Measurement and Control
How Good Are
Your Data?
Some thoughts on the measurement
and interpretation of vapor-liquid equilibrium
Scott D. Barnicki,
Eastman Chemical Co.
V apor-liquid equilibrium (VLE) is the foun-
dation for a variety of separation methods
that are essential to the chemical process in-
dustries. These include simple distillation,
azeotropic distillation, extractive distillation and flash
operations. The use of process simulators to predict the
performance of these unit operations has become
widespread in the last decade, but often little effort is
expended to assure that the underlying VLE data are of
acceptable quality. Considerable published data are
available (1, 2) and most simulators have regression fea-
tures to treat raw experimental data. In the absence of
data, estimation via group contribution methods, such
as UNIFAC, is often possible. However, the simulators
provide few checks on the quality of their output. Al-
though it may be tempting to accept estimated or pub-
lished data, or regression analyses of such data at face
value, there is no guarantee on the data’s quality.
Measuring behavior
Six possible variables can be measured in order to
characterize the behavior of a binary vapor-liquid sys-
tem. They are temperature, pressure, two liquid mole
fractions, and two vapor mole fractions. Only one of
each of the vapor and liquid mole fractions is indepen-
dent (they must sum to unity, although both are usually
measured as a consistency check). This leaves four in-
dependent variables. However, the phase rule states
that the system is completely specified when any two
of the four independent variables are fixed.
58 www.cepmagazine.org June 2002 CEP
In practice, this means for each data point, a
charge of arbitrary composition of the binary pair
(usually liquid) is placed in an experimental appara-
tus at either a fixed temperature or pressure. During
the experiment, the second phase (usually vapor) is
generated by the addition of heat or adjustment of
pressure, and the system is allowed to come to a
true equilibrium at controlled temperature and pres-
sure conditions before sampling. Changing the com-
position of the initial charge, with either the temper-
ature or pressure kept constant between data points,
generates a complete data set. Development of ef-
fective, reproducible methods for achievement of
true equilibrium has continued unabated for the past
75 years. Two of the more important techniques can
be broadly classified as static and circulation meth-
ods (3, 4).
In the static methods, a liquid mixture typically is
agitated in a closed, evacuated, thermostatted vessel
and allowed to come to equilibrium with its generat-
ed vapor. In their natural mode of operation, static
cells provide isothermal data, with a complete P-x-y
data set generated by varying the composition of the
initial mixture. Vapor mole fractions are difficult to
measure and tend to be less reliable. At low pres-
sures, the vapor volume is quite small and the pro-
cess of obtaining a sample of sufficient quantity for
analysis may disturb the equilibrium. Although cer-
tainly possible, the direct solution of the equilibrium
expressiion:

 RT (
yi Pφˆi = xiγ iφisat exp i P − Pi sat 
v
 ) (1)
for the activity coefficients may give misleading results be-
cause of the inaccuracies in the vapor mole fractions.
A number of indirect methods of correlating isothermal
P-x-y data have been formulated based on the simplified
form of the Gibbs-Duhem equation:
− H E dT V E dP
∑ xi d ln γ i = + (2)
RT 2 RT
in which the right-hand side is set to zero (e.g., the Barker
method (4)). Under isothermal conditions, the heat of mix-
ing term, HE, is zero. With the weak dependency of activity
coefficients on pressure, the excess volume term is usually
negligible (typically less than 4% of the magnitude of the
tion. The regressed bij parameters often show some variation
with temperature that is not captured by the simple form of
Eq 5. This is especially true if the binary pair is wide-boiling
or if the fit is extrapolated to temperatures outside of the
range covered in the data set. Extrapolation to extreme tem-
perature must be done with care, but in practice, the simple
form of Eq. 5 usually is adequate and is markedly superior
to those for expressions with no built-in temperature depen-
dency (e.g., Van Laar and Margules). A better fit to experi-
mental data is possible when both the aij and bij are taken as
adjustable parameters. Of course this means that the value of
aij loses its physical significance. A good or bad fit does not
necessarily indicate the quality of the data.
Minimization of an objective function of the form:
( )
n 2
Objective function = ∑ Picalc − Pi exp (6)
i =1

activity coefficient (5)). Eq. 1 can be written for each com- controls the regression algorithm. For the first iteration, the
ponent of a binary mixture and solved for the vapor mole fugacity terms in Eq. 3, φisat and φˆi, are assumed to be unity.
fraction. Upon summation, and multiplication by the pres- Values for the vapor mole fractions are back-calculated and
sure, the resulting expression is: used in the succeeding iteration until convergence is
achieved and the best fit of the parameters is obtained.
 φ sat 
φˆ
 i  RT
(
P = xi γ i Pi sat  i  exp i P − Pi sat  +

v
 )
Nomenclature
 φ sat 
 φˆi 
v
 RT (
xi γ i Pi sat  i  exp i P − Pi sat 
 ) (3)
aji, aij, bji, bij
Cpoyn
= adjustable Wilson parameters
= Poynting correction
DRK = difference term in the Redlich-Kister area test
After substitution of the binary Wilson activity coeffi- DW = deviation function defined by Wisniak
cient equation: HE = heat of mixing
∆HiBP = heat of vaporization at system pressure
M = Herrington empirical boiling range parameter
x j Λ ji
GiE
RT
(
= ln γ = 1 − ln xi + x j Λij − )xi
−
xi + x j Λij x j + xi Λ ji
( 4) P
R
= pressure
= ideal gas constant
T = temperature
for the activity coefficients, all terms on the right-hand side VE = volume change of mixing
of Eq. 3 are known except for the Wilson molecular interac- xi = liquid mole fraction
tion terms, Λij and Λji. They are unique for each binary pair yi = vapor mole fraction
and must be determined from experimental data. Note that Greek letters
Λjj = Λii = 1. In the original derivation by Wilson, these inter- γ∞ = infinite-dilution activity coefficient
action terms are dependent on the liquid molar volumes, νi γi = activity coefficient
and νj, of the species and on the system temperature, leaving Λij, Λji = molecular interaction terms for the Wilson equation
two adjustable parameters, bij and bji, for each binary pair (5). νi = liquid molar volumes
φisat = saturated vapor fugacity
φˆi = mixture vapor fugacity
bij vj
ln Λij = aij + where aij = (5) Subscripts
T vi
i = component i
j = component j
When aij is not adjustable, the Wilson equation always max = maximum
predicts that the activity coefficient increases monotonically min = minimum
n = point n
as xi approaches infinite dilution. Although relatively rare,
some real systems exhibit a maxima. Superscripts
With the Wilson equation, all temperature-dependent be- BP = boiling point
havior is lumped into the interaction term. This is not neces- calc = calculated
sarily an accurate representation of the true effect of temper- exp = experimental
ature on the heat of mixing term in the Gibbs-Duhem equa- sat = saturated

CEP June 2002 www.cepmagazine.org 59

Measurement and Control

In the circulation methods for measuring VLE, a vessel is Table 1. Physical properties for an acetone-water system (11).
partially filled with liquid that is brought to a boil by the ad-
Water Acetone
dition of heat. The generated vapors are disengaged from the
liquid, conveyed to a separate condensing surface, and even- Liquid Molar Volume, cc/gmol 18.07 74.05
Critical Volume, cc/gmol 55.9 208.45
tually returned to the pot. After some circulation time, the Critical Temperature, K 647.3 508.1
system will reach equilibrium and small samples of the pot Critical Compressibility 0.229 0.232
liquid and condensed vapor are collected. The separation of Critical Pressure, torr 165,376 35,264
vapor from the liquid before condensation and recirculation Pitzer Acentric Factor 0.344 0.309
Heat of Vaporization 9,717 6,960
of the condensed vapor are critical. If the vapor is condensed at Normal Boiling Point, cal/gmol•K
in the same location where the equilibration occurs, then
Antoine Constants
some fractionation will occur. The results will be skewed,
which is why a simple flask with a reflux condenser tend to A 18.3036 16.6513
B 3,816.44 2,940.48
give give poor results for VLE measurements. C –46.13 –35.93
In their natural mode of operation, circulating stills (3,
Note: ln Pisat = A – [B / (T + C)], where T is the temperature
4, 6) provide isobaric data, with a complete T-x-y data set
(K) and P is the pressure (torr).
generated by varying the composition of the initial mix-
ture. Operation of a circulating-still is quite simple and the
proper sampling technique is relatively easy to master. Two In practice, knowledge of the temperature-dependency of
modes of data regression are possible. Vapor analysis is activity coefficients is required for accurate representation
more reliable (although still somewhat difficult) because of of distillation behavior, especially with wider boiling mix-
the larger volume of condensed vapor. Eq. 1 can be solved tures. Therefore, application of either isothermal or isobaric
directly to determine the values of the activity coefficients data reduces ultimately to the same problem: How can the
for each set of measurements. This is the simplest and most temperature-dependency of activity-coefficient data best be
straightforward method for finding the activity coefficients. represented? It is necessary to measure two or three isother-
A less convenient approach for isobaric data is to use a mal data sets to capture the temperature effect, while only
Barker-type method, solving for temperature. Eq. 3 is not one set of isobaric data achieves the same end. If care is
implicit in temperature, so the regression of the Wilson pa- taken, either the isothermal or the isobaric method will give
rameters must be accomplished simultaneously with a so- data of acceptable accuracy. For distillation applications,
lution for the temperature by an equation-solving routine. with significant uncertainties in efficiency measurements
The objective function is the same form as Eq. 6, except and rather poor accuracy of flooding correlations, the de-
with pressure replaced by temperature. This is not an im- signer must build in some safety factor and, for ideal or
possible task by any means with modern software, but is moderately non-ideal systems, will not see much difference
more complicated than the isothermal case. in the final column design. When you can only afford the
In both approaches, one implicitly lumps all of the tem- time and expense of measuring one complete set of data
perature dependence into the chosen activity coefficient ex- and the ultimate objective is a distillation design, it is the
pression, rather than accounting for it separately via the opinion of the author that it often more practical to measure
heat of mixing term in Eq. 2. Al-
though not as large an effect as on the
vapor pressure terms, the tempera- Table 2. Experimental data set #1 — T-x-y data for an
ture-dependency of activity coeffi- acetone-water system at 760 torr (12).
cients usually cannot be ignored. It is Point Temperature, °C Liquid Mole Fraction Vapor Mole Fraction
common for the heat of mixing term Acetone Acetone
to account for at least a 6–10% varia- 1 57.1 0.850 0.917
tion in the activity coefficients. Usu- 2 57.4 0.793 0.900
ally the direct solution of Eq. 1 is the 3 58.5 0.661 0.860
most convenient and practical ap- 4 58.9 0.609 0.847
proach, as heat-of-mixing data are 5 59.5 0.538 0.840
rarely available or measured. When 6 59.7 0.506 0.837
an activity coefficient expression is 7 60.0 0.444 0.832
8 61.1 0.300 0.809
used that has a built-in temperature- 9 61.8 0.266 0.802
dependence (e.g., the Wilson equa- 10 66.2 0.120 0.756
tion), the error associated with lump- 11 76.5 0.041 0.585
ing the heat of mixing term into the 12 83.0 0.023 0.462
activity coefficient expression is usu- 13 87.8 0.010 0.335
ally acceptable.

60 www.cepmagazine.org June 2002 CEP

 Table 3. Calculated fugacity coefficients and Poynting corrections for data set #1.
Point Saturated Vapor Fugacity Mixture Fugacity Poynting Correction
Water Acetone Water Acetone Water Acetone

1 0.997 0.965 0.980 0.966 1.001 1.000

2 0.997 0.965 0.981 0.966 1.001 1.000
3 0.997 0.964 0.981 0.967 1.001 1.000
4 0.997 0.964 0.981 0.967 1.001 1.000
5 0.996 0.963 0.981 0.967 1.001 1.000
6 0.996 0.963 0.981 0.967 1.001 1.000
7 0.996 0.963 0.981 0.967 1.001 1.000
8 0.996 0.962 0.982 0.968 1.001 1.000
9 0.996 0.961 0.982 0.968 1.001 0.999
10 0.996 0.958 0.983 0.969 1.000 0.999
11 0.994 0.948 0.985 0.972 1.000 0.998
12 0.993 0.941 0.986 0.974 1.000 0.997
13 0.991 0.936 0.987 0.975 1.000 0.996

an isobaric set. This is especially true if the measurements the activity coefficients and regress the adjustable Wilson
can be made at the nominal pressure at which the distilla- equation parameters, bij and bji. Many forms of the objective
tion will be operated. function for controlling the regression have been suggested
in the literature (9). The objective function chosen here is the
Regression of data form used for isobaric data developed by Gmehling and
The steps involved in data regression and consistency Onken for the Dechema VLE and LLE data collection (1):
testing are demonstrated for the binary system acetone- 2
water for both isobaric and isothermal data sets. The system  γ calc 
Objective function = ∑ 1 − iexp  (7)
exhibits a severe pinch at the acetone-rich end of the x-y di-  γi 
agram. The activity of water in high concentrations of ace-
tone is quite large, substantially increasing water’s volatility The regression capabilities built into most spreadsheets,
as compared to the ideal liquid case. First, a detailed analy- such as Microsoft Excel, can be used to solve this mini-
sis of an isobaric data set, with and without the inclusion of mization problem. Pairs of bij and bji parameters are not
vapor-phase non-idealities and the Poynting correction, is unique and a range of values will fit the data equally well.
done. The results from this analysis are compared to anoth- The actual pair determined in the regression calculation de-
er set of data ostensibly measured under the same condi- pends on the algorithm and initial guesses. Unfortunately,
tions with similar equipment. Finally, the regression of the bij parameters have no real physical significance, so
isothermal data is compared to the results from the isobaric they are hard to guess a priori. Typical values range from
work. Pertinent physical properties for acetone and water about 3,500 to –500 with little observable pattern based on
are given in Table 1 (7). type of system or deviations from ideality. Initial guesses
Data set #1 was measured at 760 torr with an Othmer- can be found using:
type vacuum-jacketed circulation-still with a liquid holdup bij
of about 500 mL (8). This is an example of the good-quali- 0 = 1 − ln γ i∞ − aij − BP
ty data that can be measured with a circulating-still. The Tj
complete T-x-y data set is given in Table 2. To illustrate the  T BP   bij  ∞

− exp a ji + i BP 1 − a ji − exp aij + T BP  − ln γ i  
most general procedure, all terms in Eq. 1 will be calculat- (8)
ed rigorously by the direct method. Results assuming ideal-  Tj   i   
gas-phase behavior (or equivalently that φˆi ≈ φisat), are in-
cluded for comparison. Mixture and saturated vapor fugac- This equation will have multiple roots only when both of
ities were calculated using the Pitzer method (9). Although the infinite-dilution activity coefficients are less than unity.
water is rather polar, the Pitzer correlation yields accept- The second parameter is solved by substitution:
able results. Results for each data point are given in Table
3. As expected the acetone-water system does not exhibit   bij  
b ji = Ti BP 1 − a ji − exp aij + BP  − ln γ ∞j  (9)
much deviation from ideality at 760 torr. The Poynting cor-  
 Ti 
rection is essentially negligible. The fugacity coefficients
of water and acetone deviate only about 0.5 to 6 % from
ideal behavior. The goal-seek function built into most spreadsheets can
It is now possible to solve for the experimental values of readily solve such a set of equations. Although the experi-

CEP June 2002 www.cepmagazine.org 61

Measurement and Control

Table 4. Bubble temperature and activity coefficient values for data set #1.
Point Bubble Temperature, K Experimental Activity Coefficients Regressed Activity Coefficients
Experimental Regressed Water Acetone Water Acetone

1 57.1 57.4 3.17 1.05 3.36 1.03

2 57.4 57.9 2.73 1.10 2.93 1.06
3 58.5 59.2 2.22 1.21 2.23 1.18
4 58.9 59.7 2.06 1.28 2.03 1.24
5 59.5 60.3 1.77 1.41 1.80 1.36
6 59.7 60.6 1.68 1.48 1.71 1.42
7 60.0 61.2 1.51 1.66 1.56 1.57
8 61.1 62.7 1.30 2.31 1.29 2.17
9 61.8 63.1 1.25 2.52 1.24 2.40
10 66.2 66.7 1.05 4.60 1.07 4.49
11 76.5 75.9 1.06 7.70 1.01 8.17
12 83.0 82.4 1.04 9.05 1.00 9.77
13 87.8 90.3 1.05 13.27 1.00 11.36

both the rigorous and ideal vapor cases. Results are virtual-
Table 5. Wilson bij and bji parameters for data set #1.
ly indistinguishable. The main differences are in the value
Rigorous Ideal Vapor of infinite-dilution activity coefficients. Once the Wilson
aA-W –1.411 –1.411 parameters have been determined, additional T-P-x-y plots
aW-A 1.411 1.411 can be constructed at any desired conditions. However,
Initial Guess for bA-W –427.81 –419.16
Initial Guess for bW-A –537.17 –528.87
caution must be exercised when the extrapolating far out-
Regressed Value for bA-W –267.82 –244.63 side of the temperature range of the data set.
Regressed Value for bW-A –710.65 –719.70
Sum of Squares of Error 0.0629 0.0574 Verification of data consistency
Calculated Infinite Dilution Experimental data should be both internally consistent
Activity Coefficients and meet the thermodynamic constraints laid out in the
Acetone 12.40 11.83 Gibbs-Duhem equation. This equation forms the basis for
Water 5.15 5.25 one type of test for the thermodynamic consistency of VLE
data. With isothermal data, the heat of mixing term drops
mental infinite-dilution activity coefficients are unknown, out of Eq. 2 and the effect of pressure on the activity coef-
essentially any experimental values from the dilute compo- ficients is usually minor. Integration of the simplified form
sition regions will suffice to give a reasonable initial guess. across the whole composition range gives:
When the selected values are far removed from the true in-
finite-dilution values, the initial guesses and final values γ 
may not be of the same sign or magnitude. For our system, 0 = ∫01 ln i  dxi (10)
γ j 
the initial guesses are based on the activity coefficients cal-
culated from data points 1 and 13. Experimental activity The value of the logarithmic function passes from posi-
coefficients, bubble point temperatures, and regression re- tive to negative as the composition space is traversed. If the
sults are given in Table 4. The initial guesses, final values, negative area of the integral is equal to the positive area,
and sum of errors from the regression analysis are reported then the data are consistent by this test. The method,
in Table 5 for both the rigorous and ideal vapor cases. Re- known as the Redlich-Kister area test, has been used exten-
sults between the two cases are comparable. sively for many years. Gmehling and Onken (8) suggest
The goodness of fit of the data is checked by back-cal- that the difference between the positive (P) and negative
culation of temperature, vapor mole fractions, fugacities, (N) areas should be less than a few percentage points,
and activity coefficients for each experimental liquid mole where the difference is defined as
fraction and the system pressure. With isobaric data, this
entails an iterative solution of the bubble point equation for P− N
DRK = 100 (11)
each experimental point, using any number of well-be- P+ N
haved objective functions, such as that by Rachford and
Rice (9). Again this can be done readily with a spreadsheet Van Ness, et. al. (10) have shown that pressure cancels
goal-seek function. Figure 1 shows the experimental data when taking the ratio of the activity coefficients in Eq. 10.
on the same graph with the calculated x-y envelope for Thus, the Redlich-Kister area test does not make use of

62 www.cepmagazine.org June 2002 CEP

 1 on observation of only a few systems and its efficacy has
never been rigorously reviewed.
Wisniak recently proposed a more thermodynamically
sound isobaric-consistency test based on the relationship be-
Vapor Mole Fraction, Acetone

0.8
tween a mixture’s Gibbs excess energy and boiling point, via
the Clausius-Clapeyron equation. The Wisniak test can be
used as both a point-by-point check and an area test. Infor-
0.6
mation required for the evaluation includes the experimental
activity coefficient, temperature, and liquid mole fraction
data, heat of vaporization as a function of temperature, and
0.4 Y Experimental boiling points at the system pressure. Each data point, n, is
Rigorous
Ideal
consistent if the functions Ln and Wn are equal. Where:
0.2
Ti BP xi,n ∆sio,n + TjBP x j ,n ∆s oj ,n
Ln = − Tn =
xi,n ∆sio,n + x j ,n ∆s oj ,n

[( ) ]=W
0
0 0.2 0.4 0.6 0.8 1 RTn xi,n ln γ i,n + x j ,n ln γ j ,n − wn
(13)
xi,n ∆sio,n x j ,n ∆s oj ,n
n
Liquid Mole Fraction, Acetone +

■ Figure 1. x-y diagram for acetone-water data set #1 at a system pressure

of 760 torr. TiBP and TjBP are the boiling points of the pure components
i and j at the system pressure. The wn and ∆s°j, n terms are
pressure data and is inconclusive in many instances. Van defined as:
Ness, et. al. outline a point-to-point test that measures the
 yi,nφˆi,n   y j ,nφˆ j ,n 
internal consistency of the data. The difference between the w = xi,n ln +
 x j ,n 
ln  (14)
 xi,nφi,n C poyn   x j ,nφ j ,n C poyn 
sat sat
experimental and calculated vapor mole fractions is plotted
vs. the liquid mole fraction. The data should scatter around
the origin, if the error in the experimental measurements is Cpoyn is the Poynting correction, φˆi,n and φi,nsat are the mixture
truly random. Otherwise, there may be a systematic error and pure component fugacity coefficients at point n, and:
in either the experimental technique, or a poor fit of the ac-
tivity coefficient, fugacity coefficient, and/or vapor pres- ∆HiBP
sure models. Sometimes it is difficult to tell which is the ∆sio,n = (15)
Ti BP
cause. Gmehling and Onken suggest that the data are prob-
ably internally consistent when the average error in vapor The heat of vaporization of at the system pressure,
mole fractions is less than or equal to about 0.01. Outliers ∆HiBP, can be found from the heat of vaporization at the
are easy to identify. At times, a few points may not meet normal boiling point using the Watson correlation (7). The
this criterion, but the overall set may still be of adequate area test is derived by integrating the Ln and Wn functions
quality. over the composition space:
The integrated Gibbs-Duhem equation is of limited value
for the checking the quality of isobaric data. The heat of
L = ∫01 Ln dxi = ∫01 Wn dxi = W (16)
mixing term in Eq. 2 usually cannot be ignored. In an early
semi-empirical isobaric consistency test by Herrington (8)
the heat of mixing term was correlated to the total boiling- Due to experimental error, real data will deviate from the
point range of the mixture, estimated empirically as M: criteria set forth in Eqs. 13 and 14. Wisniak defines a devi-
ation function, Dw, to evaluate the quality of data.
∆Tmax
M = 150 (12)
Tmin L−W
Dw = 100 (17)
L+W
where ∆Tmax is the difference between the lowest and high-
est boiling point of the mixture and Tmin is the lowest mea-
sured boiling point of the mixture. Herrington suggested When Dw is less than about 3 to 5, the data are assumed to
that the same DRK parameter be used as in the area test. The be consistent. This holds for individual points, as well as
data set is judged to be consistent when DRK–M ≤ 10. The for the overall L and W integrals. A complete derivation is
test is simple to use, but the empirical component is based given by Wisniak (11).

CEP June 2002 www.cepmagazine.org 63

Measurement and Control

3 35

2 Data Set #1 30
Data Set #2
25
ln (γ /γ )

1
W

20
A

Li or Wi
0
15
Li Data Set #1
-1 10 Wi Data Set #1
Li Data Set #2
-2 5 Wi Data Set #2
0 0.2 0.4 0.6 0.8 1
0
Liquid Mole Fraction, Acetone
-5
0 0.2 0.4 0.6 0.8 1
■ Figure 2. Results from the Redlich-Kister area test.
Liquid Mole Fraction, Acetone

0.1 ■ Figure 4. Results from the Wisniak area test.

Error in Vapor Mole Fraction, Acetone

0.08 35
30
Wisniak Point Deviation Term, DW

0.06 Inconsistent Data

Data Set #1 25
Data Set #2
0.04 Data Set #1
20 Data Set #2

0.02 15

10
0
Consistent Data
5
-0.02
0
-0.04
0 0.2 0.4 0.6 0.8 1 -5
0 0.2 0.4 0.6 0.8 1
Liquid Mole Fraction, Acetone
Liquid Mole Fraction, Acetone

■ Figure 3. Results from the ∆y point test. ■ Figure 5. Results from the Wisniak point test.

How does data set #1 stand up against these criteria?
Plots of the Redlich-Kister area test, ∆y point test, Wisniak
area test, and Wisniak point test are shown in Figures 2–5,
respectively. The values of the deviation functions, Dw, DRK,
M, and average ∆y are given in Table 6. The N and P areas
for the Redlich-Kister-Herrington test were found by fitting
a 5th-order polynomial to the data points and integrating
analytically. The data set passes both area tests. The average
error in vapor mole fraction was within the allotted value of
0.01, and the point value Dwi’s were within two. However,
both the Wisniak and ∆y point test seem to indicate some
slight systematic error. According to the ∆y point test, two
points in the dilute acetone region are probably less reliable.
As we stated earlier, the dilute regions are the most impor-
tant for activity coefficient information, but are also the
most susceptible to the error. Overall this is a good set of
data, which can be used with reasonable confidence.
Gmehling and Onken (1) also report another set of iso-
baric acetone-water data at 760 torr measured by York and
Holmes (12). Data set #2 was measured with an older ver-
sion of an Othmer still without a vacuum jacket. The T-x-y
data are given in Table 7. Obvious problems are apparent

64 www.cepmagazine.org June 2002 CEP

 Table 6. Results of consistency tests for data sets #1 and #2.
1
Data Set #1 Data Set #2
Redlich-Kister-Herrington
Area Test
0.8 N area –0.419 –0.393
P area 0.452 0.456
Vapor Mole Fraction, Acetone

DRK 3.73 7.33

M 13.5 18.0
DRK–M –9.72 –10.67
0.6 Consistent? Yes Yes
∆y Point Test
Maximum Deviation 0.032 0.091
0.4 Absolute Average Deviation 0.00487 0.0331
Consistent? Yes No
Data Set #1 Wisniak Area Test
Data Set #2
L 17.50 15.38
0.2 W 18.04 5.81
Dw 1.52 1.37
Consistent? Yes Yes
Wisniak Point Test
0 Maximum Dwi 1.61 7.80
0 0.2 0.4 0.6 0.8 1
Range of Dwi 1.01–1.62 1.05–7.80
Liquid Mole Fraction, Acetone Consistent? Yes No

■ Figure 6. Comparison of x-y diagrams of isobaric data sets #1 and #2. is the largest. The Wisniak point test indicates that the
point nearest to infinite dilution of water in acetone is high-
ly inconsistent. Thus, the calculated infinite-dilution value
upon examination of the data. The temperature does not is probably unreliable. Again, fractionation on the vapor
decrease monotonically with the acetone mole fraction, al- dome would cause acetone to appear to be more volatile at
though no azeotrope is reported. Upon regression, the ba-w the acetone-rich end, lowering the observed infinite-dilu-
and bw-a parameters are found to be 27.429 and –736.786 tion value of water in acetone. This will be shown below to
vs. –267.82 and –710.65 for data set #2. Infinite-dilution be quite important for design work. Overall, data set #2 is
activity coefficients are calculated as 5.86 for acetone and not very good and the causes are apparent from the consis-
4.76 for water as compared to 12.4 and 5.15 from data set tency analysis.
#1. The x-y diagram generated from these Wilson parame- We make a final comparison between isobaric data set #1,
ters is somewhat different than that calculated from data set isothermal data, and UNIFAC-generated parameters. Three
#1, but not exceedingly different (Figure
6). Results of the consistency tests on
data set #2 are given in Table 6. Table 7. Experimental data set #2 — T-x-y data for an acetone-water system at 760 Torr (15).
Data set #2 passes both area tests Point Temperature, °C Liquid Mole Fraction Vapor Mole Fraction
(just barely), but fails the point tests. The Acetone Acetone
inconsistencies are particularly obvious 1 60.0 0.941 0.959
with the ∆y point test. Almost all of the 2 60.4 0.795 0.896
acetone vapor mole fractions are seri- 3 63.3 0.622 0.855
ously underpredicted compared to the 4 62.4 0.445 0.833
experimental data. The deviations are 5 63.8 0.330 0.825
probably due to fractionation on the sur- 6 64.0 0.194 0.795
face of the vapor dome, causing the 7 64.6 0.156 0.791
8 68.3 0.104 0.730
vapor to become enriched in acetone
9 75.1 0.082 0.678
(the original reference indicates that a 10 75.0 0.054 0.615
non-jacketed vessel was used). In partic- 11 84.7 0.025 0.422
ular, fractionation is seen to be most se- 12 100.0 0.011 0.204
vere for water-rich compositions, where 13 100.0 0.003 0.044
the relative volatility of acetone to water

CEP June 2002 www.cepmagazine.org 65

Measurement and Control

Table 8. Comparison of Wilson parameters and McCabe-Thiele results at P = 760 torr. lution activity coefficients at 760 torr were
generated by the Dortmund-modified UNI-
Isobaric Isobaric Isothermal Dortmund
DataSet #1 Data Set #2 Data Set #3 UNIFAC FAC (13). UNIFAC predicts slightly lower
infinite-dilution values for acetone and
Wilson bA-W –267.82 27.43 –146.59 –109.13
Wilson bW-A –710.65 –736.79 –728.74 –769.14
slightly higher values for water than either
γ ∞acetone 12.40 5.86 9.23 8.86 the isothermal or isobaric data. These re-
γ ∞water 5.15 4.76 5.18 5.75 sults are also presented in Table 8.
Minimum Reflux 1.41 0.95 1.28 1.64 The x-y diagrams generated from each
Actual Reflux Ratio 1.77 1.18 1.60 2.05 of the three data sets and the UNIFAC
Theoretical Stages 28 22 27 32 method are very similar in shape in all
Feed Stage 27 20 26 31 cases, except at the water-rich end. Each
Relative Vapor Rate 1.00 0.79 0.95 1.10
shows a pinch at high acetone concentra-
tions. The predicted infinite-dilution activi-
ty coefficients vary somewhat between
sets of isothermal data at temperatures of 35°C, 50°C and these data sets from 4.76 to 5.75 for water and 5.86 to 12.4
100°C were regressed individually by the Barker method (1, for acetone. Do such variations affect distillation design?
2). All were found consistent by the ∆y point test and the The answer is definitely yes.
area test. These data were then re-regressed as a single set Consider the separation of a 50-50 molar mixture of
(data set #3) in the same fashion as the isobaric data to cap- acetone and water, with a desired distillate purity of 98.5
ture the temperature-dependence of the Wilson parameters. mole% acetone and bottoms purity of 99% water. Results
The resulting Wilson parameters and infinite-dilution activi- of a McCabe-Thiele distillation analysis for each system
ty coefficients are given in Table 8. The Wilson parameters are given in Table 8. The number of theoretical stages (in-
were used next to predict the x-y diagram at 760 torr. The re- cluding the reboiler) varies from a low of 22 (data set #2)
sulting x-y plot is compared to the x-y diagram for data set to a high of 32 (UNIFAC). Reflux ratios (at 1.25 times
#1 in Figure 7. The curves are very similar with a slight dif- minimum reflux) also vary considerably, from 1.18 (data
ference at the water-rich end. Infinite-dilution predictions for set #2) to 2.05 (UNIFAC). Relative vapor traffic varies by
water are very close between the isobaric and isothermally- 30% from highest to lowest values. Isobaric data set #1 and
derived data sets (5.15 vs. 5.18), but somewhat different for the composite isothermal data set #3 give very similar in-
acetone (12.40 vs. 9.23). Wilson parameters and infinite-di- termediate results for all values. Most of the variation be-
tween data set #2 and UNIFAC estimates (and the similari-
ty between data sets #2 and #3) is directly attributable to
the differences in the prediction of the activity coefficient
1.0
of water at dilute concentrations in acetone. This value has
a large effect on the “tightness” of the equilibrium pinch
point, and thus on the design of the rectifying section of
Vapor Mole Fraction, Acetone

0.8 the column.

The stripping section design changes very little between
data sets. The acetone is quite volatile in the dilute region
0.6 and only two or three stages are required in the stripping
section (including the reboiler). Due to the lower predicted
volatility of acetone with data set #2 (infinite-dilution value
0.4 of only 5.86 vs. 9–12 for the other data sets), an additional
Isothermal
Isobaric stripping stage is required. Since data sets #1 and #3 were
shown to be internally and thermodynamically consistent,
0.2
most likely the data are reliable.
A column designed on the basis of either data set would
probably function as expected. On the other hand, a column
design based on the inconsistent data set #2 could not reach
0
0 0.2 0.4 0.6 0.8 1 specified purities at design reflux and throughput. With only
20 theoretical stages, a reflux ratio of 2.05 is required to
Liquid Mole Fraction, Acetone meet distillate and bottoms purity specifications. The higher
reflux would increase the vapor load by 40% over the origi-
■ Figure 7. x-y diagrams generated from isothermally derived parameters nal design case. Assuming the column were designed at the
and data set #1 (acetone-water system at a system pressure of 760 torr). typical 70 to 80% of flood velocity, the throughput would
have to be reduced substantially to prevent flooding.

66 www.cepmagazine.org June 2002 CEP

 Prediction of VLE by UNIFAC is not a substitute for ex-
perimentally measured data. For the more common groups
(e.g., ketones), the UNIFAC parameters are fit to a large
number of data sets of many different compounds (with the
acetone-water data undoubtedly included). Because of this
averaging, the accuracy in predicting any particular experi-
mental system is not guaranteed. For the acetone-water sys-
tem, the resulting column design is conservative. However,
References
1. Gmehling, J., and U. Onken, “Vapor-Liquid Equilibrium Data Col-
lection, Aqueous-Organic Systems,” Vol. I, Part 1, DECHEMA Data
Series, Frankfurt/Main, W. Germany (1977).
2. Gmehling, J., et al., “Vapor-Liquid Equilibrium Data Collection,
Aqueous Systems,” Vol. I, Part 1b, DECHEMA Data Series, Frank-
furt/Main, W. Germany (1988).
3. Raal, J. D., and A. L. Mühlbauer, “Phase Equilibria, Measurement
and Computation,” Taylor & Francis, Washington, DC (1998).
4. Hála, E., et al., “Vapor-Liquid Equilibrium,” 2nd English ed., Perga-
mon Press, Oxford, U.K.(1967).
5. Van Ness, H. C., and M. M. Abbott, “Classical Thermodynamics
of Nonelectrolyte Solutions,” McGraw-Hill, New York, NY (1982).
6. Zudkevitch, D., “Temperature Differences Inside Vapor-Liquid Ap-
parati: The Quest for the Isothermal Othmer Still,” Chem. Eng.
Comm., 116, pp. 41–65 (1992).
7. Reid, R. C., et al., “The Properties of Gases and Liquids,” 3rd Ed.,
McGraw-Hill Book Co., New York, NY (1977).
8. Othmer, D. F., et al., “Binary and Ternary Systems of Acetone,
MEK, and Water,” Ind. Eng. Chem., 44 (8), pp. 1872–1881 (1952).
9. Walas, S. M., “Phase Equilibria in Chemical Engineering,” Butter-
worth, Boston, MA (1985).
10. Van Ness, H. C., et al., “Vapor-Liquid Equilibrium: Part I. An Ap-
praisal of Data Reduction Methods,” AIChE J., 19 (2), pp. 238–244
(1973).
11. Wisniak, J., “A New Test for the Thermodynamic Consistency of
Vapor-Liquid Equilibrium,” Ind. Eng. Chem. Res., 32, pp. 1531–1533
(1993).
12. York, R., and R. C. Holmes, Ind. Eng. Chem., 34 (3), pp. 345–356
(1942).
13. Gmehling, J., et al., “A Modified UNIFAC Model 2 Present Parame-
ter Matrix and Results for Different Thermodynamic Properties,” Ind.
Eng. Chem. Res., 32 (1), pp. 178–193 (1993).
Further Reading
Smith, J. M., and H. C. Van Ness, “Introduction to Chemical Engi-
neering Thermodynamics,” 3rd ed., McGraw-Hill, New York, NY
(1975).
Kyle, B. G., “Chemical and Process Thermodynamics,” Prentice-Hall,
Englewood Cliffs, NJ (1984).
Prausnitz, J. M., et al., “Molecular Thermodynamics of Fluid-Phase
Equilibria,” 2nd ed., Prentice-Hall, Englewood Cliffs, NJ (1986).
Prausnitz, J., et al., “Computer Calculations for Multicomponent
Vapor-Liquid and Liquid-Liquid Equilibria,” Prentice-Hall, Engle-
wood Cliffs, NJ (1980).
unnecessary over-design costs money as does under-design.
Use UNIFAC only when experimental data are unavailable
and cannot be measured for some reason.
The foregoing analysis illustrates a very important point
about systems that show positive deviations from Raoult’s
law. Accurate knowledge of the activity coefficient of the
high-boiling component is key to accurate representation of
the VLE for distillation design. The activity of the heavy
component at low concentrations in the light component
dictates whether the system exhibits a pinch or an
azeotrope. The smaller the difference between the boiling
points of the pure components, the more the effect is mag-
nified. Seemingly small errors in the prediction of the activ-
ity coefficient of the high-boiler at dilute concentrations can
greatly change the shape of the equilibrium curve. The esti-
mate of the required reflux and theoretical stages for a given
separation will vary accordingly. For pinched or azeotropic
systems, the error can be manifested in dramatic fashion. It
can mean the difference between a column that performs as
advertised and one that will not work as designed.
The estimate of the activity of the light component at di-
lute concentrations usually is not as critical when the differ-
ence in pure-component boiling points is already large (e.g.,
20°C or more). The volatility of the system is already high
(i.e., the equilibrium steps are large) and errors in the activ-
ity of the low-boiler at dilute concentrations does not lead
to as much error in the distillation design. The opposite is
true for systems that show negative deviations from
Raoult’s law. Negative deviations tend to lower the volatili-
ty of the low-boiler, leading to maximum-boiling
azeotropes. Consequently, prediction of the activity of the
low-boiler at dilute concentrations is most important. CEP
SCOTT BARNICKI is a research associate with Eastman Chemical Co.,
Kingsport, TN (Phone: (423) 224-0341; Fax: (423) 229-4558; E-mail:
sbarnicki@eastman.com). His responsibilities have concentrated on
the initial development and demonstration of process concepts for the
manufacture of specialty and commodity chemicals. He has done ex-
tensive laboratory, pilot-plant and modeling work in the areas of non-
ideal distillation, reaction engineering, process synthesis and phase
equilibrium. Barnicki received a BS in chemical engineering from Ohio
State Univ. and a PhD from the Univ. of Texas. He is the holder of 13
patents and has co-authored the sections “Enhanced Distillation” in
the 7th edition of “Perry’s Chemical Engineer’s Handbook” and “Sepa-
rations Process Synthesis” in the 4th edition of the “Kirk-Othmer Ency-
clopedia of Chemical Technology.” He is a registered professional engi-
neer in the state of Tennessee.
Acknowledgements
The author would like to thank Eastman Chemical Co. for allowing the
publication of this article and for continuing to offer challenging as-
signments with the opportunity to learn.
CEP June 2002 www.cepmagazine.org 67