Evolutionary Search For An Optimal Limiting Process Flowsheet"
Evolutionary Search For An Optimal Limiting Process Flowsheet"
Evolutionary Search For An Optimal Limiting Process Flowsheet"
Abstract-The depth-first heuristic search for a process flowsheet proposed earlier by the authors is now extended
to generate good (optimal) limiting process flowsheets. A look-back strategy follows the initial look-ahead
procedure to eliminate the fragmentation of process tasks. Then an evolutionary search improves the process
structure by eliminating the earlier dependence on the order of consideration of individual product goals followed by
the implementation of by-product recycling to reduce raw material consumption. In recycling, preference is given
either to those by-product streams requiring the smallest changes to the current flowsheet or to those which maximize
the mass of recycled streams. Six look-back heuristics and three evolutionary rules are described.
Examples of evolutionary structure generation are chosen from the production of multicomponent hydrocarbon
mixtures and the production of monochlorodecane.
Scope-Defining precise procedures that can generate feasible chemical processing plant structures to achieve
optimal production goals is a lively area for basic research. Some of the procedures, especially in the area of
process separation trains and energy management networks, i.e. heat exchangers, etc., are probably close to in-
dustrial exploitation. Indeed, the latter are useful subprocedures to a general process synthesis package as we have
shown in this paper in the case of the Thompson & King[ll] separation scheme generating package. General
process synthesis programs are much further removed from immediate commercial potential. Part of the difficulty
is the heterogeneity and amount of the data with which general process structuring is concerned. Another
constraint is the fact that very few really new chemical enterprises are designed in any individual industrial group
in any given year. The experience and engineering expertise of people relative to any one of these few projects
rapidly outgrows the present capabilities of general synthesis programs.
The most significant prior work on general process synthesis is the AIDES procedure[ 101.We have already
contrasted the latter with our BALTAZAR procedure[6]. The AIDES procedure requires the user to determine a
species allocation scheme before the program is initiated. This requirement imposes tight constraints on any
subsequent possibilities of structure generation by the program in that a conceptual structure must already be
predicated. The tightening of structural constraints is compounded further after linear programming is applied to
each species allocation scheme in the system which imposes additional constraints on the fine structure of the
process. In BALTAZAR the search procedure allocates resources and matches goals stream by stream. No
preliminary species allocation is required, hence no implied structural constraints are given by a problem
specification. Intermediate results allow discarding a number of flowsheets without detailed design whereas AIDES
can only dispose of the same flowsheets after detailed design.
Nevertheless, in the work accomplished so far, a logical framework and a set of successful search procedures
have been developed which devise good material balanced and energy balanced process structures with the help of
a computer. The economic objective function, i.e. minimize cost or maximize profit, is considered only indirectly.
That is, heuristics or rules are employed which emphasize the best economic performance. The latter can only be
really judged when detailed design of alternatives is complete but the notion that mass flow in the plant must be
minimal, that raw material consumption is to be reduced and energy is to be conserved can not be disputed as
worthy intermediate objectives. The manner of their application in this work has been tested against ultimate
economic objectives and found to lead to optimal solutions.
What remains to be done beyond the further enrichment of the process structural selection rules are interfaces.
Three major interfaces are recognizable: (1) the human interface which allows an engineer to manipulate strategy,
to understand the results and to innovate new technology; (2) the data interface which provides extensive access to
reaction chemistry, to physical properties, to design standards and to safety and reliability criteria and (3)
interfaces to a number of homogeneous synthesis procedures such as separation trains, energy networks, reaction
paths, catalyst design and control systems.
Conclusionsand Significance-A procedure for the design of chemical processing systems can be decomposed into
three major steps:
(1) Determination of the overall material balance constraints and an initial feasible flowsheet.
(2) Generation of a limiting optimal process flowsheet containing alI of the processing tasks.
(3) Detailed design of the individual processing tasks to reduce them to hardware elements.
In the first step, sources of species are identified either as raw materials or reaction chemistries. Then a
one-step, look-ahead heuristic strategy[6] leads very quickly to a feasible flowsheet.
*Presented at: 82nd National Meeting, Session 23a, American Institute of Chemical Engineers, Atlantic City, New
Jersey, 29 August-l September 1976. _
tPresent address: College of Technology, University of Zagreb, Zagreb, Yugoslavia.
*Person to whom correspondence should be addressed.
149
CACE
Vol. 1. No. 2-D
150 V. MAHALEC
and R. L. MOTARD
In the second step, which is the major contribution of this paper, undesirable features of the initially feasible
flowsheet can be successfully eliminated by an analysis of already generated portions of a process structure using a
look-back strategy. This look-back strategy is supported by an evaluation function or criterion which leads to the
optimal integration of separation tasks (C&S).Finally, the structure is further refined by evolutionary search for
optimality implemented by redefining the substructure producing each particular (intermediate) product. A second
evaluation function (GL) is considered in the evolutionary procedure to identify opportunities for the recycling of
by-products and resealing. The substructure redefined during an evolutionary search is that subset of the process
tasks between the product and the raw materials. If such a substructure is also used to satisfy some other
requirements, they too are to be met by the new substructure. The evolutionary procedure combined with the
depth-first search strategy leads very quickly to a good limiting process flowsheet.
The thud step follows standard chemicalengineeringpractice, however the generation of a limiting process
flowsheet as a prelude to hardware design focusses the attention of the designer on any opportunities for
innovations which are outside of existingtechnology.
1
A chemical processing system is very often evaluated Total mass of
according to the cost of production per unit mass of the GSt
processed material
product. This criterion can be used only if a rather GL=
Total mass of + Total mass of
detailed design of the system is known. Rapid evaluation
of different limiting process structures and a selection of useful products I [ recycled streams 3
1
some of them to be designed in detail, requires a Total mass load
criterion based on variables which are known to be
[ of reactors
significant without a detailed design. t
[ 1
At the highest level of abstraction, a flowsheet con- Total mass of ’
sisting of tasks only (i.e. neither material nor energy useful products
balances are available) can be evaluated by the number
of separations it includes. In a flowsheet with a known The total mass term reflects the expenditure on the raw
material balance, the most expensive operations in a materials. The mass of recycled streams term is in-
chemical processing system are separations and chemical troduced since recycled streams are useful intermediate
reactions. The mass load on separations and their num- products requiring additional processing operations with
ber can be combined to estimate the quality of the respect to a serial structure. The term including total
design. A total mass load on separation tasks by itself is mass load of reactors is introduced in order to evaluate
not a sufficient criterion. It is possible to have two the same processing system operating under different
flowsheets with the same total mass load on separation conversion levels in reactors.
tasks but one of them may include a much larger number
of separation units than the other. The investment for the
first flowsheet is significantly higher than for the second DERH-FIRST SEARCH FOR A LIMITING
one. PROCESS FLOWSHEET
Since the initial specifications are based only on A heuristic procedure for the quick generation of
material balance considerations, the internal structure of limiting process flowsheets has been developed by Ma-
the separation sequences should not be specified at this halec & Motard[6]. The objective of the procedure is to
level of problem decomposition. The separation minimize the number of separation sequences and their
sequence structures are to be defined through a pro- mass load. The procedure is based on a successive
cedure considering the material balance and the physical comparison of available raw materials and reaction
properties. Therefore, a good performance criterion products as species sources with the required products
favors the process flowsheets which do not determine and reactor inlets as species sinks. All species in the
the internal structure of separation sequences, but in- source are simultaneously compared with all species in
stead of that develop only good material balance con- the goal. A species source most similar to a particular
straints on entire separation trains. sink is used to satisfy the requirements.
Let SEP be a set of separation subsequences in a Let s be a stream selected to achieve G. The total
process such that there are no interlinks between any available quantity of s is qso. The quantity qs of the
two elements of SEP. stream, s, to be processed is calculated via a tie com-
As a measure of the investment and operating cost of a ponent, a species present in both s and G. Based on the
process, the following criterion can be used: consideration of concentration in the stream, s and the
product, goal G, there are three possible choices for a tie
Criterion GLOBAL-SEPARATION, GS
1Ix[ I[ 1’
Total number of Number of separation Number of outlet Mass load of the ith
GS=
separation subsequences tasks in subsequence j iej streams in i minus one separation task
If each separation has only two outlet streams, the GLOBAL-SEPARATION, GS criterion reduces to:
12 [ I[ 1’
Total number of Number of separation Total mass load
separation subsequences j=SEP tasks in subsequence j of units in j
b) LUMPING OF MIXERS,
t
RULE 2
I
5
20
+ 5
GS=60
3”
considered streams, is introduced. S 4.33 13.0
(6) Two separations can not be parallelled. BC
These rules are illustrated by Fig. 2.
To demonstrate the characteristics of the DFSFF al-
AB 067
gorithm, the following example will be used.
Quantity
Species (lb moles/h)
Case I C3 Cd CS AC 2.0
Case II Cd C, CS r
Case III C4 CS C3 PRdDUCT I
Raw material 1 0.168 0.833 1200.0
Raw material 2 0.385 0.385 0.321 1300.0
Product 1 0.5 0.5 200.0
GS=20.6
Product 2 0.417 0.417 0.167 1200.0 GL=l.82
Product 3 0.091 0.909 - 1100.0 PRODUCT 3
PRODUCT 3
PRODUCT 2
6.0
-r
1.2 8.67
A 1.33 AB
II
JlII
GS = 26.9
d) PRODUCT 2 GL = 2.01
The problem has been solved with all possible order- EVOLUTIONARYSEARCH FOR AN OPTIMAL
ings of products considered, by employing the Key SR LIMITING PROCESS F‘LOWSHRRT
look-back strategy as shown in Fig. 3. Even though some Importance of mixing constraints
identical solutions may be obtained with different goal Let us consider a process flowsheet generated by some
ordering, the dependence of the solution on the goal solution procedure, for instance DFSFF. Is it possible to
ordering still exists. improve the flowsheet, in particular the separation
In order to demonstrate the validity of the limiting operations?
process flowsheet concept and the evaluation functions An analysis of any process flowsheet reveals that the
GS and G, let us associate real substances with the separation operations are determined by the mixing of
species in the 2R3P problem. In particular, species A, B intermediates into final products. In other words, once a
and C are chosen among normal paraffins C3, C4 and Cs. decision what to mix has been made, rather tight material
All six possible permutations of the species are ex- balance constraints have been imposed on the possible
amined. It is assumed that in all cases raw materials are process structure. Therefore, in order to find an optimal
at 250 psia pressure and 122°F temperature. Data is given structure, mixing tasks in a process have to be re-
in Table 2 and represent a resealing of 2R3P. examined. A single stream product can be considered as
Several different solutions to 2R3P were presented a special case of mixing (mixed with itself). The evolu-
earlier. The best solutions are shown in Fig. 3(b) and (c). tionary search for an optimal solution is based on the
They represent material balance constraints on a real redefinition of mixing tasks and their corresponding
process structure. In order to examine these solutions, an substructures.
existing procedure [ll] has been used to design the se- If the initial solution is obtained by any procedure
paration sequences. The basic cost of a distillation tower which considers assigned products in a sequential man-
is taken from a graph in Peters & Timmerhaus [7, p. 6591. ner, then the solution to the first goal is obtained when
Utility costs are: water $0.02/1000gal, electricity no process structure is yet available. On the other hand,
$O.Ol/kWh, 500 psia steam %0.60/1000lb, 100psia steam the final product is attempted when solutions to all other
$0.5O/lOOOlband exhaust steam $0.25/lOOOIb.The cost goals are known. Previously generated portions of the
of the solution obtained by separating all mixtures into structure can be used to achieve part of or the complete
pure substances and then mixing them in required ratios final product. It can be expected that the largest im-
is given for comparison. The four design solutions, Fig. provements of the existing flowsheet will be obtained by
Criterion GL Species A B C
Structure Mass load C3C4Cs C4CICS C&x, C&G CFX‘l C3CsC4
364 2.01 I: 29 49 53 75 45 33
833.0 0: 47 38 28 45 40 29
3(blt 1.82 I: 39* 48 36 43* 46 38
1033.0 0: 36 52 21 26 33 23
wt 1.82 I: 50 38* 36* 61 42* 29*
1033.0 0: 34 31 18 34 23 24
3(d) 2.01 I: 48 41 56 71 47 30
913.0 0: 31 30 15 40 36 25
Complete 4.00 I: 73 73 62 138 95 78
separation 2500.0 0: 63 63 50 110 100 68
3(a)-(d), for each selection of A, B and C in terms of regeneration of the first generated, the oldest part of the
annual costs are contained in Table 3. structure.
Detailed design of limiting process flowsheets confirms Let PR = J&r,, pr2, . . .,prk} be a set of mixtures in a
that the flowsheets (b) and (c), optimal according to the process flowsheet which impose the material balance
GS and GL criteria, are indeed optimal solutions. constraints on the process. An element of PR can be
either a single product stream, or an outlet from a mixer
Table 3. Production of monochlorodecane,MCDC (40% con- or a reacting mixture. The following rule is defined.
version of DC)
Evolutionary Rule 1. Replace a substructure produc-
Item T, “F P, psia Composition Quantity ing pri by another substructure satisfying the same
requirements.
nn Cl* 75.0 44.1 1.0 Cl 400.0 The following strategy is employed to implement
n?lDC 75.0 14.7 1.0 DC 600.0 Evolutionary Rule 1.
pr MCDC 75.0 14.7 1.0 MCDC 71.5 (1) Let pri be the oldest stream in PR not yet con-
Reacting
mkture 212.0 350.0 0.4 Cl, 0.6 DC 1.0 sidered as an evolutionary goal. Remove pri from the
process flowsheet and convert it into a goal.
Reaction 0.36 DC, 0.282 HCl
(2) Destroy the structure upstream from pr,. Destruc-
product 212.0 210.0 0.118C1,0.198MCDC, 1.0
tion is carried out until raw materials are encountered.
0.042 DCDC
(3) Outlet streams of a destroyed upstream path used
Evolutionary search for an optimal limiting process flowsheet 155
T
(4) Arrange the generated subgoals in the first-
generated first-attempted sequence, i.e. the subgoal cor-
responding to pn is attempted first.
s 2
(5) Find a solution.
(6) Designate pn as the considered mixture, stream. s5
Repeat (l)-(6) until all elements of pr have been con- PRODUCT 3, AE PRODUCT 2, ABC
sidered. a) INITIAL SOLUTION
(7) If the same sequence of solutions has been
successively generated during execution, exit; a local
optimum has been found. Otherwise proceed.
(8) Designate all streams in PR as unconsidered. Go to
% GI
(I).
PRODUCT I
It may seem unnecessary to destroy a process struc-
ture up to the raw materials. In particular, if a unit in the PRODUCT 3
upstream path is used to satisfy some other substructure, PRODUCT 2. ABC
it may seem that the destruction should stop at that unit.
b) STRUCTURE IN a) DESTROYED FROM s3 UP
However, the remaining process structure is so rigid that
in most cases the same solution will be found. The sl5 A s6 C
second step ensures that enough freedom will be created
by the structure-destroying procedure. The third step
ensures that the new solution will satisfy the same
requirements as the old solution. The solution procedure
applied in Step 5 is the Key SR look-back DFSFF. s13
Steps l-6 are repeated in the evolutionary search in s5
Y
order to find a solution containing all dependencies PRODUCT 2 , ABC
between productions of different mixtures. For instance,
the DFSFF initial solution is such that the last attempted T sl6
PRODUCT3, AE
GS=3B.O
GLc2.52
goal is achieved when all other solutions are already
c) SOLUTION $5 b)
known. On the other hand, the first attempted goal is 62
achieved when none of the solutions to other goals is yet
known. This also motivates the first-generated first-at-
tempted ordering of regenerated subgoals.
PRODUCT I rm2
initial solution is the Key LR (3, 2, 1) goal ordering d)STRUCTURE IN c)DESTROYED FROM s5 UP
solution. After two evolutionary steps the optimal solu-
tion is found, Fig. 4(e). This particular solution has also
been found by the Key SR (3, 1,2) goal ordering DFSFF
procedure, Fig. 3-b.
It is to be emphasized that the ordering of goals (3, 2,
1, 3, 2) employed in the evolutionary steps is different
from the ordering (3, 1, 2) which leads to the same
solution by the DFSFF. The next evolutionary step
GS=20.6
generates the optimal solution shown in Fig. 4(g) which is
GL=l.B2
the same as Fig. 3(c). This solution is found in the
succeeding evolutionary steps. Hence, the search PRDDlkT 3 PRODUCT 2
procedure terminates. e)SOLUTION OF d)
Recirculation 51 s12
The procedures described so far have dealt explicitly PRODUCT I
with a generation of serial substructures. In particular
s20
the best combination of the species sources and a 63
process structure has been sought. An analysis of the M
evolutionary search procedure described in the previous
sl6
section, shows that a cyclic structure may be generated -_
during evolutionary synthesis. Let us consider a process
consisting of a reactor and a separation sequence, Fig. f)STRUCTURE IN e)DESTROYED FROM s7 UP
156 V. MAHALEC
and R. L. MOTARD
AC PRODUCT
_ A PRODUCT I --J_
R S
---f
01 ACYCLIC STRUCTURE
PRODUCT
PRODUCT 3 PRODUCT 2 R S
g)SOLUTION OF f)
-+I
b) INTERMEDIATE INCOMPLETE STRUCTURE
G2
PRODUCT
PRODUCT I
M R S
the design specification on the separation train is: se- Distillation is chosen by the program as the operation
parate MCDC from the remaining components. The employed to separate chlorine from hydrogen chloride.
initial structure found by DFSFF has been further Since the purpose of this study is to develop an overall
improved by evolutionary search. The recycle of decane process structure, such a design has been accepted
is introduced as the first recycle loop, Fig. 6(b). The cor- without an attempt to design some other separation
responding flowsheet is shown in Fig. 7(b). Criterion GL process, e.g. absorption. The cost of processing units is
indicates that it is profitable to recycle decane. The next summarized in Table 5. Sample cost calculations for
evolutionary step results in the recirculation of chlorine, flowsheets (b) and (c) (CIZ:DC = 4.6 feed composition
Figs. 6(c) and 7(c). Again, criterion GL evaluates the new and 0.4 decane conversion) are given in Table 6.
structure as better than the two previous solutions.
The preliminary evaluation of the monochlorodecane
manufacturing processes is carried out by the com- Table 5. Capital cost for manufacturing MCDC
putation of the criterion G. Figure 8 shows G vs the
conversion of chlorine for two different feed ratios and Production: 1 x 10’lb MCDClyr
flowsheets 7(b) and (c). Criterion G indicates that the Operation: 330dayslyr
Conversion of DC: 0.3-0.7, 212°F 350psia
conversion of chlorine should not be below 0.3, since
then the amount of unreacted material becomes ex- Plowsheets (a) and (b) Installed costs, $
cessively large. For conversion 0.3 through 0.6, both Tanks, steel, cylindrical
chlorine and decane are to be recycled, flowsheet 7(c), 2-150,000gal
while at conversion above 0.6, only decane should be l-15,aoogai
recycled, Rowsheet 7(b). One has to accept these es- 1-8aoogal loo,wo
timates with caution and confirm them by detailed Pumps, steel, centrifugal
design. l-100 gal/min
l-40 gal/min
t-60 gal/min 20,000
Distillation
I
, RECIRCULATION OF DC
AND Cl DC distillation 80,ooO
il MCDC distillation ~,~
GL Flash drum 25,000
Reactor 450,ooo
Additional heat exchangers 23,OOG
TOTAL 1,t00,Oocl
Contingency, misc. @ 32%
I
Total fixed capital 1,450,OOo
, \;,RECIRCULATION OF DC
4 Plowsheet (c)
Chlorine distillation 300,000
Other equipment cost 1,100,Ooa
TOTAL 1,406~
Total tixed capital 1,850,OOo
Conversion of DC 0.1-0.2