Vous êtes sur la page 1sur 15

Fluid Phase Equilibria 358 (2013) 304318

Contents lists available at ScienceDirect

Fluid Phase Equilibria


journal homepage: www.elsevier.com/locate/fluid

Application of a global optimization algorithm to phase stability


and liquidliquid equilibrium calculations
Stanislaw K. Wasylkiewicz a, , Yau Kun Li a , Marco A. Satyro a,b ,
Monika J. Wasylkiewicz a,b,1
a
Virtual Materials Group Inc., Calgary, Alberta, Canada
b
The Schulich School of Engineering, University of Calgary, Calgary, Alberta, Canada

a r t i c l e i n f o a b s t r a c t

Article history: A global optimization algorithm was developed for nding all singular points (minima, maxima and
Received 29 May 2013 saddles) of an objective function by exploring the natural connectedness that exists between their singular
Received in revised form 19 August 2013 points. The idea of following ridges and valleys using information gathered along the way was signicantly
Accepted 25 August 2013
enhanced by applying the arc length continuation method. The algorithm was applied to the Gibbs tangent
Available online 8 September 2013
plane stability test for multiphase liquid mixtures. The global optimization algorithm gives an efcient
and robust scheme for locating all stationary points of the tangent plane distance function predicted by
Keywords:
any thermodynamic model. Since it provides very good initial estimates for the liquidliquid equilibrium
Global optimization
Gibbs tangent plane stability test
calculations, it became an integral part of a combined phase equilibrium and stability algorithm. The
Liquidliquid equilibrium combined algorithm is self-starting and signicantly improves reliability and robustness of multiphase
Arc length continuation equilibrium calculations. It was successfully tested on a variety of problems and is applicable to any
components mixtures and any number of liquid phases. Solutions have been found for the entire phase
diagram.
2013 Elsevier B.V. All rights reserved.

1. Introduction various methods for nding one or more solutions resulting from
model equations, but they do not guarantee that the solution corre-
Geometric methods based on residue curve map (RCM) technol- sponds to a tangible physical criterion such as minimum free energy
ogy are efcient tools for distillation process synthesis and design or that all solutions have been found. This can be of paramount
[1,2]. In modern process simulators, designers need tools for gen- importance for some modeling problems of physical systems. For
eration of RCMs and liquidliquid equilibrium (LLE) diagrams to example in liquid phase stability analysis, we have to nd all sta-
help visualize the phase behavior from the global thermodynamic tionary points of a tangent plane to the Gibbs free energy surface
space perspective. Unfortunately, most of these tools fail if more to determine whether the examined liquid is stable or not.
than two liquid phases are present at equilibrium [3]. Even if such In this paper, we describe a general approach for nding all
phenomena are not observable in reality, it is important to cal- singular points of mathematical models by exploring the natural
culate a stable LLE solution, as predicted by the thermodynamic connectedness that exists between singular points of an objective
model. Incorrect single liquid solutions or unstable liquid phases function. The idea of following ridges and valleys using informa-
may result in program failures or erroneous results. To nd a glob- tion gathered along the way is not new and was used in many of the
ally stable equilibrium solution, a stability test in terms of the hill climbing or direct search optimization techniques developed in
tangent plane criterion [48] is commonly performed using global the 1960s [18]. However, since these optimization techniques do
optimization techniques [917]. not use derivative information to determine the search direction,
Finding all solutions of mathematical models is still a challeng- they have difculties tracking strong curvature and are not useful
ing problem in many areas of science and engineering. There are for large-scale problems. They are generally considered to be less
efcient than derivative-based methods.
Van Dongen and Doherty [18] used ideas from differential
geometry and the theory of differential equations to describe the
Corresponding author. Present address: Computer Modelling Group Ltd., 150, dynamics of distillation processes. They dened a ridge on a topo-
3553 31 Street NW, Calgary, Alberta T2L 2K7, Canada. Tel.: +1 403 531 1307. graphic map of a boiling temperature surface that contains contour
E-mail addresses: stan.wasylkiewicz@cmgl.ca, stan.wasylkiewicz@gmail.com
lines for different temperature values as a location where the con-
(S.K. Wasylkiewicz).
1
Current address: Department of Chemical and Materials Engineering, University
tour lines exhibit maximum curvature. They gave also another
of Alberta, Edmonton, Alberta, Canada. denition for ridges (valleys) as maxima (minima) of an objective

0378-3812/$ see front matter 2013 Elsevier B.V. All rights reserved.
http://dx.doi.org/10.1016/j.uid.2013.08.030
S.K. Wasylkiewicz et al. / Fluid Phase Equilibria 358 (2013) 304318 305

its ridges and valleys. They conjectured that the ridges and val-
List of symbols leys of F(x) are all connected to each other and that this connected
domain contains all the stationary points. This was later proved up
aij , bij binary interaction parameter in the VLE model, K or to R3 by Cheng-Dong and Xiao-Xia [20].
cal/mol Valleys and ridges of the least-squares landscape are specic
c contour line value of the potential function, F(x) integral curves of the gradient vector eld and have been dened
F(x) tangent plane distance function by Lucia and Feng [21] as a collection of constrained extrema over
F( x) gradient of the potential function F(x) a set of level curves. Based on this denition they developed a so
G(x, , t, s) augmented homotopy function, Eq. (19) called global terrain method, which consists of a series of down-
g(x) molar Gibbs free energy, J/mol hill, equation-solving computations and uphill, predictorcorrector
g(x, , t, s) pseudo arc length equation, Eq. (20) calculations. They noticed that the nonlinear optimization prob-
H Hessian matrix lems used to dene valleys and ridges can be quite difcult to solve
Hj (2 , . . ., p ) RachfordRice equation, Eq. (A.5) because they are inherently poorly scaled.
J Jacobian matrix To overcome the difculties with tracking of ridges and valleys
K liquidliquid equilibrium constant [13,21], we developed a new arc length continuation method that is
L(x, z) tangent plane to the Gibbs energy surface at com- based on a method originally developed by Wasylkiewicz et al. [22]
position z to nd all homogeneous and heterogeneous azeotropes predicted
L( x, ) Lagrangian function for the optimization problem, by a thermodynamic model. First we created the Lagrangian func-
Eq. (17) tion for the optimization problem and applied the KuhnTucker
N dimension of the optimization problem, N = Nc 1 conditions. Then we added an additional equation dening the arc
Nc number of components length to the system of original equations to form an augmented
Nmax number of maxima system. During branch tracking we look for bifurcation points, from
Nmin number of minima which additional branches are later started in eigendirections [23].
Np number of liquid phases We applied this new global optimization algorithm to the Gibbs
Nsad even number of even saddles tangent plane stability test for multiphase liquid mixtures. The
Nsad odd number of odd saddles algorithm relies on a new arc length continuation method that gives
Ntng number of tangent nodes a robust scheme for locating all stationary points of the tangent
MAX maximum plane distance function predicted by a thermodynamic model. With
MIN minimum all stationary points located the corresponding Gibbs free ener-
R gas constant, J/mol K gies can be calculated and it is a simple matter to determine the
SA saddle point that corresponds to the most stable solution. The algorithm
s arc length can be applied to any components mixtures and any number of
T temperature, K liquid phases, is self-starting and signicantly improves reliability
t continuation parameter and robustness of multiphase equilibrium calculations. The devel-
p pressure, kPa oped technique was successfully tested on a variety of problems. In
x mole fractions of the liquid phase this article we show examples of three different mixtures, ranging
z overall composition of the liquid mixture from a ternary system with three liquid phases to a ve component
system with two liquid phases. LLE and VLLE solutions have been
Greek symbols
successfully found for the entire phase diagrams.
ij binary interaction parameter in NRTL model
i activity coefcient of component i
ij Kronecker delta function 2. Tangent plane stability test
 tuning factor in pseudo arc length equation
 Lagrange multiplier Gibbs [4,5] showed that, at a xed temperature T and pressure p,
i chemical potential of component i, J/mol a necessary and sufcient condition for absolute stability of a mix-
 time ture of overall composition z is that the Gibbs free energy surface
liquid phase fraction g(x) be at no point below the plane L(x, z) tangent to the surface at
the composition z. This can be expressed mathematically as follows
Subscripts
i component index
b derivative computed at the bifurcation point for the F(x) g(x) L(x, z)0 x (1)
old branch
0 derivative computed at the bifurcation point for the The tangent plane distance function F(x) is the distance from the
new branch Gibbs free energy surface to the tangent plane at z, evaluated at a
composition x.
Superscripts
The necessary and sufcient condition for phase stability is given
k liquid phase index
by theorems 3 and 4 of Baker et al. [6]. The mixture is stable if and
T transposed vector or matrix
only if the tangent plane distance function F(x) is non-negative at
all stationary points.
The molar Gibbs free energy of a mixture of overall composi-
function F(x) taken in the constraint plane normal to the ridge-top tion z can be expressed in terms of chemical potentials i ( z) of all
locus. The second denition, together with gradient system prop- components as follows
erties, was used by Wasylkiewicz et al. [13,19] to develop a robust
global algorithm for implementation of the Gibbs tangent plane 
Nc

analysis for complex liquid mixtures. The algorithm locates all the g(z) = zi i (z) (3)
stationary points of the tangent plane distance function by tracking i=1
306 S.K. Wasylkiewicz et al. / Fluid Phase Equilibria 358 (2013) 304318

The equation of the plane tangent at z to the Gibbs energy surface yields the stationary conditions, Eq. (12), written assuming the use
is given by of an activity coefcient model.
  g(x) 
Nc 1
F(x)
L(x, z) = g(z) + (xi zi ) (4) = j (x) Nc (x) j (z) + Nc (z)
xi xj
i=1 x=z
 
By applying the BakhuisRooseboom equation [24] xj j (x) zNc Nc (z)
= RT ln =0 (12)
xNc Nc (x) zj j (z)
g(x)
= i (x) Nc (x) (5)
xi
j = 1, . . ., Nc 1
We have

Nc

Nc 1

Nc The linear term L(x, z) in Eq. (1) does not inuence the second
L(x, z) = zi i (z) + [i (z) Nc (z)] (xi zi ) = zi i (z) derivative of F(x). Therefore, the Hessian matrix of F(x) and g(x) are
i=1 i=1 i=1
identical [25], Eq. (13):
   

Nc 1

Nc 1

Nc F(x) i (x) Nc (x)
+ i (z)(xi zi ) Nc (z) (xi zi ) = zi i (z) J [ F(x)] = H [F(x)] = = = H [g(x)]
xi xj xj xj
i=1 i=1 i=1 (13)


Nc 1 Nc

+ i (z)(xi zi ) + Nc (z)(xNc zNc ) = xi i (z) (6)   


F(x) xi i (x)
i=1 i=1 = RT ln
xi xj xj xNc Nc (x)
Finally, the expression for the tangent plane distance function  
ij ln i (x) 1 ln Nc (x)
in terms of chemical potential can be written as Eq. (7). = RT + + (14)
xi xj xNc xj

Nc

F(x) = xi [i (x) i (z)] (7) ij is the Kronecker delta function. The partial derivatives of
i=1 activity coefcients ln j (x)/xj in Eq. (9) are calculated analyti-
where the chemical potential of the i-th component of the mixture cally or numerically for any suitable activity coefcient model (e.g.
i ( z) is calculated at the overall composition z and i ( x) at any NRTL, UNIQUAC, etc.) using a routine [26] specially designed to give
point from the composition space, Eq. (8). smooth derivatives close to composition boundaries and compo-
Nc nent vertices.
xi 0 i = 1, . . ., Nc , xi = 1 (8)
i=1
3. Global optimization in process design
Michelsen [7,8] noted that all of the minima in F(x) are located
in the interior of the permissible region. Therefore, F(x) will be Optimization methods are now routinely applied at each level
non-negative if it is non-negative at all stationary points where of system and process design. Optimization is also progressively
its derivatives with respect to all of the independent variables (xi becoming more important in plant operations. Modern mathemat-
i = 1, . . ., Nc 1) vanish. The mole fraction of component c can be ical optimization methods can deliver adequate globally optimal
calculated from Eq. (8) and substituted into Eq. (7) giving to the solutions [27] if the proper method is selected for a particular prob-
following expression for the tangent plane distance function, Eq. lem.
(9). The objective of global optimization is to nd the best solution of
  highly nonlinear models that may have a number of locally optimal

Nc 1 
Nc 1
F(x) = xi [i (x) i (z)] + 1 xi [Nc (x) Nc (z)] solutions. For some optimization problems, global optimization can
be extremely difcult and frequently engineers are forced to set-
i=1 i=1
tle for solutions that are good enough because nding the best

Nc 1
solution could be prohibitively expensive.
= xi [i (x) Nc (x) i (z) + Nc (z)] + Nc (x) Nc (z)
Finding an arbitrary local minimum is relatively straightforward
i=1
by using classical local minimization methods. Finding the global
(9)
minimum of a function that has a large number of local minima and
maxima is far more difcult. Numerical solution strategies used for
Differentiation of F(x) with respect to the Nc 1 independent global optimization [28] can be grouped into the following three
mole fractions gives Eq. (10). categories:
 Nc 1  
F(x) i (x) Nc (x) Nc (x)
= j (x) Nc (x) j (z) + Nc (z) + xi + = j (x) Nc (x) j (z) + Nc (z)
xj xj xj xj
i=1
 
  (x)
Nc 1 
Nc 1
Nc (x)   (x) Nc
i i
+ xi + 1 xi = j (x) Nc (x) j (z) + Nc (z) + xi j = 1, . . ., Nc 1 (10)
xj xj xj
i=1 i=1 i=1

along with the use of the GibbsDuhem equation at constant tem-


perature and pressure, Eq. (11),


Nc   Deterministic methods (e.g. inner and outer approximation, cut-
i (x) ting plane methods, branch and bound methods [12], interval
xi =0 (11)
xj methods [17], homotopy continuation [16], methods based on
i=1 T,p,x
differential geometry [13,21]).
S.K. Wasylkiewicz et al. / Fluid Phase Equilibria 358 (2013) 304318 307

Stochastic methods (e.g. simulated annealing [29], Monte-Carlo Trajectories of Eq. (15) converge toward stable manifolds and
sampling, stochastic tunneling [30], parallel tempering). diverge from unstable manifolds.
Heuristic strategies that search the optimization domain in a The ow of trajectories is always downhill (i.e., F(x) decreases
more or less intelligent way (e.g. genetic algorithms [31,32], evo- along every solution curve) and is perpendicular to the level
lution strategies, swarm-based optimization algorithms [33], ant curves of F(x).
colony optimization, reactive search optimization, differential Ridge-top and valleys-bottom trajectories are tangent to eigendi-
evolution [34], harmony search [35], cuckoo search [36]). rections of the Jacobian J[F(x)] or Hessian H[F(x)] matrix at the
stationary points.
The deterministic methods theoretically guarantee convergence There are no periodic solutions of Eq. (15).
to the global solution. Unfortunately, most current deterministic Level curves of F(x) on a topographic map that contains con-
global optimization methods are too expensive in terms of com- tour lines for different elevations exhibit maximum curvature at
putational effort even for moderate size problems. Also, for some ridges and valleys of F(x).
methods, problem reformulation is needed depending on the ther- Ridges (valleys) are maxima (minima) of the potential func-
modynamic model being used [912]. tion F(x) taken in the constraint planes normal to the ridge-top
The stochastic and heuristic global optimization methods are (valleys-bottom) locus [18].
easy to implement and use, are independent of a model and can Ridges and valleys of F(x) are all connected to each other and this
handle large number of decision variables. They can locate good connected domain contains all the stationary points [20].
solutions within a reasonable CPU time although they often face
difculties in nding the global solution [37] and do not guarantee The connectedness that exists between singular points of the
the global optimality. If it is acceptable for a particular optimization potential function is essential for our global optimization algo-
problem, these methods may offer the best compromise between rithm. We explore this connectedness by following ridges and
quality of solution and efciency for multivariable optimization. valleys of the potential function.
Especially if calculations of a single point of an objective function is
very time consuming (e.g. in reservoir modeling and optimization 5. Ridges and valleys of a potential function
[38]), stochastic or heuristic global optimization methods are the
only reasonable choice. On the other hand, if the objective function Ridges and valleys of a potential function F(x) are specic inte-
calculation is quick and partial derivates can be easily calculated gral curves of the gradient vector eld and have been dened by Van
analytically or numerically, the more reliable deterministic meth- Dongen and Doherty [18] as a locus on a topographic map where the
ods should be used. contour lines for different elevations exhibit maximum curvature.
Heuristic global optimization methods have been frequently Later, Lucia and Feng [21] dened ridges and valleys as a collection
used for solving parameter estimation problems in phase equi- of constrained extrema over a set of level curves. Mathematically,
librium modeling [34]. Using locally instead of globally optimal these denitions can be expressed as the following constrained
parameters for thermodynamic models are potential sources of optimization problem [21,23], Eq. (16):
problems for design of separation processes, resulting in errors and

uncertainties in equipment design and erroneous model perfor- opt F(x)T F(x) (16)
x
mance [39,40].
Failure to nd the globally optimal solutions in phase stability s.t. F(x) = c
calculations can have even more serious consequences in subse-
where F(x) is the gradient of the potential function F(x) at point
quent process calculations. Using an erroneous VLE solution instead
x and c is any given value (or level, or elevation) of the potential
of a correct VLLE one, or vice versa, or switching between them dur-
function F(x). For any given contour line at a specied value of c,
ing distillation column simulation, can cause convergence failure
we can nd a point on the line that corresponds to a local opti-
in the column or give erroneous conclusions about simulated plant
mum in the gradient norm by solving the constrained optimization
performance.
problem described by Eq. (16). The collection of these optima for
all level values contains all of the ridges-tops and valleys-bottoms.
4. Stationary points of a potential function
Cheng-Dong and Xiao-Xia [20] showed that all stationary points of
a potential function F(x) are connected by a network of ridges and
The tangent plane distance function (Eq. (7)) is a potential func-
valleys. Therefore we can quickly nd all the stationary points by
tion. Some properties of this type of functions are essential for our
exploring the problem domain not by random guesses but rather by
global optimization algorithm and are summarized in this para-
systematic movement along ridges and valleys. This is an important
graph.
advantage viz.-a-viz. traditional phase stability algorithms.
A gradient system in RN is an autonomous ordinary differential
By applying the KuhnTucker conditions [41] to the Lagrangian
equation, Eq. (15):
function L(x,) for the optimization problem described by Eq. (16)
dx
= F(x) (15) L(x, ) = F(x)T F(x) +  [F(x) c] (17)
d
dened by the gradient F( x) of a potential function F(x) where we can get Eq. (18)
F(x): RN R and F(x) C .
Lx (x, ) = 2H[F(x)]T F(x) +  F(x) = 0 (18)
The gradient system has the following useful properties:
L (x, ) = F(x) c = 0
The singular points of Eq. (15) occur at stationary points in F(x)
and vice versa. Eq. (18) contains N + 1 unknowns (independent variables x1 ,
The Jacobian is a symmetric matrix and has all real eigenvalues. x2 ,..., xN , and Lagrange multiplier ) and one parameter (contour
Unstable nodes occur at maxima in F(x), stable nodes occur at line value c). A current branch that follows valley-bottoms or
minima in F(x), and saddle points occur at saddles in F(x). ridge-tops can be tracked by gradually increasing or decreasing
Ridges in F(x) correspond to the unstable manifolds of Eq. (15), parameter c and solving Eq. (18). The contour line value is then an
and valleys correspond to the stable manifolds. ideal continuation parameter for this global optimization problem.
308 S.K. Wasylkiewicz et al. / Fluid Phase Equilibria 358 (2013) 304318

Various algorithms were previously used [13,18,21] to follow where the subscript b indicates derivatives computed at the bifur-
ridges and valleys of the potential function. To enhance reliabil- cation point for the old branch and the subscript 0 for the new
ity and efciency of the algorithm we developed a new arc length branch. The top left part of the above matrix can be calculated [23]
continuation method described in details in the next paragraph. from the following equation, Eq. (22).
 
6. Tracking continuation branches Lxx (x, , t) = 2J H[F(x)]T F(x) + H [F(x)] (22)

Efcient tracking of continuation branches is crucial for the suc- The Jacobian matrix in Eq. (22) is calculated numerically using
cess of the global optimization method. One can attempt to track a routine [43] specially designed to give smooth derivatives close
the branches of Eq. (18) starting from any stationary point of F(x), to the composition boundaries and component vertices. Instead of
gradually increasing or decreasing the value of the continuation switching suddenly from the forward to the backward difference
parameter t and using the most recently calculated point as an ini- formula just before independent variable reaches the boundary, we
tial guess for the next one on the branch. However, this type of use a difference formula that does this gradually. Since very close
parameter continuation can fail at turning points or other singular- to a vertex this difference formula cannot be applied, we choose
ities where the continuation parameter must start to decrease or two points close to the vertex, calculate two derivatives and then
increase to follow the branch. To resolve these problems, a mod- extrapolate them to the original point.
ied arc length continuation was used [22], where an additional The algorithm for tracking of continuation branches to nd all
equation g(x, , t, s) dening the arc length has been added to the stationary points inside the problem domain is summarized as fol-
system of Eq. (18) to form the following augmented system, Eq. lows.
(19).
1. The algorithm starts from any stationary point inside the
Lx (x, , t)
problem domain. For the tangent plane distance function, the

G(x, , t, s) = L (x, , t) =0 (19) composition of the original mixture z is usually selected as the
starting point for the algorithm since at this point F(z) = 0 and
g(x, , t, s)
F(z) = 0.
Eq. (19) contains N + 2 unknowns (x1 , x2 ,..., xN , , t) and one 2. Calculate the Hessian matrix H[F(x)] at the stationary point, and
parameter, the arc length (s). The current branch can be tracked then its eigenvalues and eigenvectors.
by gradually increasing parameter s and solving Eq. (19). Since the 3. Add new stationary point to the set of found stationary points
continuation parameter t is one of the unknowns it can increase or (SFSP).
decrease, while the arc length always increases. 4. Add new stationary point coordinates and corresponding eigen-
Since the arc length is a nonlinear function of the independent vectors to the set of initial conditions (SIC) for subsequent
variables [22], for the arc length relationship g(x, , t, s), we adopted branches to be started from.
the pseudo arc length equation dened by Keller [42] 5. Take one point from SIC and start a new branch in one of the
  eigendirections. Follow the branch up to a new stationary point
N dxi (sk ) d(sk )
g(x, , t, s) =  [xi xi (sk )] + [ (sk )] or to a boundary of the domain. Check for bifurcations along the
i=1 ds ds
continuation branch. If a bifurcation has been detected, add the
  dt(sk ) bifurcation point coordinates and corresponding eigenvectors to
+ 1  [t t(sk )] (s sk ) (20) SIC for later exploration.
ds
6. If a newly found stationary point is not in SFSP, go to step 2.
where  is a tuning factor selected close to 0.99 to place more
emphasis on the original problem independent variables than on 7. If there are some unused initial conditions in SIC, go to step 5.
the continuation parameter. 8. Check the topological criterion (see Appendix B) for all stationary
xi (sk ), (sk ) and t(sk ) are known from the previously calculated points from SFSP.
point k for an arc length sk . The vector tangent to the continuation
path at this point [dxT (sk )/ds, d(sk )/ds, dt(sk )/ds] can be approxi- The reliability and efciency of the algorithm is crucial for the
mated by the secant vector connecting points k 1 and k. Therefore success of the search for all stationary points of the tangent plane
Eq. (20) is linear with regard to the independent variables xi and  distance function F(x). To speed up the branch tracking, at rst,
and continuation parameter t. This secant vector is used in a pre- we explore only directions that correspond to dominant geomet-
dictor step during branch tracking. In a corrector step, we solve Eq. ric distortion on the objective function [21]. For uphill movement,
(19) by the NewtonRaphson method. We optimize the Newton it corresponds to the smallest positive eigenvalue of the Hessian
step length and take the problem constraints into account. matrix, for downhill movement to the most negative eigenvalue.
The arc step size (s sk ) is set initially to 0.01 and then reduced If after exploring all dominant geometric distortion directions the
for strongly nonlinear paths or enlarged for nearly linear ones far topological criterion is not fullled, other directions from the set of
from bifurcation points. While tracking the continuation branches, initial conditions are explored.
we monitor the determinant of the Jacobian matrix of Eq. (19). If
it changes sign between two consecutive points on the branch, we 7. Combined phase equilibrium and stability algorithm
calculate a new bifurcation point [22]. The direction of the sec-
ond branch that intersects the current brunch at this bifurcation Without good initial estimates for the number of liquid phases
point is then calculated from the Kellers [42] approximation of the and their compositions, the liquidliquid equilibrium calculations
algebraic bifurcation equation can easily fail to converge to a stable solution. Also, the number
of liquid phases in equilibrium cannot be determined on the basis
dx0
Lxx (x, , t) F(x) 0 of tangent plane analysis alone. The stability test usually does not
ds
deliver the exact compositions of phases in equilibrium and some-
F(x)T 0 1 d0
=0 (21) times predicts incorrect number of phases. Fortunately, it provides
dxT ds good initial estimates for the LLE calculations. For that reasons,
b db dtb
  (1 ) dt0 the combined phase equilibrium and stability algorithm should be
ds ds ds
ds used. It is summarized as follows.
S.K. Wasylkiewicz et al. / Fluid Phase Equilibria 358 (2013) 304318 309

1. For a liquid of an overall composition z, at temperature T and


pressure p, nd all stationary points of the tangent plane dis-
tance function F(x) using algorithm described in the previous
paragraph and check the stability of the liquid. If the liquid is
globally stable, then end.
2. Otherwise, save the coordinates of all minima of F(x) and go to
step 3.
3. Use the coordinates of the minima of F(x) (up to the limit allowed
by the phase rule) as an initial guess in calculations of the liq-
uid phases in equilibrium. Make use of the rigorous LLE ash
calculation routine described in Appendix A.
4. If one or more of the phase fractions converge to zero, eliminate
the corresponding liquid phases from the solution.
5. Check the global stability of one of the liquid phases present
in the LLE calculation solution. If the calculated equilibrium is
globally stable, then end.
6. Otherwise, save the coordinates of all minima of F(x) found dur-
ing the last stability test calculations and go to step 3.

In the following paragraph, we show in details all steps of the


combined phase equilibrium and stability algorithm applied to
Fig. 1. Branches for tracking ridges and valleys of the dimensionless tangent plane
one ve-component and two ternary mixtures. All selected com- distance function F(x, z)/RT calculated for the overall composition z = [0.2, 0.4, 0.4]
positions were unstable and end up at two or three liquid phase at temperature 338.48 K.
equilibria. Notice that for the two examples of liquidliquidliquid
equilibrium the nal stable three phase equilibrium was found after
previously nding unstable two liquid phase equilibrium.

8. Examples

8.1. Ternary mixture of ethanol, water and benzene

Vaporliquidliquid equilibrium (VLLE) for the mixture of


ethanol, water and benzene at pressure 101.325 kPa was calculated
by NRTL/Ideal/Chemical model using the chemical process simula-
tor VMGSim [44]. The dimensionless interaction parameters  ij in
the NRTL model are related to the interaction energy parameters
gij and dened as  ij = gij /RT = aij + bij /T + cij ln(T). The nonzero
binary interaction parameters are shown in Table 1. They were
calculated using experimental data from [45,46].
To create a VLLE diagram, as in Fig. 4, we repeatedly applied bub-
ble point calculations to cover all possible heterogeneous regions in
the composition space. An important part of VLLE calculations is the
liquid stability test and LLE ash calculations, which are imbedded
in the search for a correct bubble point temperature for a particular
mixture. In this example, we show a comprehensive procedure for Fig. 2. Initial (LLEStrtPnt) and nal (LLE) tie lines for LLE calculations, and branches
correctly calculating the LLE for the overall composition z = [0.2, 0.4, for tracking ridges and valleys of the dimensionless tangent plane distance func-
tion F(x, z)/RT calculated to test stability of one liquid phase L1 = [0.10761, 0.03325,
0.4] at a saturation temperature of 338.48 K. We use our stability
0.85914] at temperature 338.48 K.
test several times along with rigorous LLE ash calculations to make
sure that the LLE is correct and all liquid phases in equilibrium are
stable. (see Appendix B), we conclude that all stationary points in this case
We start with the stability test for the overall composition have been found.
z = [0.2, 0.4, 0.4] as shown in Table 2 and Fig. 1. In this case the Since values of the dimensionless tangent plane distance func-
starting point for branch tracking (overall composition z) is a sad- tion F(x, z)/RT are negative at both minima (Table 2), the overall
dle point on the dimensionless tangent plane distance function composition z represents an unstable liquid. Now we switch to
F(x, z)/RT. One eigenvalue at this stationary point is positive, our rigorous LLE ash calculation routine described in Appendix
the other one is negative. Using our global optimization routine, A. From all stationary points found in the previous stability test
two other stationary points (minima) have been found in the step (Table 2) we select both minima which compositions are good
entire composition space. Since the topological criterion is satised initial guesses for LLE ash calculations (see Table 3 and Fig. 2). The

Table 1
NRTL/Ideal/Chemical binary parameters for the mixture of ethanol (1), water (2) and benzene (3). Binary interaction parameters bij are in K.

i bi1 bi2 bi3 i1 i2 i3

1 71.36832 82.33937 0.29814 0.22474


2 683.40577 1690.67004 0.29814 0.30000
3 571.40053 1357.21198 0.22474 0.30000
310 S.K. Wasylkiewicz et al. / Fluid Phase Equilibria 358 (2013) 304318

Table 2
Results of the search for all stationary points for the mixture of ethanol (1), water (2) and benzene (3). Temperature 338.48 K. NRTL/Ideal/Chemical model. z = [0.2, 0.4, 0.4].

Type Index F(x, z)/RT x1 x2 x3 Eigenvalues

SA 1 0.0 0.2 0.4 0.4 0.49665 4.85742


MIN 1 0.06755 0.05166 0.01617 0.93217 46.83048 12.7859
MIN 1 0.07999 0.05908 0.93398 0.00694 4.78161 264.0048

Topological criterion: Nmax + (1)N Nmin + Nsad even Nsad odd = (1)N is satised (N = 2)
Number of minima Nmin = 2
Number of odd saddles Nsad odd = 1

Table 3
NEWTON LLFLSH calculations for the mixture of ethanol (1), water (2) and benzene (3). Temperature 338.48 K. NRTL/Ideal/Chemical model. z = [0.2, 0.4, 0.4].

x1 x2 x3 k g(x)/RT

Overall composition z 0.2 0.4 0.4 0.26525


Starting point L1 0.05166 0.01617 0.93217
Starting point L2 0.05908 0.93398 0.00694
L1 at LLE 0.10761 0.03325 0.85914 0.38409 0.19063
L2 at LLE 0.25762 0.62871 0.11367 0.61591 0.34836
Mixture of L1 and L2 0.28778

Table 4
Results of the search for all stationary points for the mixture of ethanol (1), water (2) and benzene (3). Temperature 338.48 K. NRTL/Ideal/Chemical model. z = [0.10761,
0.03325, 0.85914].

Type Index F(x, z)/RT x1 x2 x3 Eigenvalues

MIN 1 0.0 0.10761 0.03325 0.85914 20.5528 2.98595


SA 1 7.77E-03 0.22844 0.16898 0.60258 0.54572 4.72213
MIN 1 6.44E-05 0.28444 0.54716 0.16840 0.05590 6.66389
SA 1 1.92E-08 0.25762 0.62871 0.11367 0.04733 10.5054
MIN 1 6.79E-04 0.17381 0.78856 0.03763 0.23057 39.4499

Topological criterion: Nmax + (1)N Nmin + Nsad even Nsad odd = (1)N is satised (N = 2)
Number of minima Nmin = 3
Number of odd saddles Nsad odd = 2

calculation routine converges quickly to the two-phase LLE solution mixture and then compare to the value calculated for the overall
shown in Table 3. We calculate the Gibbs free energies for both liq- composition z. Since the value for the mixture is lower than that
uid phases (L1 and L2) and their mixture, and then compare to the for the overall composition, we conclude that the liquid z will pos-
value calculated for the overall composition z. Since the value for sibly split into L1, L2 and L3. Notice that the Gibbs free energy for
the mixture is lower than that for the overall composition, we con- the two-liquid mixture in Table 3 is higher than that for the three-
clude that the liquid z will possibly split into L1 and L2. Now we liquid mixture in Table 5. Now we check if the liquids L1, L2 and L3
have to check if the liquids L1 and L2 are stable. are stable.
We perform the stability test for one liquid phase (e.g. liquid To be sure that we have found a stable three liquid phase solu-
L1 from Table 3). In this case the starting point for branch tracking tion, we choose one liquid phase (e.g. liquid L1 from Table 5) and
is a minimum on the dimensionless tangent plane distance func- perform the stability test. In this case the starting point for branch
tion (see Table 4). Using our global optimization routine, four other tracking is a minimum on the dimensionless tangent plane distance
stationary points (minima and saddle points) have been found in function (see Table 6 and Fig. 3). Using our global optimization rou-
the whole composition space (Fig. 2). Since values of F(x, z)/RT are tine, four other stationary points have been found in the whole
negative at both minima, liquid L1 is unstable. composition space. Since values of F(x, z)/RT are non-negative
Now we switch again to our rigorous LLE ash calculation rou- (within assumed tolerance 1.0E-06) at all stationary points, liquid
tine. The compositions of three minima found in the previous L1 is stable. There is no need to check stability of the other two liquid
stability test step (Table 4) are very good initial guesses for LLE ash phases in equilibrium (L2 and L3 from Table 5) since the points rep-
calculations (see Table 5). Now the calculation routine converges resenting these liquids must have a common tangent plane to the
to the three-phase LLE solution shown in Table 5. We calculate the Gibbs free energy surface (necessary conditions for equilibrium).
Gibbs free energies for all liquid phases (L1, L2 and L3) and their In other words, the number, locations, and values of all stationary

Table 5
NEWTON LLFLSH calculations for the mixture of ethanol (1), water (2) and benzene (3). Temperature 338.48 K. NRTL/Ideal/Chemical model. z = [0.2, 0.4, 0.4].

x1 x2 x3 k g(x)/RT

Overall composition z 0.2 0.4 0.4 0.26525


Starting point L1 0.10761 0.03325 0.85914 0.45715
Starting point L2 0.28444 0.54716 0.16840 0.36833
Starting point L3 0.17381 0.78856 0.03763 0.17452
L1 at LLE 0.29063 0.52356 0.18581 0.41589 0.38052
L2 at LLE 0.10827 0.03347 0.85826 0.36548 0.19130
L3 at LLE 0.18094 0.77769 0.04137 0.21863 0.27339
Mixture of L1, L2 and L3 0.28794
S.K. Wasylkiewicz et al. / Fluid Phase Equilibria 358 (2013) 304318 311

Table 6
Results of the search for all stationary points for the mixture of ethanol (1), water (2) and benzene (3). Temperature 338.48 K. NRTL/Ideal/Chemical model. z = [0.29063,
0.52356, 0.18581].

Type Index F(x, z)/RT x1 x2 x3 Eigenvalues

MIN 1 0.0 0.29063 0.52356 0.18581 0.09429 5.97733


SA 1 3.07E-04 0.24292 0.66400 0.09308 0.06444 13.33974
MIN 1 3.85E-07 0.18095 0.77768 0.04137 0.16635 35.27175
SA 1 7.57E-03 0.22859 0.16793 0.60348 0.54349 4.73123
MIN 1 2.02E-07 0.10827 0.03347 0.85826 20.40576 2.93791

Topological criterion: Nmax + (1)N Nmin + Nsad even Nsad odd = (1)N is satised (N = 2)
Number of minima Nmin = 3
Number of odd saddles Nsad odd = 2

points of tangent plane distance function must be the same for each Contour lines of the dimensionless tangent plane distance func-
liquid phase in equilibrium and a stability test for any one of the tion F(x, z)/RT calculated for the overall composition z = [0.27078,
compositions satisfying the LLE is sufcient [13]. 0.47302, 0.25620] at temperature 295 K are shown in Fig. 6 and
Fig. 7. The UNIQUAC model was used for calculations. The binary
8.2. Ternary mixture of ethylene glycol, lauryl alcohol and and pure component parameters are shown in Table 7.
nitromethane Efcient tracking of ridges and valleys of the potential func-
tion F(x, z)/RT is crucial for the success of our global optimization
The ethylene glycol lauryl alcohol nitromethane system is method. We start at an overall composition z (trivial solution for
known to have three liquid phases in equilibrium in part of the the tangent plane stability problem) in a dominant eigendirection
composition space (see e.g. Fig. 5). The system was investigated and follow the valley up to the lauryl alcohol vertex using the con-
by several researchers [9,4751]. They tested phase stability for a tinuation method (see branch 1 in Fig. 8). Then we follow the valley
number of mixture compositions using various calculation meth- in the opposite direction (branch 2) up to a stationary point, in this
ods. The pure component and binary parameters calculated based case a saddle point. We store this point as another starting point
on data from [9,50] are presented in Table 7. The dimensionless for branch tracking and calculate all eigenvalues and eigenvectors
interaction parameters  ij in the UNIQUAC equation are related necessary for determination of the type of the stationary point
to the interaction energy parameters uij as  ij = exp( uij /RT), as well as possible directions to start new branches. We continue
where R = 1.9858 cal/(mol K). branch tracking until all starting points have been examined and

Fig. 3. LLLE calculation results, and branches for tracking ridges and valleys of the dimensionless tangent plane distance function F(x, z)/RT calculated to test stability of one
liquid phase L1 = [0.29063, 0.52356, 0.18581] at temperature 338.48 K.
312 S.K. Wasylkiewicz et al. / Fluid Phase Equilibria 358 (2013) 304318

Table 7
UNIQUAC binary and pure component parameters for the mixture of ethylene glycol (1), lauryl alcohol (2) and nitromethane (3). Binary interaction parameters uij are in
cal/mol.

i ui1 ui2 ui3 ri qi

1 490.8898 108.6252 2.4088 2.248


2 138.3904 606.7016 8.8495 7.372
3 929.1161 264.4887 2.0086 1.868

Table 8
Results of the search for all stationary points for the mixture of ethylene glycol (1), lauryl alcohol (2) and nitromethane (3). Temperature 295 K. UNIQUAC model. z = [0.27078,
0.47302, 0.25620].

Type Index F(x, z)/RT x1 x2 x3 Eigenvalues

MIN 1 0.0 0.27078 0.47302 0.25620 1.01885 7.80981


SA 1 0.02522 0.53802 0.18824 0.27374 0.96906 3.92990
SA 1 0.01677 0.51886 0.01731 0.46383 26.3081 0.90359
MIN 1 0.01689 0.90590 0.01366 0.08044 72.9342 3.12370
MIN 1 0.00200 0.24249 0.00167 0.75584 530.977 1.48015

Topological criterion: Nmax + (1)N Nmin + Nsad even Nsad odd = (1)N is satised (N = 2)
Number of maxima Nmax = 0
Number of minima Nmin = 3
Number of even saddles Nsad even = 0
Number of odd saddles Nsad odd = 2
Number of tangent nodes Ntng = 0

the topological criterion is fullled. In this case we have found two Table 9
UNIQUAC/Ideal/Chemical binary parameters for the mixture of ethylene glycol (1),
more minima and one more saddle point (see Fig. 7 and Fig. 9). All
lauryl alcohol (2) and nitromethane (3). Binary interaction parameters bij are in K.
stationary points found in the process are summarized in Table 8.
Liquidliquid equilibrium (LLE) for the mixture of ethylene gly- i bi1 bi2 bi3
col, lauryl alcohol and nitromethane at a temperature of 295 K 1 118.48063 478.35228
(Fig. 5) was calculated by UNIQUAC/Ideal/Chemical model using 2 162.55850 516.51748
VMGSim [44] with the binary interaction parameters shown in 3 140.25308 72.59988
Table 9. The dimensionless interaction parameters  ij in the UNI-
QUAC equation are related to the interaction energy parameters
We start with the stability test for the overall composition
uij and dened as ln( ij ) = uij /RT = aij + bij /T + cij ln(T).
z = [0.4, 0.2, 0.4] (see Table 10). In this case the starting point
In the following few tables, we show an example of the detailed
for branch tracking (overall composition z) is a saddle point on
procedure for correct calculations of LLE for the overall composition
the dimensionless tangent plane distance function F(x, z)/RT. One
z = [0.4, 0.2, 0.4] at a temperature of 295 K. We use our stability test
eigenvalue at this stationary point (trivial solution z) is positive,
and LLE ash calculations to make sure that the LLE is correct and
the other is negative. Using our global optimization routine, two
all liquid phases in equilibrium are stable.
other (nontrivial) stationary points (minima) have been found in
the whole composition space. Since the topological criterion is sat-
ised (see Table 10), we conclude that all stationary points in this
case have been found.

Fig. 4. Vaporliquidliquid equilibrium for the mixture of ethanol, water and ben- Fig. 5. Liquidliquid equilibrium for the mixture of ethylene glycol, lauryl alcohol
zene at pressure 101.325 kPa. and nitromethane at temperature 295 K.
S.K. Wasylkiewicz et al. / Fluid Phase Equilibria 358 (2013) 304318 313

Table 10
Results of the search for all stationary points for the mixture of ethylene glycol (1), lauryl alcohol (2) and nitromethane (3). Temperature 295 K. UNIQUAC/Ideal/Chemical
model. z = [0.4, 0.2, 0.4].

Type Index F(x, z)/RT x1 x2 x3 Eigenvalues

SA 1 0.0 0.4 0.2 0.4 0.51755 0.47273


MIN 1 0.00631 0.20268 0.10520 0.69212 4.31620 1.14639
MIN 1 0.15559 0.97101 0.00445 0.02454 244.348 23.1294

Topological criterion: Nmax + (1)N Nmin + Nsad even Nsad odd = (1)N is satised (N = 2)
Number of minima Nmin = 2
Number of odd saddles Nsad odd = 1

Table 11
NEWTON LLFLSH calculations for the mixture of ethylene glycol (1), lauryl alcohol (2) and nitromethane (3). Temperature 295 K. UNIQUAC/Ideal/Chemical model. z = [0.4,
0.2, 0.4].

x1 x2 x3 k g(x)/RT

Overall composition z 0.4 0.2 0.4 0.14696


Starting point L1 0.20268 0.10520 0.69212 0.73383
Starting point L2 0.97101 0.00445 0.02454 0.26617
L1 at LLE 0.21198 0.26469 0.52333 0.74885 0.22162
L2 at LLE 0.96063 0.00711 0.03226 0.25115 0.03840
Mixture of L1 and L2 0.17560

Since values of dimensionless tangent plane distance function


F(x, z)/RT are negative at both minima (Table 10), the overall com-
position z represents an unstable liquid. Now we switch to our
rigorous LLE ash calculation routine (see Appendix A). The compo-
sitions of minima found in the previous stability test step (Table 10)
are good initial guesses for LLE ash calculations (see Table 11).
The NEWTON LLFLSH routine converges to two-phase LLE solution
shown in Table 11. We calculate the Gibbs free energies for both
liquid phases (L1 and L2) and their mixture and then compare to
the value calculated for the overall composition z. Since the value
for the mixture is lower than that for the overall composition, we
conclude that the liquid z will possibly split into L1 and L2. Now
we check if the liquids L1 and L2 are stable.
We perform the stability test for one liquid phase (e.g. liquid L1
from Table 11). In this case the starting point for branch tracking is a
saddle point on the dimensionless tangent plane distance function
(see Table 12). Using our global optimization routine, two other sta-
tionary points (minima) have been found in the whole composition
space. Since values of F(x, z)/RT are negative at both minima, L1 is
unstable.

Fig. 6. Contour lines of the dimensionless tangent plane distance function F(x, z)/RT
calculated for the overall composition z = [0.27078, 0.47302, 0.25620] at tempera-
ture 295 K.

Fig. 7. Contour lines of the dimensionless tangent plane distance function F(x, z)/RT Fig. 8. Branches for tracking ridges and valleys of the dimensionless tangent plane
calculated for the overall composition z = [0.27078, 0.47302, 0.25620] at tempera- distance function F(x, z)/RT calculated for the overall composition z = [0.27078,
ture 295 K. Lower part of the composition space. 0.47302, 0.25620] at temperature 295 K.
314 S.K. Wasylkiewicz et al. / Fluid Phase Equilibria 358 (2013) 304318

Table 12
Results of the search for all stationary points for the mixture of ethylene glycol (1), lauryl alcohol (2) and nitromethane (3). Temperature 295 K. UNIQUAC/Ideal/Chemical
model. z = [0.21198, 0.26469, 0.52333].

Type Index F(x, z)/RT x1 x2 x3 Eigenvalues

SA 1 0.0 0.21198 0.26469 0.52333 0.66599 2.48297


MIN 1 0.01025 0.28617 0.48392 0.22991 1.22381 2.74535
MIN 1 0.00330 0.14699 0.13711 0.71590 1.35732 4.29993

Topological criterion: Nmax + (1)N Nmin + Nsad even Nsad odd = (1)N is satised (N = 2)
Number of minima Nmin = 2
Number of odd saddles Nsad odd = 1

Table 13
NEWTON LLFLSH calculations for the mixture of ethylene glycol (1), lauryl alcohol (2) and nitromethane (3). Temperature 295 K. UNIQUAC/Ideal/Chemical model. z = [0.4,
0.2, 0.4].

x1 x2 x3 k g(x)/RT

Overall composition z 0.4 0.2 0.4 0.14696


Starting point L1 0.28617 0.48393 0.22991 0.45698
Starting point L2 0.14699 0.13711 0.71590 0.27995
Starting point L3 0.96063 0.00711 0.03226 0.26307
L1 at LLE 0.14327 0.12684 0.72989 0.43450 0.18771
L2 at LLE 0.28944 0.46722 0.24334 0.30627 0.28926
L3 at LLE 0.96094 0.00691 0.03215 0.25923 0.03829
Mixture of L1, L2 and L3 0.18008

Table 14
Results of the search for all stationary points for the mixture of ethylene glycol (1), lauryl alcohol (2) and nitromethane (3). Temperature 295 K. UNIQUAC/Ideal/Chemical
model. z = [0. 14327, 0. 12684, 0. 72989].

Type Index F(x, z)/RT x1 x2 x3 Eigenvalues

MIN 1 0.0 0.14327 0.12684 0.72989 1.74497 4.54342


SA 1 6.20e-3 0.22688 0.28727 0.48585 0.68808 2.24902
MIN 1 3.83e-8 0.28944 0.46722 0.24334 1.11368 2.37591

Topological criterion: Nmax + (1)N Nmin + Nsad even Nsad odd = (1)N is satised (N = 2)
Number of minima Nmin = 2
Number of odd saddles Nsad odd = 1

Now we switch again to our LLE ash calculation routine. and their mixture and then compare to the value calculated for the
The compositions of minima found in the previous stability test overall composition z. Since the value for the mixture is lower than
step (Table 12) combined with the composition of liquid L2 from that for the overall composition, we conclude that the liquid z will
Table 11 (not tested for stability) are good initial guesses for LLE possibly split into L1, L2 and L3. Notice that Gibbs free energy for
ash calculations (see Table 13). Now the NEWTON LLFLSH rou- the two-liquid mixture in Table 11 is higher than that for the three-
tine converges to three-phase LLE solution shown in Table 13. We liquid mixture in Table 13. Now we have to check if the liquids L1,
calculate the Gibbs free energies for all liquid phases (L1, L2 and L3) L2 and L3 are stable.
We perform the stability test for one liquid phase (e.g. liquid L1
from Table 13). In this case the starting point for branch tracking is
a minimum on the dimensionless tangent plane distance function
(see Table 14). Using our global optimization routine, two other
stationary points have been found in the whole composition space.
Since values of F(x, z)/RT are non-negative at all stationary points,
L1 is stable.

8.3. Five-component mixture

This example involves the ve-component mixture of acetic acid


(1), benzene (2), furfural (3), cyclohexane (4) and water (5). The
UNIQUAC model is used. The binary and pure component parame-
ters taken from [49,52] are presented in Table 15.
To calculate LLE for the overall composition z = [0.2, 0.2, 0.2, 0.2,
0.2] at a temperature of 295 K, we start with the stability test (see
Table 16). In this case the starting point for branch tracking (overall
composition z) is a saddle point on the dimensionless tangent plane
distance function F(x, z)/RT. One eigenvalue at this stationary point
is negative, the other eigenvalues are positive. Using our global
Fig. 9. Branches for tracking ridges and valleys of the dimensionless tangent plane
distance function F(x, z)/RT calculated for the overall composition z = [0.27078, optimization routine, two other stationary points (minima) have
0.47302, 0.25620] at temperature 295 K. Lower part of the composition space. been found in the whole composition space. Since the topological
S.K. Wasylkiewicz et al. / Fluid Phase Equilibria 358 (2013) 304318 315

Table 15
UNIQUAC binary and pure component parameters for the mixture of acetic acid (1), benzene (2), furfural (3), cyclohexane (4) and water (5). Binary interaction parameters
aij are in cal/mol.

i ai1 ai2 ai3 ai4 ai5 ri qi

1 137.068 711.466 94.32388 255.456 2.2024 2.072


2 0.16405 11.836 66.07903 1382.577 3.1878 2.400
3 674.2505 137.8886 22.39649 294.5019 3.1680 2.484
4 409.1192 53.8012 784.1499 1421.175 4.0464 3.240
5 523.082 1172.771 216.8131 957.9283 0.9200 1.400

Table 16
Results of the search for all stationary points for the mixture of acetic acid (1), benzene (2), furfural (3), cyclohexane (4) and water (5). Temperature 295 K. UNIQUAC model.
z = [0.2, 0.2, 0.2, 0.2, 0.2].

Type Index F(x, z)/RT x1 x2 x3 x4 x5 Eigenvalues

SA 1 6.22E-16 0.2 0.2 0.2 0.2 0.2 33.1187 1.0370 7.3246 5.4159
MIN 1 2.64E-02 0.09356 0.26321 0.07430 0.53290 0.03603 14.1313 151.200 4.8843 1.1669
MIN 1 2.75E-02 0.23356 0.06767 0.26305 0.01925 0.41648 5.6783 16.585 7.9320 37.810

Topological criterion: Nmax + (1)N Nmin + Nsad even Nsad odd = (1)N is satised (N = 4)
Number of minima Nmin = 2
Number of odd saddles Nsad odd = 1

criterion is satised (see Table 16), we conclude that all stationary Now we switch to our rigorous LLE ash calculation routine. We
points in this case have been found. use the compositions of minima found in the previous stability test
Since values of the dimensionless tangent plane distance func- step (Table 19) as the initial guess for the calculations. The LLE ash
tion F(x, z)/RT are negative at both minima (Table 16) the overall calculation routine does not converge to a three-phase solution but
composition z represents an unstable liquid. Now we switch to our stops at the boundary shown in Table 20. The phase fraction 1 for
LLE ash calculation routine (NEWTON LLFLSH). The compositions liquid phase L1 converged to zero.
of minima found in the previous stability test step (Table 16) are Now we eliminate the liquid phase L1 from the solution and
good initial guesses for LLE ash calculations. The NEWTON LLFLSH repeat the LLE ash calculations using the two remaining phases as
routine converges to two-phase LLE solution shown in Table 17. We the initial guess. The routine converges very quickly to two-phase
calculate the Gibbs free energies for both liquid phases (L1 and L2) LLE solution shown in Table 21. We calculate Gibbs free energy for
and their mixture and then compare to the value calculated for the both liquid phases at equilibrium and their mixture and then com-
overall composition z. Since the value for the mixture is lower than pare to the value calculated for the overall composition z. Since the
that for the overall composition, we conclude that the liquid z will value for the mixture is lower than that for the overall composition,
possibly split into L1 and L2. Now we have to check if the liquids we conclude that the liquid z will possibly split into L1 and L2.
L1 and L2 are stable. To be sure that we have found a stable two liquid phase solu-
We perform the stability test for one liquid phase (e.g. liquid L1 tion, we choose one liquid phase (e.g. liquid L1 from Table 21)
from Table 17). In this case the starting point for branch tracking is and perform stability test. In this case the starting point for branch
a minimum on the dimensionless tangent plane distance function tracking is a minimum on the dimensionless tangent plane distance
(see Table 18). Using our global optimization routine, two other function (see Table 22). Using our global optimization routine, four
stationary points have been found in the whole composition space. other stationary points have been found in the whole composition
Since values of F(x, z)/RT are non-negative at all stationary points, space. Since values of F(x, z)/RT are non-negative (within assumed
L1 is stable. tolerance 1.0E-06) at all stationary points, the liquid L1 is stable.

8.4. Ternary mixture close to liquid region boundaries 9. Conclusions

This example involves the three-component mixture of A general approach has been developed for nding all sin-
ethanol, water and benzene, the same as in example 8.1. The gular points of a mathematical model by exploring the natural
NRTL/Ideal/Chemical model is used with the binary interaction connectedness that exists between singular points of an objective
parameters shown in Table 1. An overall composition z = [0.28, 0.48, function. The idea of following ridges and valleys using informa-
0.24] was chosen to illustrate how the step 4 of the combined phase tion gathered along the way, which was previously used in the hill
equilibrium and stability algorithm, described in paragraph 7, was climbing optimization techniques, has been signicantly enhanced.
applied to nd correct number of liquid phases in equilibrium. The We developed a new arc length continuation method to overcome
overall composition is in the two-liquid region (see Fig. 4). It is the difculties with tracking of ridges and valleys of an objective
very close to the three-liquid region as well as to the homogeneous function. It gives an efcient and robust scheme for nding all sin-
liquid region. gular points of a mathematical model.
We start with stability test for the overall composition z = [0.28, The global optimization algorithm has been applied to the Gibbs
0.48, 0.24] (see Table 19). In this case the starting point for branch tangent plane stability test for multiphase liquid mixtures. The
tracking (overall composition z) is a saddle point on the dimen- algorithm quickly nds all stationary points of the tangent plane
sionless tangent plane distance function F(x, z)/RT. Using our global distance function predicted by a thermodynamic model. The algo-
optimization routine, four other stationary points (minima and sad- rithm is self-starting and signicantly improves reliability and
dle points) have been found in the whole composition space. Since robustness of multiphase equilibrium calculations. The developed
values of dimensionless tangent plane distance function F(x, z)/RT technique was tested on three different mixtures, ranging from a
are negative at all minima, the overall composition z represents an ternary system with three liquid phases to a ve component system
unstable liquid. with two liquid phases.
316 S.K. Wasylkiewicz et al. / Fluid Phase Equilibria 358 (2013) 304318

Table 17
NEWTON LLFLSH calculations for the mixture of acetic acid (1), benzene (2), furfural (3), cyclohexane (4) and water (5). Temperature 295 K. UNIQUAC model. z = [0.2, 0.2, 0.2,
0.2, 0.2].

x1 x2 x3 x4 x5 k g(x)/RT

Overall composition z 0.2 0.2 0.2 0.2 0.2 1.71021


Starting point Liquid 1 0.09356 0.26321 0.07430 0.53290 0.03603 0.61574
Starting point Liquid 2 0.23356 0.06767 0.26305 0.01925 0.41648 0.38426
Liquid 1 at LLE 0.25477 0.12346 0.26894 0.04017 0.31266 0.56269 2.11206
Liquid 2 at LLE 0.12953 0.29849 0.11129 0.40565 0.05504 0.43731 1.22591
Mixture of L1 and L2 1.72454

Table 18
Results of the search for all stationary points for the mixture of acetic acid (1), benzene (2), furfural (3), cyclohexane (4) and water (5). Temperature 295 K. UNIQUAC model.
z = [0.25477, 0.12346, 0.26894, 0.04017, 0.31266].

Type Index F(x, z)/RT x1 x2 x3 x4 x5 Eigenvalues

MIN 1 0.0 0.25477 0.12346 0.26894 0.04017 0.31266 5.8570 20.8035 2.2215 15.1064
SA 1 7.07E-03 0.21286 0.23380 0.20694 0.17911 0.16729 35.3976 0.5109 7.0040 5.0370
MIN 1 9.06E-07 0.12952 0.29849 0.11129 0.40566 0.05504 10.6419 98.131 3.9641 0.6471

Topological criterion: Nmax + (1)N Nmin + Nsad even Nsad odd = (1)N is satised (N = 4)
Number of minima Nmin = 2
Number of odd saddles Nsad odd = 1

Table 19
Results of the search for all stationary points for the mixture of ethanol (1), water (2) and benzene (3). Temperature 338.48 K. NRTL/Ideal/Chemical model. z = [0.28, 0.48,
0.24].

Type Index F(x, z)/RT x1 x2 x3 Eigenvalues

SA 1 0.0 0.28 0.48 0.24 0.0122 4.77476


MIN 1 7.18E-03 0.12993 0.84847 0.02160 0.84230 74.93841
MIN 1 2.28E-06 0.28186 0.44457 0.27357 0.00947 4.30751
SA 1 3.27E-05 0.27499 0.35109 0.37392 0.03929 3.72136
MIN 1 2.44E-02 0.08783 0.02500 0.88717 28.40719 5.16810
Topological criterion: Nmax + (1)N Nmin + Nsad even Nsad odd = (1)N is satised (N = 2)
Number of minima Nmin = 3
Number of odd saddles Nsad odd = 2
Table 20
NEWTON LLFLSH calculations for the mixture of ethanol (1), water (2) and benzene (3). Temperature 338.48 K. NRTL/Ideal/Chemical model. z = [0.28, 0.48, 0.24].

x1 x2 x3 k

Starting point L1 0.12993 0.84847 0.02160


Starting point L2 0.28186 0.44457 0.27357
Starting point L3 0.08783 0.02500 0.88717
L1 at boundary 0.18001 0.77977 0.04022 0.0
L2 at boundary 0.29256 0.51726 0.19018 0.92323
L3 at boundary 0.10850 0.03333 0.85817 0.07677

Table 21
NEWTON LLFLSH calculations for the mixture of ethanol (1), water (2) and benzene (3). Temperature 338.48 K. NRTL/Ideal/Chemical model. z = [0.28, 0.48, 0.24].

x1 x2 x3 k g(x)/RT

Overall composition z 0.28 0.48 0.24 0.36836


Starting point L1 0.29256 0.51726 0.19018
Starting point L2 0.10850 0.03333 0.85817
L1 at LLE 0.29302 0.51391 0.19307 0.92939 0.38279
L2 at LLE 0.10868 0.03360 0.85772 0.07061 0.19171
Mixture of L1 and L2 0.36930

Table 22
Results of the search for all stationary points for the mixture of ethanol (1), water (2) and benzene (3). Temperature 338.48 K. NRTL/Ideal/Chemical model. z = [0.29302,
0.51391, 0.19307].

Type Index F(x, z)/RT x1 x2 x3 Eigenvalues

MIN 1 0.0 0.29302 0.51391 0.19307 0.11061 5.73339


SA 1 5.02E-04 0.23418 0.68295 0.08287 0.06212 15.34030
MIN 1 3.57E-04 0.18653 0.76894 0.04453 0.12266 32.32635
SA 1 7.45E-03 0.22864 0.16724 0.60412 0.54250 4.73831
MIN 1 -5.89E-08 0.10868 0.03360 0.85772 20.31344 2.90827

Topological criterion: Nmax + (1)N Nmin + Nsad even Nsad odd = (1)N is satised (N = 2)
Number of minima Nmin = 3
Number of odd saddles Nsad odd = 2
S.K. Wasylkiewicz et al. / Fluid Phase Equilibria 358 (2013) 304318 317

In the paper we show that the number of liquid phases in equi- 6. Calculate and normalize new mole fractions x ik (i = 1, . . ., Nc ) in
librium cannot be determined on the basis of tangent plane analysis all other phases (k = 2, . . ., Np ) using Eq. (A.1).
alone and the combined phase equilibrium and stability algorithm 7. Calculate differences between guessed and calculated mole frac-
should be used. We also check global consistency of each tangent tions
plane stability test.
fj xik x ik i = 1, . . ., Nc 1 k = 1, . . ., Np

Acknowledgment j = 1, . . ., (Nc 1) Np (A.7)

and solve the set of equations


Virtual Materials Group Inc. is gratefully acknowledged for
nancial support of the project. fj = 0 j = 1, . . ., (Nc 1)p (A.8)

Appendix A. Calculation of liquidliquid equilibrium Using the NewtonRaphson (NR) method.


The algorithm is efcient at solving the LLE problems for two
Let us consider the liquidliquid equilibrium (LLE) problem for a liquid phases [26,54], where the algorithm reduces to only one
mixture of Nc components, at a specied temperature and pressure, RachfordRice equation (in point No. 4). In this case it can be solved
and overall composition z. The mole fractions (xi k , i = 1, . . ., Nc , k = 1, using a one-variable solver that is usually faster and more robust
. . ., Np ) and phase fractions (k , k = 1, . . ., Np ) for all Np liquid phases than multi-variable NR method. Even for this simplest case there
(2 Np Nc ) have to be calculated. To this end, the following set of are two issues when solving these types of equations [55]. The
equations has to be solved. existence of vertical asymptotes splits the equation domain into
Equilibrium equations a set of Nc 1 regions. Each one contains a different solution but
only one has physical meaning. Also, conventional solution meth-
i1 ods often exhibit slow convergence. A lot of efforts have been made
xik = xi1 = Kik xi1 i = 1, . . ., Nc k = 2, . . ., Np (A.1)
ik to overcome these problems, including utilizing proper function
transformation to make the transformed problem as linear as pos-
Summation equations sible [54], applying some special function classes [55], and using
the block-algebra approach [56] or various initialization schemes

Nc
[57].
xik = 1 k = 1, . . ., Np (A.2) To overcome numerical problems with the imbedding inner
i=1 loop NR solver (for phase fraction calculations in point No. 4) inside
the outer loop NR solver (for mole fraction calculations) for the
Material balances
LLE problems with more than two liquid phases, the set of Eqs.

Np (A.1)(A.4) has been rearranged as follows [19].
zi = k xik i = 1, . . ., Nc (A.3)

Np
k=1 xi1 k (xik xi1 ) zi = 0 i = 1, . . ., Nc 1 (A.9)
Summation equation for phase fractions k=2


Np
xik Kik xi1 = 0 i = 1, . . ., Nc 1 k = 2, . . ., Np (A.10)
k = 1 (A.4)
 
k=1

Nc 1 
Nc 1
KNk c 1 xi1 zi + xik 1 = 0 k = 2, . . ., Np (A.11)
The RachfordRice [53] formulation is commonly used to solve
this problem. The algorithm to calculate multiple-liquid equilib- i=1 i=1

rium (Np 2) is a follows [26,54] First we eliminated xN k (k = 1, . . ., N ) from the set of independent
c
p
variables using summation equations Eq. (A.2). The liquid phase
1. Guess xik , i = 1, . . ., Nc 1, k = 1, . . ., Np . fraction for the reference phase 1 was also eliminated using Eq.
k , k = 1, . . ., N , using Eq. (A.2).
2. Calculate xN p (A.4). Then we modied Eq. (A.3) to get Eq. (A.9), and Eq. (A.1) to get
c
3. Calculate Kik , i = 1, . . ., Nc , k = 1, . . ., Np , using Eq. (A.1). Eq. (A.10) for all independent mole fractions (xi k , i = 1, . . ., Nc 1,
4. Solve the set of p 1 RachfordRice equations obtained by com- k = 1, . . ., Np ) and Eq. (A.11) for the last component Nc .
bining equilibrium, summation and material balance equations. Now, the set of (Nc 1) Np + Np 1 equations (A.9)(A.11) can be
solved simultaneously to get all phase fractions and phase composi-
tions at equilibrium. The calculations are especially efcient when

Nc
j
Nc j
(Ki 1)zi a good initial estimate is provided from a rigorous Gibbs tangent
Hj (2 , . . ., p ) = (xi xi1 ) = Np plane analysis [13,23].
i=1
i=1 1 +
k=2
(Kik 1)k

= 0 j = 2, . . ., Np (A.5) Appendix B. Topological criterion for the tangent plane


distance function
for liquid phase fractions k (k = 2, . . ., p) using the
NewtonRaphson method. For the tangent plane distance function F(x), the ow of trajec-
5. Calculate and normalize new mole fractions in the reference tories is always into the permissible region. This special property
phase 1 combined with the common properties of gradient systems (15)
allowed developing the topological criterion [13], Eq. (B.1),
zi
x i1 = Np i = 1, . . ., Nc (A.6) Nmax + (1)N Nmin + Nsad Nsad = (1)N (B.1)
even odd
1+ k=2
(Kik 1)k
318 S.K. Wasylkiewicz et al. / Fluid Phase Equilibria 358 (2013) 304318

where Nmax is the number of maxima of F(x), Nmin is the number of [26] M.F. Doherty, M.F. Malone, Mayower v.9. Software for Distillation, Depart-
minima, Nsad even is the number of saddles with an even number of ment of Chemical Engineering, University of Massachusetts, Amherst, 1994.
[27] A. Martin, S. Pokutta, L. Schewe, Chem. Ing. Tech. 84 (2012) 19.
negative eigenvalues, and Nsad odd is the number of saddles with an [28] en.wikipedia.org/wiki/Global optimization, last access date (19.05.2013).
odd number of negative eigenvalues of the Jacobian matrix J[ F( [29] Y. Zhu, H. Wen, Z. Xu, Chem. Eng. Sci. 55 (2000) 34513459.
x)]. [30] M. Srinivas, G.P. Rangaiah, Comput. Chem. Eng. 30 (2006) 14001415.
[31] G.P. Rangaiah, Fluid Phase Equilib. 83 (2001) 187188.
The index of a stationary point x can be determined from the [32] A. Kabouche, A. Boultif, A. Abidi, N. Gherraf, Fluid Phase Equilib. 336 (2012)
eigenvalues i of the Jacobian matrix J[ F( x)] calculated at this 113121.
point [33] J. Kennedy, R. Eberhart, Proc. IEEE Int. Conf. Neural Netw. (2005) 19421948.
[34] R. Storn, K. Price, J. Global Optim. 11 (1997) 341359.

N [35] A. Bonilla-Petriciolet, Fluid Phase Equilib. 332 (2012) 720.
Ind(x) = sgn i (B.2) [36] V. Bhargavaa, S.E.K. Fateenb, A. Bonilla-Petriciolet, Fluid Phase Equilib. 337
(2013) 191200.
i=1 [37] A. Bonilla-Petriciolet, G.P. Rangaiah, J.G. Segovia-Hernndez, Fluid Phase Equi-
lib. 287 (2010) 111125.
[38] CMOST. Studio, Users Guide, Version 2010, Computer Modelling Group Ltd,
References 2010.
[39] C.Y. Gau, J.F. Brennecke, M.A. Stadtherr, Fluid Phase Equilib. 168 (2000)
[1] M.F. Doherty, M.F. Malone, Conceptual Design of Distillation Systems, 1st ed., 118.
McGraw-Hill, 2001. [40] G.M. Bollas, P.I. Barton, A. Mitsos, Chem. Eng. Sci. 64 (2009) 17681783.
[2] S.K. Wasylkiewicz, L.C. Kobylka, F.J.L. Castillo, Chem. Eng. J. 92 (2003) 201208. [41] G.S.G. Beveridge, R.S. Schechter, Optimization: Theory and Practice, McGraw-
[3] O. Ryll, S. Blagov, H. Hasse, Fluid Phase Equilib. 324 (2012) 108116. Hill, New York, 1970.
[4] J.W. Gibbs, Trans. Conn. Acad. Arts Sci. II (1873) 382404. [42] H.B. Keller, Numerical solution of bifurcation and nonlinear eigenvalue prob-
[5] J.W. Gibbs, Trans. Conn. Acad. Arts Sci. III (1876) 108248. lems, in: P.H. Rabinowitz (Ed.), Application of Bifurcation Theory, Academic
[6] L.E. Baker, A.C. Pierce, K.D. Luks, Soc. Pet. Eng. J. 22 (1982) 731742. Press, New York, 1977.
[7] M.L. Michelsen, Fluid Phase Equilib. 9 (1982) 119. [43] M.F. Doherty, M.F. Malone, Mayower v.10, Software for Distillation, Depart-
[8] M.L. Michelsen, Fluid Phase Equilib. 9 (1982) 2140. ment of Chemical Engineering, University of Massachusetts, Amherst, 1995.
[9] C.M. Mc Donald, C.A. Floudas, AIChE J. 41 (1995) 17981814. [44] VMGSim, VMGSim v. 7.1. VMG Process Simulator, Virtual Materials Group Inc,
[10] C.M. Mc Donald, C.A. Floudas, Comput. Chem. Eng. 19 (1995) 11111139. Calgary, 2012 (November).
[11] C.M. Mc Donald, C.A. Floudas, Ind. Eng. Chem. Res. 34 (1995) 16741687. [45] S.R.M Ellis, Trans. Inst. Chem. Eng. 30 (1952) 58.
[12] C.M. Mc Donald, C.A. Floudas, Comput. Chem. Eng. 21 (1997) 123. [46] G. Wozny, H. Cremer, Chem. Ing. Tech. 53 (1981) 89.
[13] S.K. Wasylkiewicz, L.N. Sridhar, M.F. Doherty, M.F. Malone, Ind. Eng. Chem. Res. [47] A.W. Francis, J. Phys. Chem. 60 (1956) 2027.
35 (1996) 13951408. [48] T. Chakravarty, C.W. White, W.D. Seider, AIChE J. 31 (1985) 316321.
[14] D.V. Nichita, S. Gomez, E. Luna, Fluid Phase Equilib. 194197 (2002) 411437. [49] S.R. Tessier, J.F. Brennecke, M.A. Stadtherr, Chem. Eng. Sci. 55 (2000)
[15] D.V. Nichita, S. Gomez, E. Luna, Comput. Chem. Eng. 26 (2002) 17031724. 17851796.
[16] A.C. Sun, W.D. Seider, Fluid Phase Equilib. 103 (1995) 213249. [50] R.P. Stateva, G.S. Cholakov, A.A. Galushko, W.A. Wakeham, Chem. Eng. Sci. 55
[17] J.Z. Hua, J.F. Brennecke, M.A. Stadtherr, Fluid Phase Equilib. 116 (1996) 5259. (2000) 21212129.
[18] D.B. Van Dongen, M.F. Doherty, Chem. Eng. Sci. 39 (1984) 883892. [51] J. Kangas, I. Malinen, J. Tanskanen, Ind. Eng. Chem. Res. 50 (2011) 70037018.
[19] S.K. Wasylkiewicz, LLE for Liquid Phases, Private Communication, Calgary, 2010 [52] J.M. Sorensen, W. Arlt, Liquid-liquid equilibrium data collection. Ternary Sys-
(20.03.2010). tems. Chemistry Data Series, vol. V, DECHEMA, Frankfurt, Germany, 1980.
[20] Z. Cheng-Dong, W. Xiao-Xia, Chaos Soliton Fract. 12 (2001) 153158. [53] H.H. Rachford, J.D. Rice, Trans. Am. Inst. Min. Metall. Eng. 195 (1952)
[21] A. Lucia, Y. Feng, AIChE J. 49 (2003) 25532563. 327328.
[22] S.K. Wasylkiewicz, M.F. Malone, M.F. Doherty, Ind. Eng. Chem. Res. 38 (1999) [54] S.K. Wasylkiewicz, Modelling and Simulation of Multistage Separation Pro-
49014912. cesses; Scientic Papers of the Institute of Chemical Engineering and Heating
[23] S.K. Wasylkiewicz, Search for all Stationary Points, Private Communication, Equipment, Technical University of Wrocaw, Wrocaw, Poland, 1992 (No. 65,
Calgary, 2009 (01.03.2009). Monograph No. 40).
[24] I. Prigogine, R. Defy, Chemical Thermodynamics, Longmans, London, 1967. [55] V. Gaganis, D. Marinakis, N. Varotsis, Fluid Phase Equilib. 322323 (2012) 918.
[25] S.K. Wasylkiewicz, Tangent Plane Criterion, Private Communication, Amherst, [56] R. Monroy-Loperena, Comp. Chem. Eng. 46 (2012) 217220.
1992 (20.11.1992). [57] Z. Li, A. Firoozabadi, Fluid Phase Equilib. 332 (2012) 2127.

Vous aimerez peut-être aussi