A Robust Iterative Method for Flash Calculations
Using the Soave-Redlich-Kwong or the
Peng-Robinson Equation of State
L.X. Nghiem, SPE, Computer Modelling Group
K. Aziz, * SPE, Computer Modelling Group
Y.K. Li, SPE, Computer Modelling Group
Abstract
A robust algorithm for flash calculations that uses an
equation of state (EOS) is presented. It first uses a
special version of the successive substitution (SS)
method and switches to Powell's method if poor con-
vergence is observed. Criteria are established for an effi-
cient switch from one method to the other. Experience
shows that this method converges near the critical point
and also detects the single-phase region without com-
puting the saturation pressure. The Soave-Redlich-
Kwong 1 (SRK) and the Peng-Robinson 2 (PR) EOS's are
used in this work, but the method is general and applies
to any EOS.
Introduction
The calculation of vaporlliquid equilibrium using an
EOS in multi component systems yields a system of
nonlinear equations that must be solved iteratively. The
SS method is commonly used, but it exhibits poor rate of
convergence near the critical point. To overcome con-
vergence problems, Newton's method has been used by
Fussell and Yanosik 3 to solve the equations. The
drawback of Newton's method is the necessity of com-
puting a complicated Jacobian matrix and its inverse at
every iteration. Hence, for systems removed from their
critical point it involves more work to arrive at the solu-
tion than the SS method. Furthermore, the radius of con-
vergence of Newton's method is relatively small when
compared to that of the SS method; hence, a good initial
• Now with Stanford u.
0197·7520/83/0006·8285$00.25
Copyright 1983 Society of Petroleum Engineers of AIME
JUNE 1983
guess is required before convergence can be achieved.
The single-phase region usually is determined by com-
puting the saturation pressure and comparing it with the
pressure of the system. This requires additional work,
and it is sometimes difficult to decide whether a dew-
point or bubblepoint pressure, which involve different
equations, should be computed.
This paper presents a robust iterative method for flash
calculations using either the SRK or the PR EOS, both of
which have received much interest in recent years. The
proposed method combines SS with Powell's iteration,
which is a hybrid algorithm consisting of a quasi-Newton
method and a steepest-descent method. 4 The SS method
is used initially and is replaced by Powell's method if it
demonstrates poor convergence, thus taking advantage
of the simplicity of the former method and the robustness
of the latter. The SS method has been modified so that
the single-phase region can be detected without having to
compute the saturation pressure.
The nonlinear equations to be solved by an iteration
scheme could behave differently, depending on their
form and the variables for which they are solved. In this
paper three different approaches are considered with
Powell's method. One of the three approaches is based
on the minimization of the Gibbs free energy.
The convergence properties of the proposed schemes
are demonstrated by three example problems.
Equations in Flash Calculations
At a given pressure and temperature, flash calculations
yield the number of moles of component i in the liquid
phase, NiL, and in the vapor phase, N w , given that N i ,
i = 1 ... n c represents the number of moles of each com-
521
ponent in the feed. The thermodynamic criterion of
vaporlliquid equilibrium is that the Gibbs free energy,
G, be minimized, which gives the following necessary
conditions.
aG aG
- - = - - = 0 , i=1...ne· ................ (1)
aN iL aN iV
Since NiL + NiV = N i , it easily can be shown 5 that
aG aG
- - = - - - = G iL -G iV , i=l. .. ne'
aN iL aN iV
............................. (2)
where GiL is the partial molar Gibbs free energy of com-
ponent i in the liquid phase. When an EOS is used to
compute_the pa.Itial molar Gibbs free energy, the dif-
ference GiL - G iV is related to the fugacities fiL and fiV
as follows.
- - fiL.
GiL -GiV=RT In-- , 1= 1...ne' .......... (3)
fiV
The expressions for the fugacities using the SRK and
PR EOS's are given in Appendix A.
The equilibrium conditions of Eq. I yield
InfiL =0, i=1...nc ....................... (4)
fiV
or
Fi -I =0, i= 1...ne> ....................... (5)
where Fi is the fugacity ratio,fiLlfiV' Eq. 5 represents a
system of ne nonlinear equations that must be solved
iteratively for NiL or N iV .
Instead of solving directly for NiL or N iV , it is often
more convenient to solve for mole fractions, which are
combinations of N i , NiL, and N iV .
The definitions of mole fractions are given in Appen-
dix B. We also define moles of component i in liquid
phase per mole of feed as
niL =/LXi =NiLIN, ........................ (6)
and moles of component i in vapor phase per mole of
feed as
niV=fvYi =N/N . ......................... (7)
Eq. 5 can be solved for one of the following sets of
variables.
/L,x i, i = 1...n c - I (LX iteration).
/v,Yi, i= 1...ne -I (VY iteration).
niL, i=1...nc (NLiteration).
niV, i= 1...ne (NV iteration).
It easily can be shown that the variables in each of
these sets are independent and completely define the
system. Later in this paper, we will show how Powell's
method is used to solve Eq. 5 for these sets of variables.
LX or NL iterations will be used for predominantly liq-
uid systems.
522
The equilibrium-phase compositions also can be ob-
tained by minimizing the Gibbs free energy. It is shown
below how Powell's method readily can be used to ob-
tain the solution from this approach.
Since niL =NiLIN and niV=NiVIN, we obtain from
Eqs. 2 and 3:
I aG I aG fiL
- - = - - - = I n - ......... (8)
NRT an iL NRT an iV fiV
We see from Eq. 8 that In(jiLlfiV) is equal to the gra-
dient of GINRT with respect to niL and the gradient of
-GINRT with respect to niV. Hence, by applying
Powell's method to solve Eq. 4 instead of Eq. 5, we are
minimizing GINRT by using the quasi-Newton method.
However, in doing so, we do not take advantage of the
fact that In(IiLlfiV), i= 1...n e , form the gradient of
GINRT(-GINRT) with respect to niL(niV), i=1...ne
since we merely were trying to reduce the values of
In(fiLlfiV) to O. Hence, the refinements that could be ob-
tained from this additional information 6 were not in-
troduced. We define the iterations on niL and niV,
i = 1...n e, as NLM iteration and NVM iteration, respec-
tively. NLM iteration will be used for predominantly
vapor systems and NVM iteration for predominantly liq-
uid systems.
In practice, the equilibrium ratios K i = Y /x i,
i= 1...ne> are introduced. The following equations can
be derived.
xi=Z/[I+fv(Ki-l)], i=1...ne' ......... (9)
Yi=Kiz/[I+fv(Ki-l)], i=1...ne' ...... (10)
ne
zi(K i -I)
g(fv)= ~ =0. . ............ (11)
i=i I +fv(K i -I)
Eq. 5 also can be solved iteratively for K i , i=1...ne'
Once a set of K values has been determined, fv is ob-
tained by solving Eq. II, and x i and Yi are obtained
from Eqs. 9 and 10. The SS method usually is used in
this case and is described in the next section.
Method of Successive Substitution
It can be shown that K values are related to the fugacity
coefficients ¢iL and ¢iV as follows.
K,
'= ¢iL = fiLI(XiP) F,.~ .......... (12)
¢iV fiV1(YiP) Xi
In this solution, estimated K values are used to flash
the mixture at the specified pressure and temperature.
The liquid and vapor fugacities then are determined from
the phase compositions, and new estimates for K values
are obtained from Eq. 12. The procedure is assumed
converged when
~ (Fi _1)2 <E. . ....................... (13)
As initial estimates for K values, the following em-
pirical relationship 7 is used.
exp[5.37(1 +Wi) (I-~) ]
,
K.= Tn .......... (14)
Pri
SOCIETY OF PETROLEUM ENGINEERS JOURNAL
Solution of g(fv) = 0 CASE CASE
1 3
Every time K values are determined, Eq. 11 has to be
solved for Iv(O:5lv:5 1). g(O is a monotonically
decreasing function of ~ (dgld~ <0 for any 0 with
asymptotes at ~ = I/(1-K i ), i=1...ne' Let Ke be the
largest and K s the smallest K values. A necessary but not
sufficient condition for the existence of a root of g(O in
[0,1] is that [0,1] is included in the interval [l/(1-Ke),
1/(1- Ks)]. Three cases are depicted in Fig. 1. For Case
1, g(O) ~ 0 and g(1):5 0, the root is in [0,1]; in Case 2,
g(O)<O, the root is negative; and in Case 3, g(I»O,
the root is greater than unity.
In Case 1, Iv is obtained by usin~ either the false-
position method or Newton iteration. In Case 2, Iv is
set to 0.5 at the first iteration and to 0.0 at subsequent
iterations. In Case 3, Iv is set to 0.5 at the first iteration
and to 1.0 at subsequent iterations.
In practice we find that the procedure just described
yields a root of g(~) in [0,1] for a two-phase system, and
for a single-phase system the fugacity ratios liLlfiV'
i = I ... n e' converge to a constant value that could be dif-
ferent from unity. This scheme has been tested exten-
sively and has predicted the single-phase region correct-
fig. 1-Conditions for the existence of solution of g(~) in [0,1].
ly. It was recently shown 9,10 that this procedure is
equivalent to testing the stability of a phase based on the
tangent-plane criterion. 11

Powell's Method NL Iteration or NLM Iteration

If the SS method demonstrates poor convergence, we can
~i=niLlliiL ,i=1. .. ne .
switch to another, more powerful method of solution.
We selected Powell's hybrid method,4 which is sum-
NV Iteration or NVM Iteration
marized in Appendix C. The Jacobian matrix and its in-
verse are computed numerically at the beginning of the ~i=niVlliiV, i=1...ne'
iteration and updated subsequently using rank-one
matrices according to Broyden. 12 This way the variables are properly scaled, andfL and
Let the initial estimate of the solution (obtained from Iv remain in the interval [0,1] for - 00 < ~ n < + 00. Ex-
the SS method) be: perience has shown that better convergenc~ is obtained
if the n e - 1 components used in LX iteration correspond
LX iteration-fL' xi, i=1...n e -I ; to the n e - 1 smallest value Xi, and those used in VY
iteration correspond to the n e - 1 smallest values y i'
VY iteration-fv, Yi ,i=1...n e -I ; The results from the SS method determine whether the
system is predominantly liquid or vapor. If Iv> fL
NL iteration or NLM iteration-Ii iL , i = 1...n e ; (predominantly vapor system), LX, NL, or NLM itera-
tion is used; otherwise we use VY, NV, or NVM
and iteration.
In the single-phase region around the critical point,
NV iteration or NVM iteration-Ii iV , i = I ... n e .
Powell's method will yield identical compositions for the
liquid and vapor phases. Hence, if the K values from the
We define the iteration variables ~ i, i = 1.. .ne for solution are such that IKi -II<IO- 3 , with i=l...ne, a
Powell's method as follows. single-phase system is assumed. For convenience, this
single phase is called vapor if fv > fL' and liquid
LX Iteration otherwise.
Criteria for Switching From SS
to Powell's Method
~ nc =In(2..-I)=lnlv
fL fL
. The important step of the proposed algorithm is the
switch from SS to Powell's method. This should lead to
a decrease in execution time; otherwise the switch is not
VY Iteration necessary. The criteria for an efficient switch from one
method to another were determined using a ternary
~i=y;lYi ,i=1...n e -I system and were then tested on 10- and IS-component
mixtures.
~ =In(2.. -1) =In IL . The ternary system used was C0 2 /nC 5 /nC I6 . Fig. 2
nc Iv Iv shows the calculated phase envelope at 1,500 psia (10.3
JUNE 1983 523
 2
\ 11% nC5
\

-1

-2
N
..-
I -3
u..
1-'1'-
-4
52
01
.2
-5

-6

-7
85 a Calculafed 15

-8

mole % nC 16
Fig. 2-Phase envelope of C0 2 /nC s/nC 16 system at 1,500 psia
and 122°F. Flash calculations in Table 1 are for points
on Line A.
_9L--L__L--L__L--L__L--L__L--L__~
o 5 10 15 20 25 30 35 40 45 50
No. of SST
Fig. 3-Convergence rate of 55 for different mixtures of
C0 2/nC s/nC 16 at 1,500 psia and 122°F.

MPa) and 122°F (SO°C). To determine the convergence
behavior of the SS method for mixtures with different
compositions, a series of flash calculations was per-
formed for mixtures containing 88% CO 2 and various
mole percents of nC 5 and nC 16. These mixtures lie
along Line A in Fig. 2 and for notational convenience
will be referred to using only their mole percent of nC 5.
This line contains mixtures in the fully developed two-
phase region (e.g., 0% nC 5), mixtures near the critical
point (e.g., 8% nC 5 ), and mixtures in single-phase
region (e.g., 12% nC s ).
Plots of log ~(Fi -1)2 vs. the number of SS iterations
(SSI) for diffelrent mixtures are shown in Fig. 3. The
following can be observed.
1. The convergence of the SS method deteriorates as
the critical point is approached.
two-phase region (2 % nC 5), a mixture near the critical
point (8 % nC s), and a single-phase mixture (11 % nC 5)'
For a two-phase mixture, ~(Fi _1)2 decreases, andlv
1
reaches its ultimate value rapidly. For a mixture near the
critical point, the change of ~(Fi _1)2 and Iv is slow.
1
For a single-phase mixture, the change of Iv is very
rapid and a value for Iv outside [0,1] is obtained after a
few iterations. In this case, we set Iv to either zero or
unity and proceed with the SS method as discussed
earlier.
Based on extensive testing of the flash program, the
following criteria are proposed to switch from the SS
method to Powell's method.
~[Flk)-IF/~[Flk-1)-IF>ER' ......... (lSa)
1 1

2. After a few iterations the slope of each curve re-

mains almost constant for a lar~e number of iterations. Ilv(k) -lv(k-1) 1< E V , . . . . . . . . . . . . . . . . . . . . . (ISb)

This implies that the ratio ~[Fl ) -IF/~[Flk-1) -IF

1 1 EL<~[F,.<k)-IF<Eu, .................. (lSc)
is also almost constant during that period and can be 1

used as an indication of the rate of convergence of the SS and

method for a particular mixture.
3. For the 8% nC s mixture, ~(Fi _1)2 increases after
44 iterations and then decrease~ again after S6 iterations
(not shown). This type of behavior of the SS method oc-
curs frequently in the critical region.
4. For the 11 % nC 5 mixture, the single-phase region
is detected after 2S iterations.
Fig. 4 shows the change of ~(Fi _1)2 and Iv with
the number of SSI for a mixturb in the fully developed
o<Iv(k) < 1 ............................ (ISd)
All four criteria must be satisfied before switching to
Powell's method. Eqs. ISa and ISb indicate that the con-
vergence of the SS method is slow. Experience shows
that Eqs. ISa and ISb are more representative of the con-
vergence rate than ISa alone. Eq. ISc contains the op-
timal range of values of ~(Flk) -1)2 where the switch
1

524 SOCIETY OF PETROLEUM ENGINEERS JOURNAL

 TABLE 1-FLASH CALCULATIONS FOR A THREE-COMPONENT SYSTEM USING THE PR EOS

Vapor-Phase SS Plus Powell

Composition Mole Fraction Method 1 Method 2 Method 3
(% nC s) tv SSI SSI Plus FE SSI Plus FE SSI Plus FE
0 0.527 14 (0.157) NA NA NA
1 0.532 14 (0.157) NA NA NA
2 0.537 14 (0.160) NA NA NA
3 0.544 18 (0.172) NA NA NA
4 0.553 22 (0.198) NA NA NA
5 0.565 27 (0.213) 11 + 8 (0.191) 11 + 9 (0.185) 11 + 7 (0.194)
6 0.583 34 (0.250) 9+ 17 (0.235) 9+ 14 (0.213) 9+ 14 (0.217)
7 0.613 44 (0.293) 11 + 17 (0.251) 11 + 16 (0.237) 11 + 14 (0.228)
8 0.688 85 (0.484) 13+21 (0.263) 13+20 (0.273) 13+21 (0.280)
9 0 141 (0.682) 20+21 (0.322) 20+21 (0.247) 20+24 (0.321)
10 0 49 (0.281) NA NA NA
11 0 25 (0.177) NA NA NA
12 0 12 (0.129) NA NA NA

to Powell's method is economical. The upper bound, TABLE 2-1 O-COMPONENT SYSTEM COMPOSITION
C u, avoids the switch too far from the solution. Although
the convergence of the SS method slows down as the Mole
Component Fraction
solution is approached, it is not economical to change
C1 0.6436
methods near the solution. This is prevented by the lower C2 0.0752
bound, CL, in Eq. I5c. Eq. I5d prevents the switch to C3 0.0474
Powell's method when Iv is equal to zero or unity. nC 4 0.0412
Usually we set cv=1O- 2 , cu=1O- 3 , and cL=10- 5 , nC s 0.0297
nC 6 0.0138
and C R is chosen by trial and error. Its value increases
nC 7 0.0303
with the number of components. Typical values for CR nC a 0.0371
are 0.5 and 0.75 for 3- and lO-component systems, nC g 0.0415
respectively. For practical purposes, any value of C R nC 10 0.0402
around 0.6 works reasonably well.
Examples
The combination method (SS plus Powell) was used to 2 1.1
flash a 3-, 10-, and I5-component system. For nota-
11% nCs
tional convenience, we group the different iterations by 1.0
using Powell's method as: Method I-LX or VY itera-
tion, Method 2-NL or NY iteration, and Method 0 0.9
3-NLM or NVM iteration.
The number of SSl's and function evaluations (FE's) -1 0.8 I"
in Powell's method are reported along with the execution I
..!..
time in central processing unit (CPU) seconds for dif- -2 z
0.7
ferent methods of solution. FE corresponds to the 8%nC s 0
number of times the nc functions in Eqs. 4 or 5 are I -3 ---- ---
------
0.6 ~
f=
u
evaluated. The execution time is for the CDC CYBER
.
OJ
~ u...
I72TM computer and includes input and output of infor-
-4 0.5 ~
mation. A variation of up to ±0.03 CPU seconds for the w·- 0
same problem has been observed. ::!:
0
,;; -5
8% nC 5
0.4 a:
.2 0
Three-Component System \
\ <t
a..
Results of the flash calculations using the PR EOS for -6 \
\ 0.3 >
\
mixtures along Line A in Fig. 2 are summarized in Table \
I
1. The execution times are enclosed in parentheses. A -7 I 0.2
I
value of 0.5 was used for CR in Eq. I5a. "NA" in the I
-8 I
SS-plus-Powell column means that the switch to I 0.1
Powell's method was not necessary. SSI decreases for ~ 11% nCs
\
the detection of the single-phase region as we leave the -9 00
0 5 10 15 20
critical region. The convergence criterion is ~(Fi _1)2 No. of 55I
< 10 -8 except for Method 3, where the ~riterion is Fig. 4-Behavior of SS for mixtures in the fully developed two-
~(in Fi)2 < 10- 8 . phase region (2% nC s), in the critical region (8% nC s),
I
and in the single-phase region (11 % nC s).
Ten-Component System
Consider the volatile oil shown in Table 2. The critical
point of this mixture, computed by the method of Peng
JUNE 1983 525
 TABLE3-FLASH CALCULATIONS FOR A 10-COMPONENT SYSTEM USING THE SRK EOS

Vapor-Phase SS Plus Powell

Pressure Mole Fraction Method 1 Method 2 Method 3
(psia) tv SSI SSI Plus FE SSI Plus FE SSI Plus FE
2,700 0.575 89 (1.010) 16 + 19 (0.637) 16+ 18 (0.581) 16+ 18 (0.596)
2,710 0.572 96 (1.106) 16+21 (0.643) 16+ 18 (0.569) 16+20 (0.638)
2,720 0.568 103 (1.170) 17 + 17 (0.613) 17+18 (0.584) 17+20 (0.658)
2,730 0.565 112 (1.239) 17+ 17 (0.589) 17 + 17 (0.598) 17 + 19 (0.644)
2,740 0.559 123 (1.371) 17+36 (0.871) 17+17 (0.587) 17+19 (0.641)
2,750 0.556 137 (1.491) 18+28 (0.784) 18+ 17 (0.597) 18+ 19 (0.637)
2,760 0.552 154 (1.669) 18+ 18 (0.625) 18 + 19 (0.622) 18 + 21 (0.676)
2,770 0.548 183 (1.937) 18 + 18 (0.627) 18+ 18 (0.634) 18+24 (0.734)
2,780 0.541 214 (2.247) 19 + 19 (0.670) 19+ 18 (0.645) 19+28 (0.813)
2,790 0.536 263 (2.698) 19 + 19 (0.657) 19+19 (0.654) 19+29 (0.888)
2,800 0.530 354 (3.631) 19+20 (0.712) 19+19 (0.664) 19+27 (0.809)
2,810 0.519 NC 20+21 (0.764) 20+20 (0.712) 20+20 (0.716)

and Robinson \3 and the SRK EOS, is at 2,813 psia (19.4
MPa) and 324°F (I62.2°C).
Table 3 shows the results of flash calculations using
the SRK EOS for this mixture at 324°F (162.2°C) and
different pressures from 2,700 to 2,810 psia (18.6 to
19.4 MPa). The convergence tolerance used in Eq. 13
was 10 - 8 , and the value used for ERin Eq. 15a was
0.75. The SS method did not converge at 2,810 psia
(19.4 MPa) ("NC" in the SSI column).
For the above two examples, Method 2 (NL or NV
iteration) appears to be slightly better than the other two
methods investigated.
Fifteen-Component Reservoir Oil System
The method presented here was also tested using the
reservoir oil described in Ref. 3, the composition of
which is shown in Table 4. The critical properties and
acentricities of the C 7 + fractions were estimated using
Bergman's method 14 together with the data given in
Katz and Firoozabadi. 15 The interaction coefficients be-
tween C 1 and the C 7 + fractions were adjusted to match
the bubblepoint pressure of 4,952 psia (34.4 MPa) at a
temperature of 235°F (I 12.78°C) using the SRK EOS.
Because a different EOS and a different characterization
method for the C 7 + fractions were used, the predictions
obtained were slightly different from those given in
Ref. 3.
The performance of the present algorithm was com-
pared with that of Newton's method as formulated by
Fussell and Yanosik. 3 However, the Jacobian matrix
was obtained by numerical rather than analytical dif-
ferentiation. This may not be as effective as in Fussell
and Yanosik's approach. VY iteration was used for all
calculations with a convergence tolerance E= 10 -12 and
ER=0.6 (Eq. 15a). The performances for two SS and
Powell's approaches and Newton's method are reported
in Table 5. In Col. 3, Eq. 14 was used as the initial guess
for all calculations using the present algorithm. In Cols.
4 and 5, a bubblepoint calculation was performed first,
and the results obtained were used as initial estimates in
the flash calculations at 4,950 psia (34.1 MPa). For each
subsequent pressure the initial iteration variables are the
value of these variables at the previous pressure. The
total execution times for all calculations in Cases A and
B are, respectively, 91 and 65 % ofthat corresponding to
Newton's method. A bubblepoint calculation was also
526
included in Case A to provide a fair comparison. Recent
developments on accelerated successive substitution 9
yield only 39% of the computing time of Newton's
method when Eq. 14 was used as the initial guess for all
calculations.
The results show that Powell's method is quite effi-
cient and usually takes less time than Newton's method.
For this example, the work involved in each Newton's
iteration is about 15 FE's plus an inversion of a 15 x 15
matrix.
Our experience shows that the SS method is very
stable, although its convergence deteriorates near the
critical point. From a poor initial guess, Wilson's equa-
tion in this example, it takes us to a region reasonably
close to the solution so that Powell's method can take
over efficiently. The straightforward use of Newton's
method, on the other hand, does not always converge
with a poor initial guess.
The convergence characteristics of Powell's method
are illustrated in Fig. 5 atp=4,875 and 4,750 psia (33.6
and 32.8 MPa). The first 15 FE's in Powell's method
were used to compute numerically the Jacobian, so only
the characteristics after the 16th FE are shown. At both
pressures, the initial step was steepest descent (0=0). At
the lower pressure [4,750 psia (32.8 MPa)], 0 increased
rapidly to unity (quasi-Newton step) and stayed there for
the remaining iterations. At the higher pressure, which is
closer to the saturation pressure, oscillations in the value
of 0 were observed. This type behavior occurs frequently
for difficult cases.
Conclusions
This paper presents a modification to the conventional
SS method that allows the detection of the single-phase
region without computing the saturation pressure and a
combination of the SS method and Powell's method that
yields a reliable and efficient algorithm for flash calcula-
tions. Criteria for an efficient switch from SS to Powell's
method are given. The success of the combination
method is due to the stability of SS and the robustness of
Powell's method. The algorithm is quite flexible, and
any other quasi-Newton method or Newton's method
may also be used instead of Powell's method. 16
Although the iteration scheme proposed was tested using
the SRK and PR EOS's, it is applicable to any other
EOS.
SOCIETY OF PETROLEUM ENGINEERS JOURNAL
 TABLE 4-RESERVOIR OIL COMPOSITION TABLE 5-FLASH CALCULATIONS FOR THE RESERVOIR OIL

Mole Pressure SS Plus Powell Newton

Component Fraction
0.00808
(psia) _'_v_ SSI Plus FE SSI Plus FE Iterations
N2 4,950 0.0116 22 + 111 2+55 2
C1 0.56990 4,925 0.1144 21 + 56 2+53 4
CO 2 0.01501 4,900 0.1770 21 + 49 2+30 4
C2 0.12714 4,875 0.2206 20+ 33 2+28 3
C3 0.07596 4,850 0.2536 20+ 30 2+27 3
iC 4 0.01075 4,800 0.3016 19+ 28 2+27 4
nC 4 0.03055 4,750 0.3358 18+ 26 2+25 4
iC s 0.00849 4,700 0.3623 18+ 26 2+26 3
nC s 0.01074 4,650 0.3838 17+ 25 2+25 3
Cs 0.01272 4,600 0.4019 16+ 25 2+24 3
180 BP' 0.01716 4,550 0.4175 16+ 25 2+24 3
230 BP 0.01764 4,500 0.4313 15+ 25 2+24 3
280 BP 0.01416
330 BP 0.01115
380 BP 0.07055
• BP represents boiling-point fractions of the c7•.

Nomenclature -2
,{O 4875 psio 8 {e 4875 psio
1.2
A = parameter in cubic EOS fIR, -1) t:J. 4750 psio ... 4750 PSiO

B = parameter in cubic EOS

1.0
fiL = fugacity of component i in liquid phase -4
fL = mole fraction of liquid phase
fv = mole fraction of vapor phase
C\I 0.8
Fi = fugacity ratio of component i, Fi =fiLlfiV ~
.,....
I -6
G = Gibbs free energy
LL
GiL = partial molar Gibbs free energy of compo-
nent i in liquid phase "".- ~
0.6 (t)

H = inverse of Jacobian matrix c> -8

0
J = Jacobian matrix 0.4
Ki = equilibrium ratio, Ki =Y/Xi
nc = number of components
-10
niL = moles of component i in liquid phase per 0.2
mole of feed, niL =NiL IN
N = total number of moles in feed
Ni = moles of component i in feed ~2~~~~~~~~~~~~~~~~0

NiL = moles of component i in liquid phase 16 20 24 28 32 36

No. of FE
p = pressure, psia (kPa)
R = gas constant Fig. 5-Convergence characteristics of Powell's method.

T = absolute temperature
v = molar volume
Xi = mole fraction of component i in liquid Subscripts
phase c = critical
Y i = mole fraction of component i in vapor phase i,j,k = component number
z = compressibility factor £ = subscript of largest equilibrium ratio
Zi = global mole fraction of component i L = liquid
a = Newton's step r = reduced
~ = steepest -descent step s = subscript of smallest equilibrium ratio
o = correction vector V = vapor
oij = interaction coefficient between components i
and j Superscripts
d = maximum norm allowed for 0 (k) = iteration level k
E = tolerance
T = transpose
~ i = iteration variables for Powell's method
- = results from SS method
rf> i = fugacity coefficient of component i in liquid Acknowledgments
phase This research is supported by the general members of
Wi = acentric factor of component i Computer Modelling Group and the Alberta/Canada
JUNE 1983 527
Energy Resources Research Fund administered by the R2T2c
Dept. of Energy and Natural Resources of the Province a(Tc) =0.42747-_-
of Alberta. We thank R.A. Heidemann and P.K.W. Vin- Pc
some for their helpful comments on this work. The ver-
sions of the SRK and PR equations used are from the
Hyprotech fluid properties program, HYPROP.
and
References
1. Soave, G.: "Equilibrium Constants from a Modified Redlich- m =0.480 + 1.574w -0.176w 2 .
Kwong Equation of State," Chern. Eng. Sci. (1972) 27,
1197-1203. PR
2. Peng, D.Y. and Robinson, D.B.: "A New Two-Constant Equa-
tion of State," Ind. Eng. Chern. Fund. (1976) 59-64. RT a(T)
3. Fussell, D.D. and Yanosik, J.L.: "An Iterative Sequence for p=---
Phase-Equilibria Calculations Incorporating the Redlich-Kwong v-b v(v+b)+b(v-b)
Equation of State," Soc. Pet. Eng. J. (June 1978) 173-82.
4. Powell, M.J.D.: "A Hybrid Method for Nonlinear Equations," RTc
Numerical Methods for Nonlinear Algebraic Equations, P.
b=0.07780-
Rabinowitz (ed.), Gordon and Breach, London (1970). Pc
5. Sandler, S.I.: Chemical and Engineering Thermodynamics, John
Wiley & Sons Inc., New York City (1977) 300-04.
6. Broyden, C.G.: "Quasi-Newton Methods and their Application to
Function Minimization," Math. Compo (1967) 368-81.
7. Wilson, G.: "A Modified Redlich-Kwong Equation of State, Ap- R2Tl
plication to General Physical Data Calculations," paper 15C a(Tc )=0.45724--
presented at the 1968 AIChE National Meeting, Cleveland, OH, Pc
May 4-7.
8. Jarrat, P.: "A Review of Methods for Solving Non-linear
Algebraic Equations in One Variable," Numerical Methods for
Nonlinear Algebraic Equations, P. Rabinowitz (ed.), Gordon and
Breach, London (1970). and
9. Nghiem, L.X. and Heidemann, R.A.: "General Acceleration Pro-
cedure for Multiphase Flash Calculation with Application to Oil- m=0.37464+ 1.54226w-0.26992w 2 .
Gas-Water Systems," paper presented at the 1982 European Sym-
posium on Enhanced Oil Recovery, Paris, Nov. 8-10.
10. Michelsen, M.L.: "The Isothermal Flash Problem, Parts I and If the compressibility factor, Z (=pvIR1), is intro-
II," Fluid Phase Equilibria (1982) 1-40. duced, the following cubic equations are obtained. For
11. Baker, L.E., Pierce, A.C., and Luks, K.D.: "Gibbs Energy SRK,
Analysis of Phase Equilibria," Soc. Pet. Eng. J. (Oct. 1982)
731-42.
12. Broyden, C.G.: "A Class of Methods for Solving Nonlinear Z3 -Z2 +(A -B-B 2)z-AB=0. . .......... (A-1a)
Simultaneous Equations," Math. Camp. (1965) 577-93.
13. Peng, D.Y. and Robinson, D.B.: "A Rigorous Method for Pre- For PR,
dicting the Critical Properties of Multicomponent Systems from an
Equation of State," AIChE J. (1977) 23, 137-44.
14. Bergman, D.F.: "Predicting the Phase Behavior of Natural Gas in
Z3 _(1-B)Z2 +(A -3B 2 -2B)z-(AB-B 2 -B3)=0 ,
Pipelines," PhD dissertation, U. of Michigan, Ann Arbor (1976).
15. Katz, D.L. and Firoozabadi, A.: "Predicting Phase Behavior of
. ........................ (A-1b)
Condensate/Crude-Oil Systems Using Methane Interaction Coeffi-
cients," J. Pet. Tech. (Nov. 1978) 1649-55.
16. Young, L.C. and Stephenson, R.E.: "A Generalized Composi- where A=apIR 2 T 2 and B=bpIRT.
tional Approach for Reservoir Simulation," paper SPE 10516 Eqs. A-1 a and A-1 b yield one or three real roots
presented at the 1982 SPE Symposium on Reservoir Simulation, depending on the number of phases in the system. In the
New Orleans, Feb. 1-3. two-phase region, the largest root is the compressibility
factor of the vapor, while the smallest positive root cor-
APPENDIX A responds to that of the liquid.
For mixtures, the following mixing rules are
The SRK and PR EOS's recommended. 2
The details of the development of the SRK and PR
EOS's are given in Refs. 1 and 2. The final results are b= 2.'>ibi, ......................... (A-2)
the following.

SRK a= ~ ~XiXjaij , .................... (A-3)

i
RT a(T)
P=---
v-b v(v+b) and

RTc
b=0.08664-
Pc
where 0 ij are binary interaction coefficients obtained by
fitting the predicted binary bubblepoint pressures to ex-
perimental data.
528 SOCIETY OF PETROLEUM ENGINEERS JOURNAL
 The fugacity coefficient of component k in a mixture and
can be calculated from the following equations.
2: Zi= 2:Xi= 2:Yi=1 .
i i
.......................... (B-8)

APPENDIX C
Powell's Hybrid Method
Consider the problems of finding the root x* of a system
of n nonlinear equations in n unknowns .
. . . . . . . . . . . . . . . . . . . . . . (A-5a)
f(x)=O ................................ (C-l)

given a starting iterate x (0) .

For PR,
Various methods have been studied and shown to be
bk quadratically or superlinearly convergent if x (0) is close
In ¢k=-(z-I)-ln(z-B)
b enough to x*. For many problems, however, a good in-

~ [2
__2~B
_ 2:,. Xiaik itial estimate is not available and therefore a more robust
algorithm is needed.
a - :kJ Two well-known methods for solving Eq. C-l are
Newton's method and the steepest-descent method .
Newton's method is based on a linear approximation
. In(Z+2.414B) , .............. (A-5b)
z-0.414B
of f around the kth iterate x(k) .

f[X(k) +a(k)] =f[x(k)] +J(k)a(k) ,

................................ (C-2)
where
APPENDIX B
Mole Fractions Definitions J(k)= -
af )(k)
( ax
Let
is the Jacobian matrix of f at x(k).
The Newton's step, a(k) =x(k + I) -x(k), is obtained
by setting the left side of Eq. C-2 to 0 and solving for
a(k) .
The mole fractions are defined as follows.
Global mole fraction of component i, a(k) = -J -I(k)f(k) , ..................... (C-3)

Zi =N;lN. . ............................ (B-1) where f (k) denotes f [x (k)] for convenience. Newton's
method is quadratically convergent if x (k) is close to x*.
Liquid mole fraction of component i, If x(k) is not close to x*, there is no guarantee that
x(k+ I) will be any better than x (k) and Newton's method
xi=NiLINL . ........................... (B-2) need not converge. A method that guarantees that

Vapor mole fraction of component i,

is the steepest-descent method where 11'11 denotes the
Yi =NiVIN v . ........................... (B-3)
Euclidian norm.
Consider the functional equation
Liquid mole fraction of the system,
F(x) = vzr T (x )f(x). . .................... (C-4)
fL =NLIN. . ............................ (B-4)
The steepest-descent direction g is
Vapor mole fraction of the system,
g= - \l F(x) = -JTf(x). . ................ (C-5)
fv=NvIN. . ........................... (B-5)
Given the kth iterate x (k) , the steepest -descent step p(k)
= x(k+I) - x(k) is defined by
From these definitions we obtain
p(k) =J1.(k)g(k) , •...........••.•..•...... (C-6)
Zi =fLXi +fVYi, ...................... (B-6)
with J1. (k) chosen so that the functional C-4 with f(x)
fL = f v = 1 , ........................... (B-7) replaced by f(k) + J(k)p(k) is minimized.
JUNE 1983 529
 It can be shown that the value of p, (k) that minimizes H(k+I) =H(k)

F[X(k)] = 1/2 [f(k) + p, (k) J(k)p(k)] T [&(k+ 1) _ H(k)[f(k+ I) -f(k)Jj& T(k)Hlk)

+------~~~--~--~-------
&T(k)H(k)[f(k+I) _flk)]
. [f(k) +p,(k)J(k)P(k)] ............ (C-7) ........................ (C-ll)

is given by A method using this updating scheme is called quasi-

Newton, and the final convergence is superlinear rather
than quadratic .
. . . . . . . . . . . . . . . . . . . . . (C-8)
e
The computation of in Eq. C-9 is given in the follow-
ing. Let ~ be the maximum norm of the correction vector
Although the steepest-descent method avoids &. Therefore, if
divergence, it converges very slowly near the solution.
Powell's hybrid method is a combination of a Newton- Ila(k)II~~, ........................... (C-12)
like method and the steepest-descent method. It has the
divergence-avoiding characteristic of steepest-descent e(k) is set to unity (a quasi-Newton step). Otherwise a
method and the rapidly converging property of the multiple of p is included in & in order to satisfy
Newton-like method near the solution.
Powell's algorithm computes a scalar e(k), o~e(k) Ilea+(l-e)PII=~· ..................... (C-13)
~ 1, that is used to define the step &(k) =x(k+ 1) -x(k) as
follows. Straightforward algebra gives for (j the expression

e=~ 2-IIPI1 2/[ra-p) Tp + [(a Tp _~2)2

+(llaI1 2_~2)(~2 -IIPI12)]'12). . ....... (C-14)
where elk) = 1 corresponds to the Newton-like method
and elk) =0 corresponds to the steepest-descent method. The ~ = 5 in all examples.
The Jacobian matrix J and its inverse H are computed
numerically at the beginning of the iteration and updated SI Metric Conversion Factors
at each step using rank-one matrices according to OF (OF-32)/1.8 °C
Broyden. I2
psi x 6.894 757 E+OO = kPa
f(k+I) _f(k) _J(k)&(k)&T(k)
J(k+I) =J(k) + [
. * Conversion factor is exact SPEJ
&T(k)&(k) Original manuscript received in Society of Petroleum Engineers office July 31,1979
Paper accepted for publication Aug. 19, 1982. Revised manuscripl received Feb. 7,
1983. Paper (SPE 8285) first presented at the 1979 SPE Annual Technical Can·
........................ (C-IO) ference and Exhibition held in Las Vegas, Sept. 23-26.

530 SOCIETY OF PETROLEUM ENGINEERS JOURNAL