Fluid Phase Equilibria 408 (2016) 265e272

Contents lists available at ScienceDirect

Fluid Phase Equilibria

j o u r n a l h o m e p a g e : w w w . e l s e v i e r . c o m / l o c a t e / fl u i d

Effect of VLE uncertainties on the design of separation sequences by

distillation e Study of the benzeneechloroformeacetone system*
Paul M. Mathias
Fluor Corporation, 3 Polaris Way, Aliso Viejo, CA 92698, USA

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

Article history: It is widely recognized by experts that the computer-based design of chemical processes depends
Received 10 August 2015 strongly on the correlated thermodynamic and transport properties, and that the effect of the property
Received in revised form uncertainties should be accounted for in the design. The most significant source of property uncertainties
3 September 2015
on process design comes from the correlations of mixture phase equilibrium. Many approaches to un-
Accepted 3 September 2015
certainty analysis have been proposed, but uncertainty analysis is not a routine element of today's in-
Available online 8 September 2015
dustrial practice, mainly because education and awareness are lacking, and the proposed methods are
difficult to apply.
Keywords:
Vaporeliquid equilibrium
In order to help rectify the situation, the author developed an intuitive and easy-to-apply phase-
Property uncertainty equilibrium (specifically, activity coefficient) perturbation method based upon treating the mixture, for
Azeotropic separations the purpose of uncertainty perturbation, as a set of pseudobinaries described by the Margules equation.
Residue curves The approach was applied to two case studies e (1) a propylene-propane superfractionator for which
Distillation synthesis small changes in correlated relative volatilities have a large effect on the design of the distillation col-
umn; and (2) a dehexanizer column that separates a mixture containing many close-boiling components
e and demonstrated that the proposed methodology provides quantitative insight into the effect of
property uncertainties, and helps to estimate the safety factors that should be imposed upon the design.
In this paper the Margules perturbation method is applied to the “textbook” separation of an acetone
echloroformebenzene ternary mixture; this case study was proposed in 1996 by Westerberg and
Wahnshafft, and was previously studied by Parodi and Campanella. The approach of Parodi and Cam-
panella used different levels of data fitting to identify uncertainties, while the present methodology is to
obtain the best fit of the data, and then apply uncertainties based upon estimated combined model
uncertainties, which includes experimental uncertainties and any model inadequacies. The purpose of
this paper is to further study whether the Margules approach to mixture uncertainty offers insight, and is
a useful technique to perform uncertainty analysis.
© 2015 Elsevier B.V. All rights reserved.

1. Introduction as a significant barrier and advocated the use of online properties

While experts recognize that the effect of uncertainties in
property models on process designs should be quantitatively
evaluated, this analysis is rarely, if ever, done in industrial practice.
For example, Kim et al. [1] observed that, “Its practical imple-
mentation in a variety of scientific and engineering fields has
typically seen less emphasis than it deserves.” There are several
barriers to adopting quantitative uncertainty analysis in chemical-
engineering industrial practice. Kim et al. [1] highlighted education
*
This article belongs to an upcoming special issue 19th Symposium, Guest edited
by C. McCabe and I. Siepmann. However, due to a production error it was included
in this regular issue.
E-mail address: Paul.Mathias@Fluor.com.
with clearly defined uncertainties as the means to incorporate
uncertainty analysis into undergraduate and graduate courses.
Another barrier is that the methods proposed by academic re-
searchers (e.g., Macchietto, Maduabeuke, and Szcepanski [2], and
Reed and Whiting [3]) are not readily applicable by industrial
practitioners, who predominantly use commercial simulators.
Recognizing that a significant barrier to the adoption of uncertainty
analysis is that simple, intuitive methods are not available, Mathias
[4] proposed a practical method, guided by the Margules equation,
to perturb activity coefficients calculated by any chosen model.
The method proposed by Mathias [4] is based upon the
assumption that the dominant property uncertainty results from
the uncertainty in phase equilibrium, and, in turn, the phase-
equilibrium uncertainty usually results from the liquid mixture

http://dx.doi.org/10.1016/j.fluid.2015.09.004
0378-3812/© 2015 Elsevier B.V. All rights reserved.
266 P.M. Mathias / Fluid Phase Equilibria 408 (2016) 265e272
nonideality. Mathias applied his method to two examples e (1) a
propylene-propane superfractionator for which small changes in
correlated relative volatilities have a large effect on the design of
the distillation column; and (2) a dehexanizer column that sepa-
rates a mixture containing many close-boiling components e and
demonstrated that the proposed methodology provides quantita-
tive insight into the effect of property uncertainties, and enables
quantitative estimation of the safety factors that need to be
imposed upon the design. A key conclusion by Mathias [4] is
repeated here: “The [property] uncertainties in the dehexanizer
column are about an order of magnitude higher than for the pro-
paneepropylene splitter, but the design uncertainties are much
larger for the propaneepropylene splitter. Qualitatively, this is an
expected result; however the value of the present approach is that
the design uncertainties can be quantified with modest effort.”
Mathias [4] also emphasized the point that “The first task that
must be completed in any uncertainty analysis is to quantitatively
understand the correlation uncertainties.” Correlation uncertainties
largely depend upon measurement uncertainties, and here the
problem is that uncertainties have rarely been reported in experi-
mental papers [5]. While there has been continuous progress in the
reporting of uncertainties, the majority of existing publications do
not report measurement uncertainties [5,6]. Six key journals in the
field of thermodynamics (Journal of Chemical and Engineering Data
[7], Journal of Chemical Thermodynamics [8], Fluid Phase Equilibria
[9], Thermochimica Acta [10], and The International Journal of Ther-
mophysics [11]) have mandated reporting of combined uncertainties
together with the experimental data tables, but this is a future
benefit. In this work, correlation uncertainties have been inferred by
comparing model calculations with measurements from various
sources, and by the judgment of an experienced practitioner [6].
In this paper the Margules perturbation method is applied to the
“textbook” separation of the benzeneechloroformeacetone
mixture, a ternary system containing one maximum-boiling azeo-
trope. This example was proposed by Westerberg and Wahnshafft
[12], and uncertainty analysis was previously performed by Parodi
and Campanella [13]. Parodi and Campanella executed the uncer-
tainty analysis by using a design-of-experiments approach, and
compared the simulation results using property regression from
various sets of fits to the experimental data. In the present work the
best fit of the data was obtained and perturbations were performed
using estimated uncertainties. The present approach enables a
quantitative relationship between the correlation uncertainties and
design uncertainties, and facilitates new insight into handling
process-design uncertainties.
2. Margules-based activity-coefficient perturbation scheme
The perturbation scheme for the mixture nonideality has been
presented by Mathias [4], and hence is only briefly summarized
here. The approach uses the simplification that, for the purpose of
perturbation to the activity coefficient, the mixture can be divided
into a set of pseudobinaries, which are then described by the
Margules equation.

  
lnðgi Þ ¼ ln gm
p
þ ln gi (1)
i
where gi is the effective activity coefficient after perturbation, gm
i is
the activity coefficient calculated by the model (the “best” appli-
p
cable model should be chosen), and gi represents the perturbation
due to the combined model uncertainty. We emphasize that gm i is
calculated from the chosen activity-coefficient model, and is not in
any way constrained by the pseudo-binary approximation. gpi is
given by,
2 
m  3
  ln g 
24
p
ln gi ¼ di ð1  xi Þ i  5 (2)
 
ð1  xi Þ2 þ ln gm
i 
Equation (2) is phenomenological, but has desirable character-
istics from the viewpoint of activity-coefficient perturbation. The
value of lnðgpi Þ is small in value when gm
i is close to unity (i.e., al-
ways small departures from Raoult's law), and also is low when the
mole fraction of component i is close to unity. For components at
low concentrations and with activity coefficients significantly
different from unity, the fractional change in gi resulting from the
perturbation approaches {exp(di) e 1}. In order to apply the
perturbation method, the value of di for each component must be
related to the estimated uncertainty in its activity coefficient at
infinite dilution, and this procedure is a key element of this
perturbation method.
3. Correlation of vaporeliquid equilibrium in the
acetoneebenzeneechloroform ternary mixture
As in the previous application by Mathias [4], the perturbation
scheme has been implemented as a user model in the NRTL-RK
property option in Aspen Plus V7.3. The NRTL-RK property option
uses the NRTL [14] activity-coefficient model and the Redlich-
Kwong [15] equation of state for the vapor phase. The pure-
component properties have been taken from the Aspen Plus V7.3
database and these are expected to be sufficiently accurate. The
application here is at a pressure of 1 atm, and the temperatures of
interest are approximately 50e85  C, and at these temperatures the
correlations for the vapor pressures of these three compounds from
the Aspen Plus database are expected to be better than 1%. The basis
for this confidence is that the Aspen Plus pure-component prop-
erties come from the DIPPR database [16], and the correlated vapor
pressures have been reported to be accurate to <1%. The Redlich-
Kwong equation is used for the vapor-phase fugacity coefficient,
but this thermodynamic quantity is expected to close to unity and
accurately predicted since the pressure of interest is low, z1 atm.
The Data Regression System (DRS) in Aspen Plus V7.3 has been
used to obtain optimum values of the NRTL parameters at pressures
of about 1 atm, which corresponds to the temperature range of
z50e85  C. An extremely important source of vaporeliquid
equilibrium (VLE) data for the benzeneeacetoneechloroform sys-
tem comes from the work of Kojima et al., [17], who reported
isobaric (1 atm) VLE data for the ternary system as well as the three
constituent binaries, but it should be emphasized there are many
sets of binary data available for the three constituent binary sys-
tems. Binary VLE data were obtained from NIST-TDE [18], and this
source clearly indicates that these are well-studied systems since it
lists 30, 20 and 13 publications containing VLE data for the aceto-
neechloroform, acetoneebenzene and benzeneechloroform bi-
naries, respectively. The NRTL parameter fit focused on the
50e85  C temperature range and the fitted NRTL parameters are
presented in Table 1.
The NRTL model has two binary parameters (aij and tij). aij is a
symmetric binary parameter (here assumed to be independent of
temperature), and tij is a temperature-dependent asymmetric bi-
nary parameter,

tij ¼ aij þ bij T þ eij lnðTÞ þ fij T (3)
The values of eij and fij have been defaulted to zero and the other
two temperature-dependent parameters as well as aij are fitted
values, as guided by experienced judgment, to obtain an optimum
representation of the VLE data. Table 1 also shows the infin-
iteedilution activity coefficients at 1 atm, and these values are
 P.M. Mathias / Fluid Phase Equilibria 408 (2016) 265e272 267

Table 1 85
NRTL parameters for the constituent binaries of the acetoneebenzeneechloroform
system. The table also reports the infinite-dilution activity coefficients and absolute 80 Bubble-Model
values of their logarithms. Dew-Model
Component 1 Benzene Acetone Chloroform 75 Bubble-Kojina-1991
Dew-Kojima-1991
Component 2 Chloroform Benzene Acetone

Temperature ( C)
70
a12 0 2.31 4.18
a21 0 3.15 2.16 65
b12 34.1 630.5 1054.1
b21 37.2 727.5 222.6
60
a12 0.3 0.3 0.16
g∞1 0.805 1.621 0.526
g ∞
0.815 1.376 0.425 55
2 
 ∞ 
lnðg1∞ Þ 0.22 0.48 0.64
lnðg Þ 0.21 0.32 0.86 50
2
0 0.2 0.4 0.6 0.8 1
x, y (Acetone)
important to the perturbation scheme (Equations (1) and (2)), as
Fig. 2. Txy diagram of the acetoneebenzene binary mixture at 1 atm. Comparison of
discussed later in this paper. calculated results (NRTL model with parameters in Table 1) to the data of Kojima et al.
Figs. 1e3 present comparisons of the calculated Txy diagrams at [17].
1 atm using the NRTL model with parameters in Table 1 for the
acetoneechloroform, acetoneebenzene and chloroformebenzene
binaries, respectively, to the data of Kojima et al. [17] These figures 85
suggest that the model provides an excellent fit of the three sets of Bubble-Model
binary data, but these charts do not enable uncertainty estimates. 80 Dew-Model
For this purpose we use relative-volatility charts; the relative- Bubble-Kojima-1991
volatility charts have been restricted to the 1 atm isobaric data- Dew-Kojima-1991
Temperature (°C)

75
sets where both vapor and liquid compositions have been
measured.
The K-value of component i in a mixture (Ki) at vaporeliquid 70
equilibrium is equal by the ratio of the vapor and liquid mole
fractions, and the relative volatility between components i and j
65
(aij) is defined as the ratio of the K-values.

y 60
Ki ≡ i (4)
xi 0 0.2 0.4 0.6 0.8 1
x, y (Chloroform)
Ki
aij ≡ (5) Fig. 3. Txy diagram of the chloroformebenzene binary mixture at 1 atm. Comparison
Kj of calculated results (NRTL model with parameters in Table 1) to the data of Kojima
et al. [17].
Fig. 4 compares calculated relative volatilities for the aceto-
neechloroform binary with data from Kudryavtseva and Kusarev
[19], Kogan and Deizenrot [20], Kojima et al. [17], and Segura et al.
[21] The data from all four sources agree well (except for two
outliers from Kogan and Diezenrot [20]), and strongly indicate that
the maximum model uncertainty is 5%. Fig. 5 provides similar

66

64

62
Temperature ( C)

60

Bubble - Model
58
Dew - Model
Bubble-Kojima-1991
56 Dew-Kojima-1991

54
0 0.2 0.4 0.6 0.8 1
x, y (Acetone) Fig. 4. Relative volatilities of the acetone chloroform system at 1 atm. Comparison of
model results (NRTL model with parameters in Table 1) to the data of Kudryavtseva
Fig. 1. Txy diagram of the acetoneechloroform binary mixture at 1 atm. Comparison of and Susarev [19], Kogan and Diezenrot [20], Kojima et al. [17], and Segura et al. [21]
calculated results (NRTL model with parameters in Table 1) to the data of Kojima et al. The chart also shows (dashed lines) ± 5% deviations from the calculated relative
[17]. volatility.
268 P.M. Mathias / Fluid Phase Equilibria 408 (2016) 265e272
Fig. 5. Relative volatilities of the acetone chloroform system at 1 atm. Comparison of
model results (NRTL model with parameters in Table 1) to the data of Kranich et al.
[22], Konobeev and Gurina [23], Kojima et al. [17], and Kurihara et al. [24] The chart
also shows (dashed lines) ± 5% deviations from the calculated relative volatility.
comparisons between model and data for the acetoneebenzene
system, and here the experimental data come from Kranich et al.
[22], Konobeev and Gurina [23], Kojima et al. [17], and Kurihara
et al. [24] Again we can conclude that the model uncertainty is 5%
or better. Finally, Fig. 6 presents analogous results for the chlor-
oformebenzene binary, with comparisons to the data of Bukata and
Majewski [25], Nagata and Hayashida [26], Loiseleur et al., [27],
Kojima et al. [17], and Chen et al. [28] In this case, the agreement
with the data of Kojima et al. [17] is good, but discrepancies with
the other four sets of data are larger, sometimes exceeding 10%. In
an attempt to understand the cause of the relatively large dis-
crepancies, each of the four datasets were fitted individually, but
the errors did not decrease significantly. Fig. 6 also indicates that
the Kojima et al. [17] relative volatilities are approximately equal to
the average of the other four data sets. We conclude that the
relatively high discrepancies between model and data are due to
experimental random error or bias, and therefore the model un-
certainty for the benzeneechloroform binary is 5%, but recognize
that additional VLE data may be needed to confirm the
Fig. 6. Relative volatilities of the chloroformebenzene system at 1 atm. Comparison of
model results (NRTL model with parameters in Table 1) to the data of Bukata and
Majewski [25], Nagata and Hayashida [26], Loiseleur et al., [27], Kojima et al. [17], and
Chen et al. [28] The chart also shows (dashed lines) ± 10% deviations from the
calculated relative volatility.
Fig. 7. % errors of model K-values (NRTL model with parameters in Table 1) of acetone
in comparison to the ternary VLE data of Reinders and de Minjer [29] and Kojima et al.
[17].
measurement and model uncertainties.
The above comparisons between model and data were per-
formed for binary mixtures. Kojima et al [17] measured ternary data
(total of 53 data points), and Reinders and de Minjer [29] also
presented an extensive set of ternary data for this system (total of
231 data points). In the case of ternary data, percentage differences
between model and data for component K-values are used as the
measure of combined model uncertainty, and in each of the three
cases the percentage differences are plotted against the mole
fraction of the particular component; the idea is that the errors
naturally become small when the component mole fraction ap-
proaches unity, and reach the largest and most meaningful values
when the component approaches infinite dilution. Figs. 7e9 pre-
sent these error charts for acetone, chloroform and benzene,
respectively. Examination of these charts shows that the K-value
differences do decrease and tend to zero when the particular
component's mole fraction approaches unity, and the values at low
concentration give a quantitative indication of the model K-value
uncertainty. These three figures show that the Reinders and de
Minjer data [29] have relatively high uncertainty, and this perhaps
is a consequence of their indirect analytical approach in which the
compositions were deduced by measurements of the mixture
density and refractive index rather than by direct composition
Fig. 8. % errors of model K-values (NRTL model with parameters in Table 1) of chlo-
roform in comparison to the ternary VLE data of Reinders and de Minjer [29] and
Kojima et al. [17].
 P.M. Mathias / Fluid Phase Equilibria 408 (2016) 265e272
Fig. 9. % errors of model K-values (NRTL model with parameters in Table 1) of benzene
in comparison to the ternary VLE data of Reinders and de Minjer [29] and Kojima et al.
[17].
measurements. Our conclusion is that the large experimental un-
certainty of the Reinders and de Minjer data do not enable esti-
mation of model uncertainties, and therefore these uncertainties
have been estimated based upon the Kojima et al. data alone. It is
interesting that Kojima et al. [17] did not seem to recognize the
large uncertainty of the Reinders and de Minjer data since their
paper simply stated: “The predicted vapor compositions of our
parameters agree with Reinders' data in the accuracy of 0.003 mol
fraction on the average.” [17] K-value errors of chloroform (Fig. 8)
average to about zero indicating no model bias. The K-value errors
for acetone (Fig. 7) tend to be negative, while those for benzene
(Fig. 9) tend to be positive, and the signs of these differences
indicate some amount of model (or data) bias. However, all K-value
differences between data and model tend to be small, and we may
conclude that the model agrees with the ternary data of Kojima
et al. [17] to better than 5%.
The purpose of this section has been to fit the model to the
available VLE data for the acetoneechloroformebenzene system at
about 1 atm (50e85  C), and to estimate the combined model
uncertainties. The acetoneechloroform and acetoneebenzene
MIX
F F1
Fig. 10. Two-column flowsheet for the separation of the acetoneechloroformebenzene ternary mixture.
269
binary data indicate that the model uncertainty is better than 5%,
while the chloroformebenzene binary data indicate that the model
uncertainty is larger, perhaps as much as 10%. However, we have
concluded that the larger model uncertainty for the chlor-
oformebenzene binary (Fig. 6) is due to experimental uncertainty,
and that the model uncertainty is likely 5%. The ternary data (using
only the data of Kojima et al. [17] and neglecting the data of
Reinders and de Minjer [29]) indicate that the model K-value un-
certainty is better than 5%. We conclude that the VLE uncertainty of
the model for relative volatility and K-values in the acetonee-
chloroformebenzene system is 5%, but may be as high as 10%. It
should be noted that uncertainty analysis remains subjective, and
ultimately depends upon the judgment of experienced practioners
[6].
4. Separation sequences and residue analysis
As part of an extensive “tutorial on recent advances in the
synthesis of ideal and nonideal distillation-based separation pro-
cesses,” Westerberg and Wahnschafft [12] used residue-curve
analysis [30,31] to synthesize separation schemes for the aceto-
neechloroformebenzene ternary mixture. They proposed a two-
column design and two three-column alternatives. Parodi and
Campanella [13] analyzed the effect of property uncertainties on
the two-column design and one of the three-column designs. Here
we limit our study to the two-column design since we prefer to
focus on the simplest case in order to highlight key elements of the
uncertainty analysis. The feed to the process is a 100 kmol/h liquid
mixture at 1 atm and its bubble point containing 36 mol% acetone,
24 mol% chloroform and 40 mol% benzene. In the two-column
design (for the flowsheet, see Fig. 10), the feed is mixed with a
chosen flow of recycle (stream REC) from the bottoms of the second
column (Column 2), and fed to the first column (Column 1) in which
high-purity acetone is obtained as distillate, and an essentially
acetone-free mixture of chloroform and benzene is taken as the
bottoms product. Column 2 produces high-purity chloroform as the
distillate and high-purity benzene as the bottoms product. Both
columns operate at a pressure of 1 atm, and pressure drops are
assumed to be negligible. The design parameters and specifications
D1
COL1
D2
COL2
B1
B2
REC
REC PROD
270 P.M. Mathias / Fluid Phase Equilibria 408 (2016) 265e272
Table 2
Summary of design parameters and specifications for two-column flowsheet. Both
columns have a total condenser and a partial reboiler.
Parameter Column 1 Column 2
Number of stages 68 60
Feed stage (from top) 26 30
Distillate Rate (kmol/hr) 36 24
Distillate design Acetone mole fraction Chloroform mole fraction
specification in distillate ¼ 0.99 in distillate ¼ 0.98
Variable to achieve Reflux ratio Reflux ratio
design specification
are presented in Table 2. The reflux flow rate (stream REC in the
flowsheet, Fig. 10) is considered to be an independent variable, and
for each specification of the flow rate of stream REC, the reflux
ratios, and hence the reboiler duties, in the two columns are
calculated to achieve their respective design specifications. Here
the sum of the two reboiler duties is taken as indicative of the
operating cost, and hence the simulation enables a quantitative
relationship between the operating cost and the recycle rate. The
Margules perturbation method adds an additional dimension to the
analysis since it quantifies the effect of the property-model un-
certainty on the operating cost.
Following the lead of experts in distillation synthesis, the
ternary residue-curve diagram is used to enable understanding of
the fundamental elements of the separation process [12,13,30,31]
The residue-curve diagram, including some variations that are
described later, is presented in Fig. 11. There is one distillation
boundary that divides the ternary diagram into two separate re-
gions, and this distillation boundary starts at pure benzene and
ends at the acetoneechloroform azeotrope (about 65 mol% chlo-
roform). The distillation boundary is curved, and stays at low
concentration in acetone for the initial portion as it moves away
from pure benzene. By mass balance, the two products of a distil-
lation column with a feed to the left of the distillation boundary fall
Fig. 11. Residue curve diagram for the acetoneechloroformebenzene ternary mixture
at 1 atm. Three distillation boundaries are shown: Distillation Boundary-0 (thick black
line) corresponds to the base case with the NRTL parameters in Table 1, and all three
perturbation parameters set to zero; Distillation Boundary-1 (dashed black line) has
perturbation parameters dA ¼ 0.2, dC ¼ 0.2, dB ¼ 0.2; and Distillation Boundary-2
(thick green line) has perturbation parameters dA ¼ 0.2, dC ¼ 0.2, dB ¼ 0.2. The
red lines show the feed (F), the combined feed to column 1 (F1), and the products from
the flowsheet (D1, D2 and B2). (For interpretation of the references to color in this figure
legend, the reader is referred to the web version of this article.)
on a straight line that goes through the feed composition and ends
on the distillation boundary, and it is desired that the distillate (D1)
has 99 mol% acetone. With zero flow in REC it can be seen that the
bottoms from Column 1 will have too high a concentration of
acetone, and this renders the composition specification on Column
2 impossible to achieve (Table 2); to “see” this result imagine the
intercept of a straight line between D1 and F extrapolated to the
distillation boundary. A certain minimum flow of REC is needed to
attain the three product specifications, and the reboiler duties
generally rise with an increase in REC flow because the vapor and
liquid flows in the two distillation columns will increase.
Table 3 presents the results of a study in which the recycle flow
rate has been varied. In general, increasing the recycle flow rate
causes the recycle ratios and also the reboiler duties of both col-
umns to increase. However, if the recycle flow rate is decreased
below about 55 kmol/h, the recycle ratio in Column 1 rises with a
decrease in reflux rate. This result follows because decrease in
reflux rate increases the mole fraction of acetone in the bottoms
(see the upper part of the distillation boundary in Fig. 11), and an
increase in reflux ratio is needed to meet the design-spec of 99 mol
% acetone in the distillate of Column 1. In fact, when the reflux flow
rate drops below z43 kmol/h, it becomes impossible to meet the
purity specification in Column 2 even with an infinite reflux ratio;
again, imagine a straight line from pure benzene to the distillation
boundary and its intercept on the bottom axis in Fig. 11).
The goal of this paper is to add the consequences of property
uncertainties onto these effects, and these analyses are presented
in the next section.
5. Quantification of uncertainties in the two-column design
In order to perform the uncertainty analysis, the perturbation
parameters for each of the three components (di ) need to be related
to model uncertainties. Fig. 12 uses Eq. (2) to show the percentage
change in infinite-dilution activity coefficient with the perturbation
parameter d at three values of the model infinite-dilution activity
coefficient. The three values chosen for the model infinite-dilution
activity coefficients are representative of the three binaries of the
acetoneebenzeneechloroform system (see Table 1). Fig. 12 shows
that the values of d in the range of 0.1e0.2 will give a perturbation
in
 the∞ range
 of 5% for the present ternary systems. In particular for
lnðg Þ values of 0.2 and 0.8, d values for a 5% variation in infinite-
i
dilution activity coefficients are 0.2 and 0.1, respectively.
We now turn to the sign of d. In order to be conservative
d should be negative (i.e., lower the K-value) when the component
is a light, and d should be positive when the component is a heavy.
In Column 1, acetone is a light, while chloroform and benzene are
heavies, while in Column 2, acetone and chloroform are lights and
benzene is the heavy. To control the several perturbations with a
single parameter, a perturbation scheme has been set up as follows:
Column 1 : dA ¼ d0 ; dC ¼ d0 ; dB ¼ d0 (6)
Column 2 : dA ¼ d0 ; dC ¼ d0 ; dB ¼ d0 (7)
where the subscripts A, C and B refer to acetone, chloroform and
benzene, respectively, and d0 is the single perturbation parameter.
We note that the two separate perturbations in Eq. (6) and Eq. (7)
have been implemented in Aspen Plus using the “multiple data-
set” capability [32].
Fig. 11 shows the effect of the two perturbations (Eqs. (6) and
(7)) on the distillation boundary when d0 is equal to 0.2. The
distillation boundaries do not move by very much and certainly
don't move qualitatively, which tells us that the effect of the per-
turbations on the design flowsheet structure will be small and
 P.M. Mathias / Fluid Phase Equilibria 408 (2016) 265e272 271

Table 3
Effect of recycle flow rate. Base case with NRTL parameters in Table 1, and no perturbation.

Recycle rate (kmol/hr) Column 1 Column 2 Total reboiler duty (Gcal/hr)

Reflux ratio Reboiler duty (Gcal/hr) Reflux ratio Reboiler duty (Gcal/hr)

43 4.97 1.528 9.17 1.730 3.259

44 4.85 1.496 9.23 1.739 3.236
46 4.73 1.467 9.38 1.766 3.233
48 4.69 1.455 9.56 1.795 3.250
50 4.66 1.450 9.74 1.826 3.276
55 4.66 1.449 10.21 1.906 3.355
60 4.68 1.454 10.69 1.987 3.441
70 4.75 1.472 11.64 2.148 3.620
80 4.83 1.492 12.58 2.308 3.800
90 4.92 1.514 13.51 2.466 3.979
100 5.00 1.536 14.43 2.622 4.158

have been shown as percentage changes relative to the base case:

Fig. 12. Percentage change in infinite-dilution activity coefficient with perturbation
parameter, d, at various values of the absolute value of the natural logarithm of the
model infinite-dilution activity coefficient. Eq. (2) is used for the calculation.
likely negligible, and hence distillation synthesis conclusions are
unlikely to be affected by the magnitude of the property un-
certainties expected in the present study.
Fig. 13 summarizes the results of the perturbation study. At each
value of the recycle flow, the change in total reboiler duty has been
shown as a function of the perturbation parameter, d0. All results
recycle flow of 50 kmol/h and no property perturbation. The total
reboiler duty increases monotonically with an increase in d0 and
this is because the perturbations have been performed in a con-
servative manner (Eqs. (6) and (7)). A simple result is that an in-
crease in d0 by 0.2 (about 5% uncertainty) increases the total
reboiler duty by about 15%. A new insight is that if d0 increases to a
particular value, the two distillation columns will not be able to
meet the purity specifications, and this minimum value of d0 in-
creases with the reflux flow rate. For example, if d0 is 0.2 a recycle
flow of 50 kmol/h will be inadequate, and it should be raised to
about 55 kmol/h in order to ensure that the system can meet the
purity specifications. Similarly, if the K-value uncertainty were as
high as 20%, the reflux flow would need to be 80 kmol/h, and the
total reboiler duty would rise by as much as 50%. These decisions
will, in industrial practice, be made by the design team, but the
clear value of the perturbation scheme (in particular, charts such as
Fig. 13) is that the design decisions are greatly facilitated by cal-
culations that can be performed with modest effort.
Parodi and Campanella [13] concluded that “For the two-column
sequence, analysis done to each column alone or to the whole
sequence gave the same results.” In contrast, the present study
clearly shows that property uncertainty does affect the two-column
design, and the proposed perturbation scheme provides guidance
on how the design must be modified to allow for the property
uncertainties.
6. Summary and conclusions

The conclusions of this study may be summarized as follows:

 The first step in the distillation design procedure is careful data

analysis to estimate the property uncertainties, and here the
property uncertainty uncertainties of concern are K-values and
relative volatilities. This step is essential and cannot be elimi-
nated or even shortchanged.
 The K-value uncertainties have less effect on synthesis (flow-
sheet structure) than on equipment performance (utility
consumption).
 The present Margules perturbation method provides semi-
quantitative insight into distillation operation. The two identi-
fied benefits are that it quantifies the effect on utility con-
sumption and identifies limits on operating variables to meet
design criteria (specifically minimum recycle flow).

Researchers in distillation synthesis [12] have typically identi-

Fig. 13. Effect of perturbation parameter, d0, on the percentage change in the total
reboiler duty at various values of the reflux rate (in kmol/hr). The solid circles repre-
sent the point beyond which the composition specifications cannot be attained at a
particular recycle flow rate.
fied two stages in the distillation design process: (1) Discover the
possible system structures; and (2) Perform detailed design (e.g.,
number of stages, optimum tray location, etc.). Uncertainty analysis
272 P.M. Mathias / Fluid Phase Equilibria 408 (2016) 265e272
is not currently included in Step 2, but this omission should be
rectified.
References
[1] S. Kim, J.W. Kang, K. Kroenlein, J. Magee, V. Diky, M. Frenkel, Chem. Eng. Educ.
47 (2013) 48e57.
[2] S. Macchietto, G. Maduabeuke, R. Szcepanski, Fluid Phase Equilib. 29 (1986)
59e67.
[3] M.E. Reed, W.B. Whiting, Chem. Eng. Commun. 124 (1993) 39e48.
[4] P.M. Mathias, J. Chem. Eng. Data 59 (2014) 1006e1015.
[5] Q. Dong, R.D. Chirico, X. Yan, X. Hong, M. Frenkel, J. Chem. Eng. Data 50 (2005)
546e550.
[6] R. Chirico, Implementation of Standards for the Reporting of Measurement
Uncertainty for the Worldwide Thermodynamics Community, in: Presented at
the Nineteenth Symposium on Thermophysical Properties, Boulder, CO,
21e16 June 2015.
[7] J.F. Brennecke, A.R.H. Goodwin, P. Mathias, J. Wu, J. Chem. Eng. Data 56 (2011),
4279e4279.
[8] R.D. Weir, J.P.M. Trusler, J. Padua, J. Chem. Thermodyn. 43 (2011), 1305e1305.
[9] P.T. Cummings, Th W. de Loos, J.P. O'Connell, Fluid Phase Equilib. 307 (2011)
iveiv.
[10] V. Rives, C. Schick, S. Vyazovkin, Thermochim. Acta 521 (2011), 1e1.
[11] W.M. Haynes, D.G. Friend, A. Mandelis, Int. J. Thermophys. 32 (2011)
1999e2000.
[12] A.W. Westerberg, O. Wahnschafft, Adv. Chem. Eng. 23 (1996) 63e170.
[13] C.A. Parodi, E.A. Campanella, Lat. Amer. Appl. Res. 39 (2009) 33e40.
[14] H. Renon, J.M. Prausnitz, AIChE J. 14 (1968) 135e144.
[15] O. Redlich, J.N.S. Kwong, Chem. Rev. 44 (1949) 233e244.
[16] R.L. Rowley, W.V. Wilding, J.L. Oscarson, Y. Yang, N.A. Zundel, DIPPR® Data
Compilation of Pure Chemical Properties, Design Institute for Physical Prop-
erties, AIChE, New York, NY, 2010.
[17] K. Kojima, K. Tochigi, K. Kurlhara, M. Nakamichi, J. Chem. Eng. Data 36 (1991)
343e345.
[18] M. Frenkel, R.D. Chirico, V. Diky, X. Yan, Q. Dong, C. Muzny, J. Chem. Inf. Model
45 (2005) 816e838.
[19] L.S. Kudryavtseva, M.P. Susarev, Zh. Prikl. Khim. (Leningrad) 36 (1963)
1231e1237.
[20] L.V. Kogan, I.V. Deizenrot, Zh. Prikl. Khim. (Leningrad) 48 (1975) 2757e2759.
[21] H. Segura, A. Mejia, R. Reich, J. Wisniak, S. Loras, Phys. Chem. Liq. 41 (2003)
283e301.
[22] W.L. Kranich, R.E. Wagner, D.W. Sundstrom, H. Slotnick, Ind. Eng. Chem. 48
(1956) 956e960.
[23] B.I. Konobeev, T.V. Gurina, Zh. Prikl. Khim. (Leningrad) 47 (1974) 2492e2495.
[24] K. Kurihara, H. Hori, K.J. Kojima, J. Chem. Eng. Data 43 (1998) 264e268.
[25] M. Bukata, J. Majewski, J. Przem. Chem. 8 (1952) 551.
[26] I. Nagata, H.J. Hayashida, Chem. Eng. Jpn. 3 (1970) 161e166.
[27] B. Loiseleur, P. Clechet, J.C. Merlin, Bull. Soc. Chim. Fr. 11 (1968) 4390e4398.
[28] G. Chen, Q. Wang, Z.-M. Ma, X.-H. Yan, S.-J. Han, J. Chem. Eng. Data 40 (1995)
361e366.
[29] W. Reinders, C.H. de Minjer, Rec. Trav. Chim. 59 (1940) 369e391.
[30] E.J. Hoffman, Azeotropic and Extractive Distillation, Wiley (Interscience), New
York, 1964.
[31] M.F. Doherty, J.D. Perkins, Chem. Eng. Sci. 33 (1978) 281e301.
[32] For further descriptions on how to use the multiple dataset capability refer to
the Aspen Plus documentation available from Aspen Technology, Inc.