraints These are constant coefficient linear equations in the un-
inherent in the definition of a flow splitter. A flow splitter known species flows. The balance on the key reactant is
is simply a device which divides a stream flowwise with- the only equation involving the unknown natural simula-
out disturbing the stream composition. If the split fractions tion parameter X k .
are specified, then stream compositions are conserved by Similarly, if the stream split fraction for a splitter with
assigning the same split fraction to each species in a two output streams is unknown, then the splitter equation
given output stream. If the split fraction is not known, Ns(l) = tsN,(in)
then in order to insure that stream compositions are con- s = 1, .,s
served, the splitter conditions NS(2) = (1- t s )N,(i")
and
ts = ts s = 1, . . ., s - 1

S S can be written as
must be explicitly imposed between each pair of output Ns(in) = Ns(l) + Ns(2) s = 1, ., S
streams i and 2. In general, if there are M splitter output
streams, there will be ( S - 1)( M - 1) such conditions.
Of course, these quadratic constraints can be reduced to
linear constraints if the composition of one of the streams 8 S

incident to the splitter is known, or, in the case of the and

flash unit, if the I< values and the vapor to liquid total flow Ns(l) = t,N,(in) s = 1, .,s
ratio are known. However, frequently these quantities are
not known, and then such simplifications can not be made. where the ts are unspecified natural parameters, Of these
A properly specified CMBP will thus, in general, con- three sets of equations, the first S are constant coefficient
sist of an unconstrained material balance problem with a linear equations, the next S - 1 are the quadratic splitter
certain number of unknown natural parameters and an constraints, and the last S involve the S unknown natural
equal number of constraint equations, some of which are simulation parameters. Thus, under the separator conven-
linear and others of which are nonlinear but typically tion, whenever the stream split fraction is not specified,
quadratic. The structure of the underlying unconstrained the splitter can be considered to have S unknown natural
material balance is further elucidated via the following specixations and S constraint specifications: the S - 1
elementary proposition. To simplify the statement of the quadratic splitter constraints and one additional constraint
proposition, it is assumed that separators and splitters each specifications which must be imposed to replace the un-
involve only two output streams. We refer to this assump- known stream split fraction. As an immediate consequence
tion as the separator convention. The extension to multiple of the above proposition, the following classification of the
output streams is obvious. equations associated with any ChlBP can be given, This
Proposition 1: suppose each stream vector involves S classification assumes that the separator convention given
species. Then: previously is employed and that all unspecified process
1. Any reactor module with unspecified conversions feed streams are always introduced into the flow sheet
can be represented by S - 1 constant coefficient linear through mixer modules. The latter assumption, which we
equations plus one equation involving the conversion. will call the mixer convention, imposes no loss of general-
2. Any component separator module with P unspecified ity. If an unknown feed stream is entering directly into a
species split fractions can be represented by 2s - P con- reactor or separator, a dummy mixer whose other input
stant coefficient linear equations and P equations, each in- streams are zero can always be introduced without effect-
volving one of the P split fractions. ing the solution of the CMBP. Of course, the number of
3. Any splitter module with unspecified stream split streams is increased by one and the number of balance
fraction can be represented by S constant coefficient linear equations by S.
equations, S - 1 quadratic splitter restrictions, and S Proposition 2 : Under the separator and mixer conven-
equations involving the unknown split fraction. tions, any properly specified CMBP consisting of N in-
4. Any unspecified species flow rate of a process input dependent module equations with €' constraint specifica-
stream will give rise to one equation involving that un- tions is equivalent to an equation set consisting of N - P
known flow rate and will reduce by one the set of constant constant coefficient linear equations, P equations each de-
coefficient linear equations associated with thc module to fining exactly one unknown natural simulation parameter,
which it is input. and P constraint equations.

Page 222 March, 1979 AlChE Journal (Vol. 25, No. 2)

 Since the CMBP is assumed to be properIy specified, If x is N dimensional and C has rank N - P, then it is
there must be P unknown natural specifications corre- clear that the w(i) will correspond to any set of P linearly
sponding to the given P constraint specifications. As independent solutions of the system Cx = d.
shown by proposition 1, each missing natural speci- This construction suggests that an explicit general soh-
fication in effect reduces the set of constant coefficient tion to the set of N - P constant coefficient linear equa-
linear equations by one and will introduce one equation in- tions of the CMBP can be developed directly from a suit-
volving that natural simulation parameter. Hence, proposi- able set of particular solutions. One merely needs to de-
tion 2 is verified provided that both a species flow rate vise a systematic procedure for generating particular solu-
of a process input stream and a natural parameter asso- tions which can insure that these solutions will be linearly
ciated with that species in the module to which it is input independent. A convenient and simple way of generating
are not both unspecified at the same time. However, a particular solution is to arbitrarily assign values to the
that special situation is accommodated by the second as- P simulation parameters zi in Equations (1) and ( 2 ) . As-
sumption underlying proposition 2. Since all unknown feed suming that the assigned values will constitute a proper
streams are always introduced through mixers, the above specification, the resulting unconstrained material balance
situation can not occur. Each unknown feed stream will problem can be solved using any available method. The
simply give rise to a single equation involving this un- solution obtained will, of course, depend on the values of
known natural parameter; the remaining mixer balances the zi selected, and any two solutions generated from
will simply consist of constant coefficient linear equations. two different values of the zt will be distinct. However,
As shown by the next proposition, the equations involv- they need not necessarily be linearly independent in gen-
ing the unspecified ideal parameters have a rather special eral.
form. For convenience in notation, we indicate the aggre- As will be verified in proposition 4, the following choice
gate of the species flows in all streams except the Drocess of the parameters zican be guaranteed to lead to linearly
rnput streams by the vector x. That is, x f N(2), independent solutions. Specifically, consider a CMBP with
NC3). . ., N(K)]. P unspecified simulation parameters, the first Q of which
Proposition 3: Under the mixer convention, each unspec- are associated with type 1 equations. Let the jth simula-
ified natural parameter zi will occur in one and only one tion parameter vector be given by
equation which will either have the form
=j
or the form
%(i)

KXS
= fi (Zi)%i)

aV. . x3 .-- blzi + ci

(1)

(2)
qCj)

i =
LY~ if i

otherwise
where the ai and f i t are selected so that a( # pi, (ui # 0,
j=l and, if Q # P, that
where f i ( z i ) is a linear or affine function of zi; the aij, bi,
and ci are appropriate constants; and m ( i ) , n ( i ) are suit-
able index sets.
From the proof of proposition 1, it is clear that if the
Denote by v(j) the solution to Equations (1) through (3)
module input streams are not process feed streams, then
obtained when z is set equal to z(j).
each unspecified conversion or split fraction will give rise
Proposition 4: If the specifications z(j), j = 1, ., P satis-
to an equation of the first type. If, in addition, the module
fying the conditions given above are proper, then the set
input stream is a specified feed stream, then the unknown
conversions and split fractions will give rise to an equation of solutions v(j) obtained will be linearly independent. For
of the second type. For example, if the module is a reactor, proof, see the appendix.
It should be noted that the choice of C U ~ and pi in the
then the unspecified conversion will appear in the equation
above construction is quite arbitrary. To keep the solu-
NkoUt = Nkin - XkNkin tions generated close to physical reality, it is advisable,
where Nkinis the specified input stream. but not necessary, to choose values between 0 and 1. Fur-
As a result of this very elementary analysis, we can con- thermore, the condition 2 9i/pii # l is not really very
i
clude that any CMBP must always be viewed as consist- restrictive. In practice, it is highly unlikely that any arbi-
ing of a set of N - P linear equations trary choice of mi and pi will lead to a violation of this
Gx = h (3) condition. Hence, from our computational experience, a
check of the above condition can be quite safely ignored.
a set of P equations of either the first or second type, each
By using the above choice of the z ( j ) , we are thus as-
involving one of the unknown natural simulation parame-
sured that the generated solutions will be linearly inde-
ters, and a set of P constraints which are either linear or
pendent. Thus, if one further solution vo is obtained, say
quadratic in the unknown species flow xi. This classifica-
tion of the flow sheet module equation suggests that the
by setting zi = pi for all i, then the resulting P 1 un- +
constrained solutions can be assembled to generate a gen-
solution of these equations might well be carried out in eral parametric solution
stages, beginning with a parametric solution to the N - P
constant coefficient linear equations [Equation ( 3 ) 3. x = y(0) + 2 [v‘i’ - .(0)-jAi 3 vo + wx (4)
i
PARAMETRIC SOLUTION OF THE CMBP of the CMBP. As shown in the next two sections, this
It is well known that any system of linear equations Cx general solution can be inserted into the linear and non-
= d has a solution if and only if d lies in the range of C, linear constraints to yield specific values of the A’s. These
and, if so, then all solutions are given by specilk parameter values will, when substituted back into
the above parametric solution, in turn, give the desired
= w ( o ) + X [ , C i ) - w(o)])li specific solution to the CMBP.
i
Before proceeding with that discussion, we digress briefly
where w(0) is any solution to C x = d, the w ( ~ are
) any to point out that the explicit parametric solution can itself
basis of the range of C, and the xi are scalar parameters. be used to extract useful information about the problem

AlChE Journal (Vol. 25, No. 2) March, 1979 Page 223

being solved. For instance, it frequently happens in both Note that at this stage, too, the array D can be exam-
large and small material balance problems that the stoi- ined to determine if any invariants occur which ought not
chiometry of the reactions, the nature of the separations, to be subjected to additional specifications.
and the structure of the flow sheet combine to induce
flows or ratios of flows which remain invariant of other SOLUTION IN THE PRESENCE OF NONLINEAR
specifications. Such invariants can be easily recognized CONSTRAINTS
from the parametric solution, since those components of
x will be independent of the k's or will have rows which As noted earlier, a CMBP may, in addition to the pre-
are simple multiples of each other. These situations can viously discussed linear constraints, also involve nonlinear
be readily identified and flagged during solutions by ex- constraint equations:
amination of the W array. From this information, im-
proper (that is, dependent) specifications in the constraints gi(x) =0 i = 1, , ., P -L
can frequently be avoided. If the CMBP is properly specified, then these equations
will be sufficient to solve for the unknown parameters1
SOLUTION IN THE PRESENCE OF LINEAR CONSTRAINTS in Equations ( 7 ) . However, since the equations are non-
Given the explicit parametric solution, Equation (4), linear, their solutions will require the use of a suitable
to the underlying constrained material balance problem, iterative equation solving algorithm. For instance, if the
any set of linear constraints Newton-Raphson method is employed, then the follow-
Tx = u -ing iteration formulas expressed in terms of the variables
h and the gradients of gi(x) come into play:
associated with the CMBP can be accommodated by, in
effect, substituting the parametric solution into the con- [ V,gD] ( A x ) ( k ) = - g {x 13(k)11
straints and solving the equation and -A ( k t 1 ) = x ( k ) + (AT)(k)
TWX = u - Tv0= y (5) where
If the CMBP only involves these linear constraints, g[x(x)l = g(D% 8 ) +
then assuming the rank of the product matrix TW is equal As usual, the iterations must be initiated with an initial
to P, the substituted Equation (5) will have a unique estimateXO. However, since t h e x s are mathematical con-
solution h'. Upon insertion into Equation (4),this k" veniences rather than physically identifiable quantities,
will then yield the required unique solution xu of the initial estimates of the%s are, in general, not easy to pro-
CMBP. Since TW is a P x P matrix, simultaneous solu- vide. Moreover, because of the nonlinearities, guarantees
tion of (5) only involves as many linear equations as there concerning convergence can not be given, and multiple
are constraints. Ailoreover, when any array reduction pro- solutions may well be encountered. Fortunately, most of
cedure is used to solve ( 5 ) , the rank of TW will neces- these difliculties can be overcome in the present appli-
sarily be verified, and hence the existence of any im- cation because the functions gi(x) have a very special
proper, that is, dependent specifications, will immedi-
structure. As noted earlier, the primary nonlinear con-
ately be recognized. If, in addition, row permutation in- straints occurring in material balance problems are quad-
dexes are saved during the attendant elementary row
ratic functions oh the form
operations, a set of dependent specifications can be identi-
fied so that constructive action can be taken in subsequent
problem solution attempts.
If, in addition to the linear constraints, the CMBP also
involves nonlinear constraints, then complete solution of
(5) is not possible. Instead, an array reduction procedure In this expression, each of the sums in the denominators
which involves explicit backward elimination can be em- represents the summation of the species flows associated
ployed to generate a partial solution of some components with a given stream. Hence, it is a sum over only a sub-
of X in terms of the others. In such cases, the linear con- set of the xi.
straints, in effect, serve to reduce the dimensionality of X. Since each x i ( x ) in the above functions is a known
Suppose the P constraints of L that are linear and P - L linear function of the I's, the derivatives are easily calcu-
that are nonlinear. Reduction of the augmented L x ( P + lated in general form. In particular if we let
1) matrix (TW, y ) will yield the equivalent array
(1s c, 4
where I is an L x L identity matrix, C is L X ( P - L ) , (9)
then, from Equation (7)
and d is a column vector. Corresponding to this effective
partitioning and reduction of TW, the vector X will simi-
A
larly be partitioned into h a n d 3 such that
A
h=Cx+d (6)
As a consequence, the parametric solutions to the CMBP Since the ratios on each side of Equation (8) are mole
become or weight fractions of certain species in selected streams,
x = WCT +
(wo Wd) + their values must always lie between 0 and 1. Thus, if by
(7) convention, Equation ( 8 ) is always formulated so that
=DX+R mi 1, then reasonable initial ratio estimates yio can al-
ways be obtained by selecting values in the range 0 6
and involve only P - L undetermined parameters. yio 4mi. From these estimates, it is easy to calculate initial

Page 224 March, 1979 AlChE Journal Wol. 258 No. 2)

 TABLE
1. TESTPROBLEM
CHARACTERISTICS

Number of
Problem
No. Application Source Mixers Reactors Separators Streams Species

1 Chlorobenzene Thatcher, 1962 3 2 5 17 5

2 Ammoniahitric acid Reklaitis, 1973 4 5 4 21 9
3 Cracking gas Nagiev, 1964 3 5 6 21 6
4 Natural gasoline Motard et al., 1969 8 1 9 28 10
5 Coal conversion process Parsons, 1974 18 13 32 105 17

ilq&
Kneile, 1975

BENZENE
estimates of the 2 s by simply setting each side of Equa- @ CHLORINFITION II

tion (8) equal to its respective y t . This leads to 2 ( P - L)

linear equations in the P - L variables x,,
which can be C6H .CLE
solved in normal form (Forsythe and Moler, 1967) to give
a minimum norm initial estimate of the 1;s. Alternately,
one can simply use only the right-hand ratio in Equation
(8) and obtain xo
by solving the linear system !-J.Im€ n--. ouor*llaa 13

s ( X ) = yiOmi[s%(X)] i = 1, . ..)P - L 0- 0.96 OICHLORCBENZENE ’‘

m 1Y
0.97 CHLORDEENZENE
or
Fig. 1. Chlorobenzene flow sheet.

[ 5 [5
P-L P--L

2
n=l
d k Z n +6k = yi’mi

i=l, ..., P - L
dm&+6m
I1 ever, providing adequate care is taken to save a calcula-
tion map for the unconstrained solution, the remaining
P unconstrained solutions can be obtained quite expedi-
tiously. As shown in our previous work (Sood et al.,
In either case, the initial estimates of theT;s obtained 1979), such a calculation map can readily be retained by
in this manner are, in our experience, sufficient so that saving the symbolic form of the mixer equations and the
problems of divergence or convergence to fictitious solu- precedence order for the mixer equation calculations.
tions are avoided. In fact, in all of the test cases run to
date, even the choice yio = 0 has proven to be satisfac- COMPUTER CODE AND RESULTS
tory when used in conjunction with the normal form of
A program called MBPI implementing this calculation
the equations for estimatingxo. strategy has been coded. The program solves the uncon-
Once the initial TO vector is calculated, Equations ( 7 ) strained subproblem using the mixer equation methodology
can be evaluated to obtain ~ ( 5 0 )With. this known, the of Sood et al. (1979). Provisions have been made in the
g i ( h ) and partial derivatives agi/aA, can be evaluated program for three types of constraints:
from Equations ( 9 ) and ( lo), respectively, and the usual 1. Species or total flow = constant.
Newton-Raphson iterations initiated. The key point to 2. Species or total flow = species or total flow x con-
be again noted is that in the Newton iterations only as stant.
many equations are solved as there are nonlinear con- 3. Mass Fraction of species i in stream k - mass frac-
straints, tion of species i in stream m x constant.
The last of these is the only source of nonlinear con-
SUMMARY OF THE CALCULATION STRATEGY straints. Constraints of this type always occur when the
split fraction of a flow splitter or an equilibrium flash unit
The algorithm for the solution of CMBP’s with P un- are not specified.
defined natural specifications thus can be decomposed into In implementing the above algorithm, there is some
four separate phases: choice as to how step 4 is to be carried out. If the com-
1. Solution of the underlying unconstrained material
balance problem P + 1 times to generate the general pa-
plete set of P +
1 unconstrained solutions generated in
step 1 are saved, then once the h’s are known, the final
rametric solution, Equation (4). solution can be determined from Equation ( 4 ) . However,
2. Simultaneous (partial) solution of the L linear con- for purposes of steps 2 and 3, the complete set of solutions
straints to generate the reduced parametric solution, Equa- is really not required. Hence, considerable storage savings
tion (7). can be achieved by only storing the solution corresponding
3. Iterative solution of the P - L nonlinear constraints to the xi actually involved in the constraints. In that case,
to determine the parameter vector x. however, once the A’s are determined, the P 1 material +
4. Recovery of the vector of stream flows via Equations balance solutions must be regenerated so that the com-
( 7 ) or ( 6 ) and ( 4 ) . plete final flows can be recovered. Hence, the storage
In this algorithm, the iterative calculations are clearly savings is achieved by about doubling the computation
isolated from those calculations which involve solution of time. This latter calculation mode was implemented in
linear equations. This decomposition is obtained at the MBP-I.
price of having to carry out P +
1 unconstrained soh- The program was tested on a number of flow sheet case
tions, simultaneous solution of t linear constraint equa- problems. The general characteristics of the problems are
tions, and solution of P - L linear equations to determine listed in Table 1. The smallest problem is a textbook ap-
initial estimates of the X Of these calculations, the un- plication (Figure l ) ,while the largest represents a solvent
constrained solutions are clearly the most costly. HOW- refined coal synthetic fuels complex, a flow sheet of which

AlChE Journal (Vol. 25, No. 2) March, 1979 Page 225

 62

- - - _ - -- - - ---I

Steam

Acid Gas Removal 1

80 H20
79

OIL

0 Reactors
Heavy Liquid Fuel A Separators

6 Fig. 2. SRC cool conversion process.

2. CORE AND EXECUTION

TABLE T I M E &QUIRE MENTS FOR CONSTRAINED PROBLEM SOLUTrONS

Ratio
Core of con-
Number of require- strained
equivalent Number of constraints ments Execution to uncon-
Problem linear ( K words) time' strained
No. equations Linear Nonlinear octal (CPUs ) CPU time

75 2
153 13 -1 42
47
1.37
10.2
4.2
14.6
120 - 2 43 1.1 2.8
250
1615
1
10 -9 46
105
10.6
140
11.5
14.3

CDC 6500.

Page 226 March, 1979 AlChE Journal (Vol. 25, No. 2)

is shown in Figure 2. Further details of the problems can conjunction with the sequential modular method for solv-
be found in the program user’s manual (Sood and ing the unconstrained problem.
Reklaitis, 1977) which is available from the authors upon
request. Computational statistics for the solution of the ACKNOWLEDGMENT
constrained form of these problems are summarized in The authors wish to acknowledge the advice of Professor
Table 2. The number of equivalent linear equations given H. Rubin in formulating some of the formal developments in
in the table refers to the total number of species balance this paper. The work was performed under ERDA Contract
equations associated with the flow sheet elementary mod- NO.EX-76-C-01-2275.
ules. This number is only intended to suggest the problem NOTATION
size if a simultaneous approach was used.

1
A = ( ~ j ) coefficients of the module equation of type
The CPU times reported were those required by the b = (bi) 2 containing unspecified natural parameters
Purdue Universaity CDC 6500 and include normal pro- c = (ci) zi
gram input and output overheads. The ratio of the con- c = reduction remainder of the array W
strained to unconstrained CPU times confirms that the D = WC, coefficient matrix of the reduced parametric
computation time is essentially proportional to the num- solution of the CMBP
ber of constraints. The computation time for nonlinear a = the vector 8 after array reduction
constraints does not appear to be substantially larger than e = vector with unit components
that for linear constraints, Moreover, the choice of initial f i ( 1 = affine function involving the unspecified natural
estimates does not appear to be critical; the three nonlin- parameter zi
early constrained problems were all successfully solved F = array of coefficients of module equation of type 1
with zero initial estimates. G = array of coefficients of the N - P constant coeffi-
The main alternative to the strategy proposed in this cient linear module equations associated with the
paper is the strategy conventionally used in conjunction CMBP
with the sequential modular method for solving material g o = nonlinear constraint functions
balance problems, namely, repetitive solution of the entire h = right-hand side vector associated with G
flow sheet problem with iterative adjustment of the un- K = number of unknown streams
known natural specifications until constraint satisfaction is L = number of linear constraints
achieved. Since the parametric solution strategy requires a M = number of splitter output streams
fixed, finite number of unconstrained flow sheet solutions, mi = proportionality constant in the ith nonlinear con-
while the conventional approach requires a sequence of straint
~ - ~ .
solutions that is only limited by an imposed relative error m ( i ) , n ( i ) = indexes of unknown stream variables asso-
criterion, the former strategy a p p e m to enjoy an inherent ciated with unknown natural parameters
computational advantage. Detailed computational experi- N(j) = vector of species flow in stream i
ments to confirm this superiority are tedious to carry out N = total number of flow sheet module balance equa-
because of the number of influencing factors: effects of tions
initial estimates, choice of convergence promotion method, P = number of unspecified natural parameters
etc. Fortunately, as shown by the results obtained in at- P = (pii) = diagonal matrix with elements pit =
tempting to solve problem 1 using the conventional h ( a i - Bi)
strategy, such detailed comparisons are, in fact, unneces-
sary.
4 = vector with elements 4i = ci + Bib{
Q = number of module equations of type 1 involving
Problem 1, the flow sheet for which is shown in Figure unspecified natural parameters
1, involves three unknown natural parameters: the flow s = number of species
of benzene in stream ( I ) , the conversion in reactor 1, T = coefficient matrix of the linear constraints
and the split fraction of chlorine in separator 2. Three con- t, = split fraction of components
straints are imposed: the product flow, [stream (14)], is u = right-hand side vector of the linear constraints
specified, the chlorobenzene composition of the product is v = (+I) = matrix of unconstrained solution vectors
specified, and the ratio of the chlorine mass fractions in w = matrix of homogeneous unconstrained solutions
streams (4)and (11) are specified. The problem thus has with columns 4’) - v(0)
two linear and one nonlinear constraint. This constrained Xk = conversion of species k
problem was solved by using the conventional modular x = aggregate vector of all species flows in all streams
approach with bounded Wegstein convergence promotion yij = mole or weight fraction of species i in stream j
on stream (7).Iterative adjustment of the three unspeci- z = vector of unspecified natural parameters
fied natural parameters was carried out using Brown’s
Greek Letten
algorithm (Brown, 1969; Brown and Dennis, 1971), as
arbitrary nonzero choices of values for the
implemented in subroutine ZSYSTM of the IMSL library
(see references). This algorithm is an efficient Newton
8
Z= constant
} naturalparameterzi
vector in the reduced parametric solu-
analogue for solving simultaneous nonlinear equations
which only require function values. Starting with initial tion of x
estimates of 2000, 0.2, and 0.5 for the feed, conversion, X = vector of unknown parameters in the parametric
and split fraction, ninety unconstrained solutions were solution of the CMBP
required to reach the actual values of 1919.205,0.23048, 8 = vector of scalar multipliers 8i
0.6283 within a relative error tolerance of lo-’. CPU a,, = stoichiometric coefficient of species s in reactor r
time on the Purdue CDC 6500 was 44.5 s. Using the
same tolerance, MBP-I required 1.37 s. Of course, differ-
ent times will he obtained with different initial estimates. LITERATURE CITED
Good estimates had to be provided to avoid divergence. Anon., “Demonstration I’lant : Clean Boiler Fuels from Coal,”
The problem could be solved with either four or eight un- OCR F h D Rept. No. 82, Interim Report No. 1, prepared by
constrained solutions, depending upon the storage strategy the Ralph M. Parson Company (1974).
used, if the parametric solution construction was used in Anon., The IMSL Package, International Mathematical and

AlChE Journal (Vol. 25, No. 2) March, 1979 Page 227

 Statistical Libraries, Inc., Sixth Floor-GNB Bldg., 7500 Multiplying both sides of each of Equations ( A l ) by O ( k ) and
Bellaire Blvd., Houston, Tex. 77036. summing, we have
Brown, K. M., “A quadratically convergent Newton-like method
based upon Gaussian elimination,” SIAM 3. Nun. Anal., 6,
No. 4,560 ( 1969).
, and J. E. Dennis, Jr., “On the Second Order Conver-
gence of Brown’s Derivative-Free Method for Solving Simul- From this it follows that both
taneous Nonlinear Equations,” Yale University Dept. of
Computer Science Technical Rept., pp. 71-77 (1971).
Forsythe, G., and C. B. Moler, Computer Solution of Linear
Algebraic Systems, Prentice-Hall, Englewood Cliffs, N.J. and Equation (A2)must be satisfied simultaneously. But this
( 1967). is a contradiction since it is assumed that ni $; pi and since the
f 1are bijections by construction. Hence, the first part is proved.
Henley, E. J., and E. M. Rosen, Material and Energy Balance
Computations, Wiley, New York ( 1969). To prove the linear independence of the solution vectors 19,
Hutchison, H. P., “Plant Simulation by Linear Methods,” Trans. j =Q + 1, . , , P, recall that for these cases zi(j) = pi for I =
Inst. Chem. Engrs., 52, 287 (1974). 1, . , Q. Thus, the set of equations which the P - Q solutions
KneiIe, R. G . , “Solution of Material BaIance Problems for Proc- must all satisfy contains a set of constant coefficient linear equa-
ess Design,” Ph.D. thesis, Purdue Univ., W. Lafayette, Ind. tions
Gx=h
(Dec., 1975).
Motard, R. L., H. McLee, and R. W. Barkeley, CHESS-Chem- and
ical Engineering Simulation System: User’s Guide, Technical
Fx = 0
Publishing Company, Houston, Tex. ( 1969). where the latter Q homogeneous equations are those of type 1.
Nagiev, M. F., The Theory of Recycle Process in Chemical In addition, each of these solutions must satisfy the equation
Engineering, Pergamon Press, Ltd., Oxford, England ( 1964).
Ravicz, A. E., and R. L. Norman, “Heat and Mass Balancing Ax = bz + c
on a Digital Computer,” Chem. Eng. Progr., 60,71 ( 1964). for its corresponding value z(j). The complete set of P -Q
Reklaitis, G. V., Class Notes, Purdue Univ., W. Lafayette, Ind. solutions must thus satisfy the matrix equations
( 1973).
Seader, J. D., W. D. Seider, and A. C. Pads, “Flowtran Simula-
tion: An Introduction,” CACHE Corporation, Ulrich‘s Book-
store, Ann Arbor, Mich. ( 1974).
Sood, M. K., and G. V. Reklaitis, “User’s Manual: Material Bal-
ance Program Version-I,” School of Chemical Engineering,
[:Iv=“ B] [81..1
where the elements of the diagonal matrix P are
+

Purdue Univ., W. Lafayette, Ind. (June, 1977).

, and J. M. Woods, “Solution of Material Bal- pi* = b i ( a i - p i ) i = Q + 1, . ., P,
ances for Flowsheets Modelled with Elementary Modules: the vector q has elements
The Unconstrained Case,” AlChE I., 25, 209 ( 1979).
Thatcher, C. M., Fundamentals of Chemical Engheering, pp.
356-358, C. E. Merrill Books, Columbus, Ohio (1962). and the P - Q component vector e has components unity, The
matrix ( P + q e T ) has tile same rank as a symmetric matrix
APPENDIX: PROOF OF PROPOSITION 4 consisting oi off diagonal elements equal to 1 and diagonal
Proof: Consider the system of equations
+
elements ( qi p i i ) / q i . To verify that under the assumption

Gx =h
~ m ~=i fi(zi)zn(i)
) .,Q
i = 1,
the matrix has full rank, we proceed by contradiction. Suppose
ZU$j = bizi + ci i=Q .,P
1,+ the matrix does not have maximal rank. Then, for some k,
1
there must exist scalars e ( j ) # 0, j = Q + 1, . ., P, and j # k
Note that for the first Q solutions u(j), the right-hand sides such that
of the last set of equations will be constants equal to bipi + + =z 1.
-
qk Pkk
#j)
ci. For the remaining P - Q solutions, the second set of equa- qk i#k
tions will have constant coefficients f i ( p i ) . Linear indepen- and
dence of the P solutions v(j) can thus be proved in two parts:
first by showing that each of the first Q solutions cannot he
linearly dependent on the remaining solutions, and second, by -,
showing that the other P - Q solutions are themselves Iinearly Q + l L i L P and i # k
independent.
To prove the first part, suppose that any one of the first Q Subtracting the first equation from each of the others, we ob-
tain
solutions, say v(1), can be written as a linear combination of
- (qk + pkk) - [ (qi +
the others. If so, then there must exist a set of scalars e ( k ) ,
k = 1, . , P , a n d k # l s u c h t h a t
”(1) = 2 ,g(k),,(k)
qk

qk
- Pii)

4i
- 41
1 eci)

01
k#1
e ( i )= - Pkk/qk Q t l L i L P and i + k
Consider the equation containing the Zth parameter piifqz

Xm(1) = f1(+7t(1) But this implies that

By construction

while the Ith solution must satisfy or, upon simplification, that

z(qi/pii) =1
t

Page 228 March, 1979 AlChE Journal (Vol. 25, No. 2)

which clearly contradicts the initial assumption. Thus, the right- it follows that the matrix of solutions V must have rank P - Q.
hand side of the matrix equation must have rank equal to Thus, the soIutions generated from the z ( i ) must be linearly
P - Q. Since the module equations themselves have full rank, independent, and the proposition is verified.
by hypothesis, that is

[f] hasrankN Manuscript received December 21, 1977; revision received Aclgwt 18,
and accepted September 6,1978.

Use of a Finite-Stage Transport Concept for

Analyzing Residence Time Distributions
of Continuous Processes
REED S. C. ROGERS
and
ROBIN P. GARDNER
Departments of Nuclear and Chemical Engineering
North Carolina State University
Raleigh, North Carolina 27650

This study is concerned with the derivation and application of the

finite-stage transport concept for modeling residence time distributions
(RTD's) originally proposed by Adler and Hovorka (1961). This six-param-
eter model is shown to be more versatile and accurate than the common
two-parameter tanks-in-series or axial dispersion models. Radioactive tracer
results are given for a number of particulate unit processes with special
emphasis on wet-discharge ball mills.

SCOPE
The residence time distribution (RTD) concept is be- six-parameter, finite-stage transport concept for increased
coming more arid more important to chemical engineers versatility and accuracy.
in the modeling of continuous, first-order processes by TO accomplish this objective, a complete derivation
the segregated flow approach (Himmelblau and Bischoff, of the model is included, along with an accurate and
1968; ~ ~ ~1962).~ It is~very ~advantageous
~ i to ~have l efficient
, parameter estimation method for it. In addition,
the model is incorporated into the segregated flow con-
analytical functions for the RTD's encountered as evi-
cept to model a general first-order process. Finally, radio-
denced by interest in the simple two-parameter tanks-in- active tracer results for a number of particulate unit
series and axial dispersion models. These simple models processes with on wet-discharge ball mills are
Often do not offer sufficient versatility or accuracy. There- given which indicate the increased versatility and accu-
fore, the present paper reexamines a previously proposed racy of this model over the simple two-parameter models.

CONCLUSIONS AND SIGNIFICANCE

In this study, the finite-stage transport concept of Adler explicit, closed-form analytical solutions to a wide class
and Hovorlta (1961) is investigated as a general transfer of steady state first-order processes by the use of the
function for industrial process analysis. Application of segregated flow concept with this model. Extensions of
the model is demonstrated experimentally on several the model and experimental methods that are often nec-
continuous particulate mineral processes. The study indi- essary or desirable when using radioactive tracers have
cates that the parameter estimation method proposed can also been derived and demonstrated.
provide a high degree of accuracy for fitting the finite- The finite-stage transport concept has the advantages
stage model to actual data and that it is possible to obtain over the simpler tanks-in-series and axial dispersion models
of greater versatility and improved accuracy. It provides
Correspondence concerning. this paper should be addressed to Robin
P. Gardner. these advantages while still yielding analytical models
for first-order process phenomena when the segregated
0 The American Institute of Chem-
0001-1541-79-2312-0229-$01.35,
ical Engineers, 1979. flow concept is used. The axial dispersion model does not

AiChE Journal (Vol. 25, No. 2) March, 1979 Page 229