4 Setting up Multiparameter Equations of State

for Pure Substances

When setting up multiparameter equations of state in the form of Eq. 3.5, two
separate tasks have to be fulfilled - formulations have to be developed both for aD
and for a~ Except for the fact that in both cases least squares fitting techniques are
used, which are explained in Sect. 4.1, these tasks require very different approaches.
Since the development of an equation for a f depends on the equation for aD this
formulation has to be developed or at least finalised first. Theoretically, the contributions to the Helmholtz energy of the ideal gas are well known and empirical
simplifications are usually restricted to simple fits of known functional forms to
accurate data for a single property in order to establish formulations which are
suitable for a broad variety of applications. Section 4.2 summarises the basics of
the underlying theory and common simplifications.
The difficulty lies with the development of a formulation for af. Until now, no
theory has been available which describes the residual contribution to the Helmholtz energy with sufficient accuracy for the whole range of fluid states. Thus, the
development of formulations for a f has to rely almost exclusively on experimental
results; Sect. 4.3 explains how formulations for the residual Helmholtz energy can
be developed based on data for different thermodynamic properties.
In this context, the assessment of the available experimental data is one of the
most crucial points. When setting up equations of state for certain technical applications, the question which has to be answered for each data set is "are these data
accurate enough to fulfil the formulated requirements?". For reference equations of
state, this question has to be replaced by the more difficult question "how accurate
are these data?" since it is the aim to represent all data sets within their experimental uncertainty. Today, estimated uncertainties are usually given in articles on
experimental investigations and are generally becoming more reliable on average.
Nevertheless, one still has to rely on data without or with overly optimistic information on experimental uncertainties. In these cases, an understanding of the applied experimental techniques is often necessary to find satisfactory answers.
Books on experimental techniques (see for instance McCullough and Scott, 1968;
Le Neindre and Vodar, 1975; Eder, 1981/1983; Goodwin et aI., 2000) are useful
as an introduction, but they cannot replace an intense discussion with experienced
experimentalists.
One of the characteristic features of the development of multiparameter equations of state is that little is known about their functional form from theory. While
the basic fitting techniques which are in use today have been known since the late
R. Span, Multiparameter Equations of State
Springer-Verlag Berlin Heidelberg 2000

62

4 Setting up Multiparameter Equations of State for Pure Substances

70s, there has been continual research on algorithms used for optimising the functional form of such equations. The progress in this area was essential to achieve
the accuracies which are typical for modem equations of state. The most important
optimisation algorithms are summarised in Sect. 4.4 together with some general
rules for setting up reasonable functional forms.
Finally, Sects. 4.5 and 4.6 deal with two aspects which are relevant, especially
when setting up reference equations of state: Accurate representation of properties
in the critical region and reasonable extrapolation to very high pressures and temperatures.

4.1 Linear and Nonlinear Fitting

The need for fitting empirical formulations to data arises from the fact that the
"true" physical relation ~(X, Y) =0 with its K independent variables X and the
dependent variable Y (e.g. Xl = T, X2 = P and Y = P for p(T,p)) is either not known
or unsuitable for a certain application. Thus, an empirical formulation
~(X, Y,n)::::: 0 with the parameter vector n has to be developed which approximates the true physical relation. The deviation from ~ =0 is called "residuum'?
Usually the "true" values of X and Y are not known as well and have to be replaced by data sets x and y which are subject to experimental uncertainties.
The task of fitting algorithms is to find the parameter vector n which yields the
best approximation of ~(x, y,n) =o. For certain applications, algorithms have
been developed which minimise the largest value of the residuum (see for instance
Jing and Fam, 1987) but usually the approximation with the highest statistical
probability of n is regarded as the best one. This is the basic idea of the maximumlikelihood method introduced by Fisher (1934) which aims to maximise the likelihood function and can be written as

L(n) =

n~ exp(-~2(xm,ym,n)/(2a~))
M

m=1

2na m

(4.1)

under four conditions:

The errors of the experimental results are normally distributed with an expected value of 0 both for the dependent and the independent variables.
The errors of the variables are independent of each other.
The experimental uncertainties are known and values for the corresponding
variances can be given.
The error of approximation is negligible compared to the uncertainty of the
experimental data.

In Eq. 4.1, ~ corresponds to the total variance which considers the uncertainties
both of the independent and dependent variables; for details on the weighting of
experimental data see Sect. 4.3.4. Equation 4.1 can be traced back to the classical
method of GauS (1809) which minimises the weighted sum of squares of the residua of all M experimental data,

4.1 Linear and Nonlinear Fitting

M

x2 (n)= L(~(Xm,ym,n)lam)

63

(4.2)

m=l

Although it is questionable whether the conditions given above are really fulfilled
when setting up multiparameter equations of state, the resulting procedure for the
determination of the parameter vector n is widely used - in fact the rnaximumlikelihood method has never been used without these assumptions to set up multiparameter equations of state.
To minimise the sum of squares, Eq. 4.2, for a function with / adjustable parameters, ni corresponds to the determination of the minimum of a function in an /dimensional space. The necessary condition for such a minimum is
(4.3)
for all parameters ni.
A linear set-up for ~(x, y, n) is given if the parameters ni are linear coefficients
of terms ai(x) and if the relation between the measured property y or a suitable
combination ao(x,y)l and the ai is linear. In this case ~(x, y,n) can be written as
I

~(x, y,n) = ao(x, y)-

2: nj ai(x) "" 0,

(4.4)

where the form of the / functions ai can be very different, see for instance Eqs.
4.35 -4.37. With Eq. 4.4, the sum of squares becomes

x2(n)= ~Jao(Xm,ym)- ~ni aj(Xm)r a~

(4.5)

The derivative of X2 with respect to the i-th coefficient ni which is required in Eq.
4.3 can be written as

( ~:2)
I

(4.6)

=
x,y.n);fn,

or after recombination as
nj ai(x m). a/x m
j=l

am

)).

am

(4.7)

With Eq. 4.7, the condition for the minimum of the sum of squares, Eq. 4.3, becomes

For ppT data the residuum may be based, e.g., on ao(x,y)

ao(y) =p/(pnRTn).

=Z =P / (p R 1) instead of

64

4 Setting up Multiparameter Equations of State for Pure Substances

(4.8)
Equation 4.8 defines a system of I normal equations which can be written in matrix
notation as

AN=Q

(4.9)

where the vector N holds the I coefficients nj which have to be determined. The Ixl
matrix A consists of typical elements of the form
aij =

m=1

(ai(x m) ~ /X m
am

(4.10)

and the elements of the vector Q can be written as

qi = f(ai(Xm)ao2(Xm,Ym).
m=1
am

(4.11)

Linear systems of normal equations can be solved using the well known Gauft
algorithm in the form proposed by Cholesky (Benoit, 1924). Details on such algorithms can be found in textbooks on matrix operations such as Zurmiihl (1964)
and Zurmiihl and Falk (1992) but today the required subroutines can be obtained
from numerical program packages as well.
Where necessary, constraints can be considered using the method of Lagrangian
mUltipliers as proposed by Hust and McCarty (1967). For each of the constraints,
an equation of the form
I

~(xe' Ye,n) = ao(xe, Ye)-

L ni ai(x e ) = 0

(4.12)

i=1

has to be fulfilled exactly. With these additional conditions the sum of square
which has to be minimised becomes

~Jao(Xm'Ym)- ~ ni ai(xm)J. a~

x2 (n)=

+
Setting the derivative of

~ A{ao(xe, Ye)- ~ ni ai(x e )}

with respect to a coefficient ni equal to zero yields

~[ai(Xm).~ n.1 aj(x m)+ ~-~a.(x

.J
m=1

.J

am

1=1

The derivatives of
tions of the form

(4.13)

.J
e=1

2,e

)=

~ ai(xm).ao(xm,Ym).

.J
m=1

(4.14)

with respect to A~= - Ae/2 yield C additional normal equa-

4.1 Linear and Nonlinear Fitting

65

(4.15)

2: ni a ;(x C )=ao(x c ,yJ.

i=i

In matrix notation, the resulting problem can be written as

(4.16)
where elements of the Cx! matrix C have the form
(4.17)
and the C elements of the vector Qc correspond to the constrained values
(4.18)
Solving Eq. 4.16 yields the set of coefficients ni for the least squares correlation
which fulfils the C constraints exactly.
The linear procedure described above fails either if the relation between ao and
ai becomes non-linear or if the functional form of the ai implies a non-linear relation between different parameters ni. Examples of these two types of non-linearity
are the residua
I
( n
~(x,y,n)=ao(x,Y)-2:nia;(x)+
i

i=1

and

J2 =0

2: n i a;(x, n/+ i ) = 0,
I

~(x, y, n) = ao(x, y) -

aai(x)

-aXI

(4.19)

(4.20)

i=1

respectively. For such residua, the derivatives (a~/ani)x,y,n/'ni depend on the parameter vector 0 and so does the matrix A and the vector Q in Eq. 4.9 and Eq,
4.16, respectively - the system of normal equations becomes non-linear, Such nonlinear systems of normal equations can be solved only by iterative procedures
which require a set of starting values 0, for the parameter vector o. The quality of
the starting values is a crucial point since even sophisticated algorithms may not
find the absolute minimum of the sum of squares without a suitable vector 0,.
Problem dependent strategies have to be applied in order to estimate the required
starting values as accurately as possible, But even with good starting values, convergence cannot be guaranteed for problems with complex intercorrelations between different parameters as they are typical, e.g., for a simultaneous adjustment
of reducing parameters and coefficients, see Sect. 7.2.2, or for a simultaneous fit of
coefficients and exponents of a single term, see Sects, 4.5.3 and 4.5.4.
Different strategies have been developed for the solution of non-linear systems
of normal equations; the corresponding algorithms can be found in advanced textbooks and numerical program packages again. An algorithm which has proved to
be very suitable for the development of multiparameter equations of state is the
Marquardt (1963) algorithm in the modified form proposed by Fletcher (1971).

66

4 Setting up Multiparameter Equations of State for Pure Substances

The application of this algorithm was described in detail by Ahrendts and Baehr
(1979b/1981b).

4.2 Describing the Helmholtz Energy of the Ideal Gas

As already discussed in Sect. 3.1.1.1, the description of the Helmholtz energy of
the ideal gas essentially depends on equations for the heat capacity of the ideal gas.
Based on equations with the general form

c~
* ~ * t* ~ (()k)2
-=no+ LJni T ' + LJmk R

i=l

k=l

e(()k

IT )

(4.21)

(e(()k l T)_1)2

(see also Eq. 3.9), the steps which are necessary to obtain an equation for the
Helmholtz energy of the ideal gas were also discussed in Sect. 3.1.1.1. Thus, this
section focuses on the development of correlations for the heat capacity of the
ideal gas, which can either be based on results from statistical thermodynamics and
molecular constants measured by spectroscopic means or on experimental results
for caloric properties in the low density gas phase. Although both approaches are
very different, they are discussed in a single section here, since a basic understanding of the theoretical background helps to avoid major mistakes when using the
empirical approach, and since the more theoretical approach often results in fitting
Eq. 4.21 to special data as well. Any explanations with regard to statistical thermodynamics will be highly simplified and restricted to results which are useful for
the topic of this section; for a more detailed and theoretically based description see
Lucas (1991).
From classical mechanics, a body which consists of n mass points can be described by 3n pairs of position and momentum co-ordinates. Thus, in an ideal gas,
a molecule consisting of n atoms has 3n degrees of freedom (the momentum coordinates) for movements which can contribute to the dynamic energy state of the
molecule. It is this dynamic energy state which can be observed as internal energy
of an ideal gas on a macroscopic scale and which is hence responsible for the ideal
gas heat capacity.
Assuming a rigid molecule without internal rotation, the 3n degrees of freedom
correspond to translation, external rotation and vibrational modes of the atoms in
the molecule, where neither the rotational nor the vibrational modes must move the
centre of mass of the molecule. When assuming that these effects do not interact
with each other, the heat capacity can be written as

C~.simp = C~.simp + R = c trans + crot + cvib + R

(4.22)

for this simplified model.

In an ideal gas, each molecule has 3 degrees of freedom for translational
movements which are, by definition, fully excited since the molecules must not
interact with each other or with any force fields. Each translational degree of freedom contributes Y2 R to the heat capacity and thus we find the well known relation

4.2 Describing the Helmholtz Energy of the Ideal Gas

c trans

="23 R.

67

(4.23)

The assumption of fully excited translational modes becomes questionable in the

limit of very low temperatures, but it can be shown that it is in general valid for
temperatures of T > 10- 10 K; the only restrictions apply for the quantum gases
helium and hydrogen. Monoatomic molecules of noble gases have only the three
translational degrees of freedom, and thus their heat capacity becomes completely
independent of temperature.
Molecules which consist of 2 or more atoms can rotate around 2 (linear molecules) or 3 (non-linear molecules) principle axis. When assuming that molecules
behave like rigid rotators, i.e., the molecule stretches neither with increasing rotational speed nor with increasing vibrations and that therefore the moment of inertia
does not depend on temperature, each fully excited mode of external rotation contributes V2 R to the heat capacity. Therefore the contributions for Crot are
2

for linear molecules

crot

="2 R

(4.24)
(4.25)

and for non-linear molecules

In most cases, it can be assumed that the rotational modes are fully excited in the
temperature range where multiparameter equations of state are usually applied, i.e.,
for temperatures above the triple point of the corresponding substance. For substances with small molecules or for applications at very low temperatures, this
assumption should be verified. The i-th rotational mode can be regarded as fully
excited if the temperature is much larger than the characteristic temperature of the
rotational mode, (}rot,i (T > "" 2 (}rot,;). The characteristic temperature is given by
().=
rot"

with
and

h2
8:n 2 k I.I

hcA
=--'
k

h = 6,6260755,10-34 Is
k = 1.380658,10-23 1 K- 1
C = 2.99792458,10 10 cms- 1

(4.26)
(Planck constant),
(Boltzmann constant),
(speed of light in vacuum).

For linear molecules, the relevant moments of inertia, 1;, and the relevant rotational
constants, Ai, are the same for both rotational modes due to reasons of symmetry.
The rotation around the bonding axis does not contribute to the dynamic energy
state since the corresponding principal moment of inertia is almost O. For nonlinear molecules, the smallest value for Ii or the largest value for Ai has to be considered. Evaluation of Eq. 4.26 yields (}rot,max= 40.11 K for water, (}rot,max= 7.54 K
for methane, (}rot.max= 2.36 K for difluoromethane (refrigerant 32), and
(}rot,max= 0.49 K for chlorodifluoromethane (refrigerant 22); the influence of the
size and the shape of the molecules is obvious. Listings of molecular constants can
be found in scientific literature (see Herzberg and Herzberg, 1972).
Besides translation and external rotation, 3n - 5 modes of vibration have to be
considered for rigid linear molecules. For rigid non-linear molecules 3n-6 modes

68

4 Setting up Multiparameter Equations of State for Pure Substances

have to be considered. These vibrational modes cannot be regarded as fully excited

at ambient temperatures and are therefore responsible for the temperature dependence of the ideal gas heat capacity. To consider the influence of internal vibrations
it is usually assumed that the molecules behave like harmonic oscillators. Although
this is strictly not correct since the intramolecular forces depend linearly on the
elongation of the atomic bonds, the harmonic oscillator model yields good results
for moderate temperatures.
The Bose-Einstein statistic yields the so called "Planck-Einstein" function

)2

~= ~

(O,lTl

(e(;/Tl_ 1)2

(4.27)

for the contribution of harmonic oscillations to the heat capacity, where 0i is the
characteristic temperature of the i-th vibrational mode. The plot of Eq. 4.27 is
shown in Fig. 4.1. The characteristic temperatures can be determined from spectroscopic results for the corresponding wavenumbers, Vi, according to
0.
I

= hcv i

(4.28)

k'

Extensive listings of fundamental vibrational frequencies for various molecules are

published in scientific literature (see for instance Shimanouchi, 1972; Shimanouchi, 1977; Shimanouchi et aI., 1978; Shimanouchi et aI., 1980). As experimental results, wavenumbers are of course subject to experimental uncertainties. However, typical uncertainties in Vi have only a small influence on calculated heat capacities.
Thus, assuming a rigid molecule with harmonic vibrations and excited rotational modes, the heat capacity of the ideal gas becomes

0.1
Q::;

..........

i;

/'

0.01
0.001

vibrational mode

- - - electronic energy state

0.0001
0.07

0.1

0.2

0.4

. ,.

l.l.

0.7

T l ei
Fig. 4.1. Contribution of a single vibrational mode and a single electronic energy state to the
heat capacity of the ideal gas. For the contribution from electronic energy, a degeneracy of I has
been assumed both for the ground state and for the first excited state (see Lucas, 1991).

4.2 Describing the Helmholtz Energy of the Ideal Gas

69

(4.29)

where mi is equal to 1 except for multiple vibrational modes with identical wave
numbers which may occur in symmetrical molecules. For multiple modes, mi corresponds to the order of the mode and the total number of vibrational modes IVib
has to be reduced correspondingly.
For simple molecules at ambient temperatures, Eq. 4.29 yields an accurate description of the isobaric heat capacity of the ideal gas. Unfortunately, the simple
rigid rotator, harmonic oscillator model fails at high temperatures and for complex
molecules. For these cases, additional contributions have to be considered for
excited states of the electron shell, for internal rotation and finally for corrections
to the rigid rotator, harmonic oscillator model. With these contributions the relation for the isobaric heat capacity of the ideal gas is
o
CP

= Cp,simp + cel + cirot + c eor

0

The contribution from excitation of the electron shell,

temperatures where

(4.30)
Ce],

becomes relevant for

(4.31)
is no longer fulfilled. For practical applications, this equation becomes
eel I

cell

CO

- - ' >::::=;10 for -'-<0.1%;

(4.32)

see also Fig. 4.1. The characteristic temperature of the lowest electronic state can
be calculated from the energy level of the lowest non-ground state, eel,1> or the
corresponding wavenumber, Vel,), according to

e = eel,1 = hCVel,1
el,1

k'

(4.33)

For monoatomic species such as noble gases, typical values for the wavenumber of
the lowest electronic state are in the order of 50,000 cm- l to 160,000 cm-~ resulting in characteristic temperatures of 72,000 K to 230,000 K. Thus, the contribution from excited electronic states is usually negligible for these substances. Extensive tabulations of atomic energy levels are given by Moore (1971) and Kelly
(1987). For diatomic or polyatomic molecules, the levels of single electronic states
are sometimes much lower. The oxygen molecule (0 2 ) for example has a single
(experimentally uncertain) electronic energy level with a characteristic temperature
of about 10,000 K and one energy level of nitric oxide (NO) has a characteristic
temperature which is even as low as 174 K. For diatomic or polyatomic molecules,
the concept that excitation of electronic energy levels has no significant influence
in the temperature range which is relevant for multiparameter equations of state

70

4 Setting up Multiparameter Equations of State for Pure Substances

has to be verified for each case. Tables of molecular energy levels are given by
Herzberg (1950, 1966). Where necessary, the contribution of excited electronic
states can be considered according to the relations given by Lucas (1991); a simple
correlation which takes excited electronic states into account was published by
Wagner et al. (1982).
For polyatomic molecules, groups of atoms may rotate against each other within
the molecule, where the "internal rotation" replaces one vibrational mode of the
molecule. The way in which the internal rotation is considered as a contribution to
the heat capacity depends on the potential energy barrier, Vir,max, which is used to
characterise the type of rotation,
If Vir,max1kT is large (an order of 10 or more), the potential energy barrier is so
high that only a few molecules can rotate internally around the bonding axis. The
result is an oscillation around the minimum of potential energy, which can be
treated as a harmonic oscillation like the other vibrational modes, if the wavenumber of the oscillation is known. This behaviour can be observed for example for
ethene where the double bond between the carbon atoms avoids a real internal
rotation.
If Vir,max1kT is sufficiently small, the internal rotation can be treated as an additional free rotational mode, which results in a contribution of Y2 R to the heat capacity of the ideal gas.
Unfortunately, Uir,max1kT is neither "large" nor "small" in most cases, and rotations around single bonds between groups of atoms have to be treated as hindered
internal rotations. An algorithm for treating such rotational modes is given by
Lucas (1991), but this algorithm is far too complex to be introduced in a simple
correlation for the isobaric heat capacity. Furthermore, the required spectroscopic
data are often either difficult to assess or not available at all. Thus, in many cases
an accurate theoretical calculation of contributions from internal rotation is possible only for scientists who are specialised in this kind of calculations, if it is possible at all.
For corrections to the rigid rotator, harmonic oscillator model the situation is
very similar. At very high temperatures or for molecules which consist of heavy
atoms with weak bonds, the results from the rigid rotator, harmonic oscillator
model do not match experimental results for caloric properties in the ideal gas
state. The necessary corrections to the model have to account mainly for three
effects:
Centrifugal stretching leads to an increase of the moments of inertia of the
molecule and changes therefore the energetic contribution from the rotational
modes.
With increasing amplitudes the anharmonicity of intramolecular vibrations
becomes important.
With increasing amplitudes the average distance between atoms in a molecule
increases due to the vibrational anharmonicities, The increased distances result
in an increase of the moments of inertia and therefore in a rotationalvibrational coupling,
Algorithms which account for these corrections on the basis of spectroscopic data
are well developed (see, Pennington and Kobe, 1954; Schafer, 1960; Gurvich et

4.2 Describing the Helmholtz Energy of the Ideal Gas

71

aI., 1989). But again, these algorithms are far too complex to be included directly
into simple equations for the isobaric heat capacity and the required spectroscopic
data are very often either difficult to assess or not available at all.
Due to the difficulties which are related to internal rotations and to necessary
corrections to the rigid rotator, harmonic oscillator model, it is usually more convenient to fit the parameters of the equation for the isobaric heat capacity of the
ideal gas, Eq. 4.9, to tabulated values if these data were calculated using the necessary corrections to C~,simp, Extensive tabulations which are frequently used as the
basis for correlation equations were published by Gurvich et aI. (1989), by Garvin
et aI. (1987), by Chase et aI. (1985), by the Thermodynamic Research Center at
the Texas A&M University (TRC, 1972-1993), or for individual substances in the
Journal of Physical and Chemical Reference Data (see Chao et aI., 1973; Chen et
aI., 1975).
To fit the parameters of Eq. 4.9 to data for the isobaric heat capacity of the ideal
gas, a simple residuum of the form

~2 = (c~,calc - c~,rneas J2

(4.34)

arneas

can be used, where the subscript "meas" corresponds to the tabulated values. An
accurate assignment of a meas is difficult in this case, since calculated values do not
have an experimental uncertainty in the sense discussed in Sect. 4,3.4. However,
since Eq. 4.9 is fitted only to data of a single property, the value of a meas is arbitrary, as long as it is chosen in a consistent way for all values of c~,meas' Quite often
it is convenient to set a meas equal to the aspired accuracy of the correlation Eq. 4.9.
The required derivatives of the residuum with respect to the adjustable parameters
(see Sect. 4.1) become
(4.35)
t'
R'(- a~ ) = T

an~

a rneas

(4.36)
'

(4.37)

and

R
arneas

(4.38)

72

4 Setting up Multiparameter Equations of State for Pure Substances

While Co, the ni: and the mk alone can be fitted with linear algorithms, the (h can be
adjusted only with non-linear algorithms. When using Planck-Einstein functions,
the characteristic temperatures of the vibrational modes of the molecule are useful
starting solutions for the (h. In this case the mk should be set equal to 1. For
polyatomic molecules with a large number of vibrational modes, groups of modes
with similar wavenumbers can be combined in a single term with an average value
for (h and a value for mk which corresponds to the number of combined modes.
The temperature exponents of the polynomials can either be fitted nonlinearly or
optimised by a suitable algorithm (see Sect. 4.4), but usually they are simply chosen by trial and error.
Where only a limited range of validity is required, pure polynomial forms of
Eq. 4.9 without any Planck-Einstein functions are widely used. The constant Co and
the of the equations are simply fitted to the data and the extrapolation behaviour
of the correlations is not examined at lower or higher temperatures. Such equations
are typical for Helmholtz equations for the halogenated hydrocarbon refrigerants.
When fitting Eq. 4.9 to data with a broad temperature range, the constant Co is
usually set equal to the contributions from translation, external rotation, and where
applicable, from free internal rotation, C trans , Crot. and Cirot. In this way, Eq. 4.9
yields reasonable values for the isobaric heat capacity in the limit of low temperatures, at least if the temperature exponents of the polynomials are chosen to be
positive.
For a better extrapolation behaviour at high temperatures, it is useful to avoid
polynomial terms completely. In this case, the contribution from vibration and all
corrections to C~.simp are expressed only by a summation of Planck-Einstein functions. To achieve reasonable extrapolation at high temperatures, all the coefficients
mk should be positive and their sum should roughly be equal to the number of
vibrational modes of the considered molecule. Often the lowest values of (h which
result from the fit can be assigned to the corresponding values from spectroscopic
data, see Eq. 4.28, with only small deviations, while different vibrational modes
with high characteristic temperatures can be combined in a single empirical term
with a correspondingly large value of mk - this experience justifies the starting
solution proposed above.
Wherever possible, theoretical results for the heat capacity of the ideal gas
should be compared to experimental results for caloric properties in the low density gas phase in order to exclude possible errors in the interpretation of underlying
spectroscopic data. Traditionally, experimental values for isobaric heat capacities
in the ideal gas state were determined by extrapolation of data for the gas phase
heat capacity according to

t;

n;

t;

(4.39)
Unfortunately, experimental results for the isobaric heat capacity in the low density gas phase are rather uncertain due to the small masses (mass flows) in the
calorimeters. Under these conditions, small absolute errors in the determined heat
flow result in considerable uncertainties in the measured heat capacities. While
ideal gas heat capacities which were calculated from statistical thermodynamics

4.2 Describing the Helmholtz Energy of the Ideal Gas

73

and spectroscopic data have typical uncertainties of L1c;/c; < 0.1 % for moderate
temperatures and simple molecules, good experimental results have typical uncertainties of L1c;/c; < 0.3 %- 1.0 %. Thus, experimental results according to Eq.
4.39 are very often unsuitable to test Eq. 4.9 on a high level of accuracy.
The situation has improved with the introduction of sound speed measurements
from spherical resonators. These devices allow extrapolations of the ideal gas
speed of sound,
WO

= p-+o
lim weT p)
"

(4.40)

with typical uncertainties of L1wo/wo < 0.01 %. For ideal gases, the isobaric heat
capacity can be calculated from the speed of sound according to
(4.41)
and an error analysis yields
(4.42)
Thus data from speed of sound measurements with spherical resonators are very
suitable to test results from theoretical calculations, especially for small molecules
or low temperatures where c; / R is small.
For more complex molecules like halogenated hydrocarbon refrigerants, the
available spectroscopic information is very often insufficient to calculate accurate
heat capacities from statistical thermodynamics. Ab initio calculations have been
proposed to solve this problem (see for instance Lucas et aI., 1993). However,
these results have not yet been considered when setting up multiparameter equations of state. For these fluids, it is common to fit Eq. 4.9 directly to experimental
results in exactly the same way as discussed above. Where data from spherical
resonators are available, this method is not necessarily disadvantageous with regard to accuracy, but with regard to the covered temperature range. Based on apparatuses used today, temperatures from the triple point of typical cryogenic substances up to about 450 K can be covered by spherical resonator measurements
(see for instance Estrada-Alexanders and Trusler, 1995). Usually however, reliable
data are available only in smaller temperature ranges and a direct fit to these data
restricts the range of validity of the equation for the Helmholtz energy of the ideal
gas. This method may be acceptable for equations which are intended to become
technical standards without application at higher temperatures, but it is unsatisfactory for setting up scientific property standards. For chlorodifluoromethane
(refrigerant 22) Wagner et aI. (1993; see also Marx et al., 1992) fitted only the
necessary (but not sufficiently investigated) anharmonicity corrections to experimental data while maintaining the results from the rigid rotator, harmonic oscillator model as a theoretical background. This mixed approach yielded a better extrapolation to high temperatures than simple fits to experimental data.

74

4 Setting up Multiparameter Equations of State for Pure Substances

4.3 Describing the Residual Helmholtz Energy

As discussed before, recent multiparameter equations of state are usually set-up in
the form of the reduced Helmholtz energy (see for instance Eq. 3.5), where the
residual part of the Helmholtz energy, a~ is determined from an empirical representation of experimental data. However, Helmholtz energies cannot be measured
directly and adopting the procedures described in Sect. 4.1 to such an equation
requires a multiproperty approach. Fits which involve data for derived properties
have already been used to set-up pressure explicit equations (see for instance Hust
and McCarty, 1967; Wagner, 197011972; Bender, 197011973, Jacobsen and
Stewart, 1973) for a long time, but with the introduction of Helmholtz equations
they became mandatory. For equations in form of the Helmholtz energy, Ahrendts
and Baehr (1979al 1981a) summarised the theoretical background of the required
procedures, but their application was still restricted to only a few properties (see
Ahrendts and Baehr, 1979bl 1981b). During the following years the use of multiple properties was extended more and more, reaching its current maximum for the
new reference equation for water developed by PruB and Wagner (1995; see also
Wagner and ProB, 1997/2000) which was fitted to data of 16 different properties.
The basic idea of the multiproperty approach and its application to Helmholtz
equations is described in Sect. 4.3.1. In Sects. 4.3.2-4.3.3 the required residua are
discussed and Sect. 4.3.4 describes the determination of the variances of experimental data which are needed to apply the relations given in Sect. 4.1.
4.3.1 The Multiproperty Approach

Following the nomenclature used in Sect. 4.1, the usual formulation for the residual part of the Helmholtz energy, Eq. 3.25, can be written as
I

ar(x)= Lnj

Aj(x)

(4.43)

with Xl = -r = Trl T and X2 = 0 = P I Pro In Eq. 4.43 the Aj(x) correspond to -rtj Odj for
polynomial terms and to -rtiodiexp(-rjOPi) for exponential terms; more complex
functional forms are used in reference equations of state for an improved representation of properties in the critical region, see Sects. 4.5.3 and 4.5.4. However, at no
point in Sect. 4.1 was it assumed that the aj(x) in Eq. 4.4 have the same functional
form. Thus, the Aj(x) in Eq. 4.43 can be identified with the aj(x) in Eq. 4.4.
The fact that the ao(x,y) and aj(x) may correspond to very different functional
forms makes it possible to fit equations of state to data of mUltiple derived properties. Usually, the most important information on the thermodynamic surface of
pure substances results from data for the ppT relation. When introducing Eq. 4.43
into the relation between compression factor and reduced Helmholtz energy,
Eq. 3.30, the relation for p(T,p) reads

4.3 Describing the Residual Helmholtz Energy

~=l+o(aar) =l+on (aA;(T,O))

pRT
ao
i=1
ao

75

(4.44)

and the corresponding residuum

can be written in an explicit linear form as

~ ppT

(4.45)
Thus, ppT data can be used in the linear algorithm described in Sect. 4.1 with
a

and

p-pRT
p2RT

ppT,O -

(4.46)

1(aA)
ao'

_ _ _I
ppT,i -

Pr

(4.47)

In the same way, the residuum

~c,

v
= ( -+T
a" ) - 2:I ni [ -T

i=1

a Ai
(2))
aT 0

(4.48)

-2-

can be derived from the relation between isochoric heat capacity and Helmholtz
energy, With
(4.49)

and

ac

",

i - -T

(a Ai)
2

(4.50)

-2 '
aT
0

data for the isochoric heat capacity can be used in linear fits too and since no restrictions have been made with regard to the functional forms of ai and ao in
Sect. 4.1 they can be used together with ppT data in a single fit. To illustrate this
fact, the sum of squares according to Eq. 4.5 can be rewritten as
2

a p,rn

(4.51)

for a multiproperty fit which considers P different properties. Equations 4.9-4.11

change correspondingly and the consideration of constraints is not affected by the
reformulation of Eq. 4.5.
From Eq. 4.51 the importance of the variance ai"m becomes obvious. In a multiproperty fit, residua of different properties are only comparable if they are reduced
by a suitable quantity. This quantity is usually the experimental uncertainty of the
data, see Sect. 4.3.4; for the development of equations of state for technical appli-

76

4 Setting up Multiparameter Equations of State for Pure Substances

cations, the experimental uncertainty may be replaced by the accuracy which is

aspired for the corresponding property, see Sect. 6.2.
Problems arise, e.g., for the consideration of enthalpy differences
llh = hiT2,P2) - h1(T"PI)' Starting from Eq. 3.52 and introducing the abbreviations from Eq. 3.57, the expression for enthalpy differences becomes
(4.52)
and the corresponding residuum can be written as

h(~~~I) [ra~t +[ra~l)

- ~ n{[ rea~i )0 + o( aad )J2- [r( aa~i )0 + o( aad

This residuum yields an expression for
fits, namely

a!!.h,i

)Jj.

(4.53)

which seems to be suitable for linear

(4.54)

but since enthalpy differences are usually measured as a function of temperature

and pressure, Eq. 4.53 is an implicit linear relation. For a consistent evaluation of
Eq. 4.53, the densities which correspond to Tlo PI and T2, P2 have to be calculated
from the equation of state using the current parameter vector n. Since new values
for PI and P2 exercise an influence on n and vice versa, this is only possible for
iterative procedures as they are typical for nonlinear algorithms like the one proposed by Fletcher (1971).
Other properties lead to implicit nonlinear residua. For the speed of sound one
finds, e.g.,

~ 2=(~-1)
RT
W

-ni [20(aAi)
i=1

[I

ao

02(aao2~i) 1
T

ni[o(aAaoi) - or(~)
aoar ll2
i=1

(4.55)

From the fraction in Eq. 4.55, the coefficients cannot be isolated and thus any
derivative of , with respect to a coefficient will contain the complete vector of
coefficients, n. Nonlinear algorithms are required to minimise sums of squares
which contain contributions from such properties unless linearisation procedures
are applied which are described in the following section.

ni

4.3 Describing the Residual Helmholtz Energy

77

4.3.2 Defining Residua for Linear Algorithms

Linear algorithms can be used directly to determine the coefficient vector n if only
data are used in the fit which lead to explicit linear residua. Usually these are data
for the pp T relation, for differences of the internal energy u, for the isochoric heat
capacity c v, for the second and third virial coefficient Band C, and for the second
acoustic virial coefficient f3a. The residua for these properties are given in Table
4.1 except for those of the ppT relation and the isochoric heat capacity which were
given in Eqs. 4.45 and 4.48, respectively. In Table 4.1, the subdivision into ao and
ai which was shown in detail in Eqs. 4.45-4.47, 4.48-4.50, and 4.53-4.54 is
indicated by brackets.
Data for properties which result in implicit linear residua such as enthalpy differences cannot be considered exactly when using linear algorithms, but the implicit dependency can be transformed into an explicit one by precorrelation of the
density which corresponds to the given values of temperature and pressure. Since
the uncertainty of calculated densities is usually much smaller than the uncertainty
of enthalpy differences, the corresponding densities can be calculated from preliminary equations of state without significantly affecting the representation of the
enthalpy differences. Nevertheless, this method does not yield exactly consistent
results and the resulting equation has to be compared both to the original h(T,p)
data with p calculated from the equation and to the precorrelated h(T,p) data to
prove that the influence of the precorrelation is negligible. The residuum for enthalpy differences was given in Eq. 4.55. Enthalpies of evaporation can be considered as h(T,p") - h(T,p') where the densities of the equilibrium phases can be
precorrelated either from a preliminary equation of state or from the corresponding
ancillary equations. Again, for the resulting equation, it should be proved that the
influence of this precorrelation is negligible.
Implicit linear relations are also found for the thermal properties of the vapourliquid phase equilibrium,
pi and pi: As written in Eqs. 3.71 and 3.72, the phase
equilibrium condition yields implicit residua since all three properties are contained in both conditions. Data for a single property can only be used in combination with iterative procedures where the other properties are calculated from the
equation of state depending on the current vector of coefficients, n. However,
Wagner (1970/1972) and Bender (1970/1973) found that the phase eqUilibrium
condition can be resolved into three explicit linear conditions, if values for all
three properties are available. The residua of these conditions read

Ps,

~VLE,1 = (:~ (~-~)-ln(P:))ini (A;(i,c5')- A;(i,c5"),

p
p
p

(4.56)

.=1

~
and

VLE,2

VLE,3

= (Ps - P'RT)_ ~ n. _1 (aAi (i,c5')

p

,2

RT

L.J. p
.=1

ac5

'f

= (Ps - PIlRT)_ ~ n. _1 (aAi (i,c5 )

p

112

RT

l1

L.J.
i=1

Pr

ac5

'

(4.57)

(4.58)
'f

78

4 Setting up Multiparameter Equations of State for Pure Substances

Table 4.1 Residua for properties which result in explicit linear relations. a
Property

Residuum

ao

B(1)

C(1)
/3a(1)

Except for ppT data and isochoric heat capacites for which the residua were given in Eqs. 4.45
and 4.48. respectively.

where Eq. 4.56 corresponds to the Maxwell criterion itself while Eqs. 4.57 and
4.58 represent the condition that the vapour pressure has to result from the ppT
relation both for T,p' and T,p'~ Since values for all three properties are usually not
available for exactly the same temperature from experiments. the data set for the
thermal properties at phase eqUilibrium can be calculated from the corresponding
ancillary equations - thus accurate ancillary equations for Ps. p' and p" are extremely important. This step becomes a problem for substances where the data
situation for one of the properties, usually for the saturated vapour density, is
scarce. In this case saturated vapour densities can be calculated from a simple
equation of state for the gas phase and from the ancillary equation for the vapour
pressure as p" = p(T,ps)' The result of the precorrelation has to be checked by
comparison of the resulting equation of state with original data.
The necessary precorrelations become more complicated for implicit nonlinear
data. For speeds of sound, Jacobsen et al. (1986a) introduced the residuum
(4.59)

with
This linearised residuum results directly from the relation

(4.60)

4.3 Describing the Residual Helmholtz Energy

w2(T,p)=

Cp

Cv

.(aapp )

79

(4.61)
T

which was used to calculate speeds of sound from pressure explicit equations of
state. To use speed of sound data in a linear fit, the precorrelation factor y,.,z has to
be calculated from a preliminary equation of state and precorrelated densities have
to be used to make Eq. 4.59 explicit. The calculation of y,.,z makes high demands
on the accuracy of the preliminary equation, especially if speed of sound data have
to be precorrelated in regions where no accurate data are available for the isochoric or isobaric heat capacity. These high demands have lead to problems with
regard to the representation of highly accurate speed of sound data from measurements with spherical resonators, see Sect. 4.4.6. Tegeler et al. (1997/1999) proposed a different residuum, where speed of sound data are linearised as

with

(4.63)

and

This more complex precorrelation enhances the benefit of linearised speed of

sound data since Eq. 4.62 contains information on the mixed second derivative of
the Helmholtz energy which was lost in Eq. 4.59. However, Eq. 4.62 proved to be
superior only in regions where highly accurate speed of sound data were available
and where accurate preliminary equations could be formulated.
For the isobaric heat capacities, Saul and Wagner (1989) introduced the linearised residuum
(4.64)

with

fc p

{1+oa:;-om:;tf
= (1+2oad +o2ad<l)"

(4.65)

The precorrelation factor fcp and the density which corresponds to the given values
of temperature and pressure have to be precorrelated from a preliminary equation
of state again in order to transform the implicit nonlinear relation for the isobaric
heat capacity into an explicit linear relation.

80

4 Setting up Multiparameter Equations of State for Pure Substances

Experimental results for the speed of sound and the isobaric heat capacity are
published not only for homogeneous states but also for the saturated liquid and
vapour. These data can be considered according to the relations given above with
the only exception that the density has to be precorrelated from the corresponding
ancillary equation or from an equation of state (see Sect. 3.3.5) as a function of
temperature only. When comparing an equation of state to the original data, the
corresponding densities have to be calculated from this equation for consistent
results.
Data for the heat capacity of the saturated liquid cannot be used directly in linear algorithms. These data can be converted into data for the isobaric heat capacity
of the saturated liquid according to the relation

(aH'!ft)
p)

(T)= a (T)--l'J*

cp

p'2

ap

(4.66)

c;

which is implied
Up to reduced temperatures of about T ITc = 0.95 the error in
by this conversion should be clearly less than O.5% and therefore smaller than the
uncertainty of typical data for the heat capacity of the saturated liquid if Eq. 4.66 is
calculated from a good preliminary equation (see also Span, 1993).
Until now, no direct linearisation has been proposed for data for the JouleThomson coefficient. In most cases, such data are only of minor importance. When
the data are essential for a certain substance they can be used as data for isenthalpic pressure changes (Ah(T2, P2, TJ, PI) =0). The corresponding values of TJ, PI
and T2, P2 can be calculated from

Pi.2

= Pf.l !J..p/2

and

(4.67)

where Tf.l and PI' correspond to temperature and pressure of the published JouleThomson data point. Published experimental data for the differential JouleThomson coefficient were usually measured as integral Joule-Thomson coefficients over a finite pressure difference !!.p. When this pressure difference is given
it should be used in Eq. 4.67; otherwise a reasonable value must be estimated.
4.3.3 Defining Residua for Nonlinear Algorithms
With respect to the required residua and derivatives, there is, at least on principle,
little difference between linear and nonlinear fitting routines. The sum of squares
is still defined by Eq. 4.2, and Eq. 4.3 still defines the necessary condition for the
minimum of the sum of squares. The first difference is that the fitted parameters do
not need to be linear coefficients when using nonlinear fitting algorithms. Formally
the required derivative of the sum of squares has to be given in a more general way
as

4.3 Describing the Residual Helmholtz Energy

81

(4.68)
The simplification of this relation, Eq. 4.6 extended to a multiproperty fit, is still
valid in most cases. Within certain numerical limits, Eq. 4.68 leads to different
results only if nonlinear parameters such as exponents or parameters within exponential expressions are fitted.
For explicit linear properties, neither the residua nor their derivatives change
when using them in combination with nonlinear algorithms; Eqs. 4.45, 4.48 and
the residua which are summarised in Table 4.1 can be used without changes. The
given (Eqs. 4.47 and 4.50) or indicated derivatives of the residua change only if
fitted parameters require more general formulations, see above.
For implicit data precorrelations are no longer necessary since p(T,p,n) can be
calculated directly from the equation of state. Since the calculated density depends
on the current parameter vector n in this way, an additional contribution has to be
considered when calculating the derivative of the residuum with respect to a parameter nj. For an implicit property z(T,p); the required derivative of the residuum
becomes
(4.69)

(4.70)

with
-)

The term (ap/ap)T,n can be replaced by (ap/ap)T,n' see Table 3.9. The required
derivative of z with respect to density can be calculated based on the relations
given in Table 3.9 with

(apaz) T,n = r;;1 (az)

ab

T,n'

(4.71)

For common implicit properties, these derivatives are summarised in Table 4.2.
The resulting relations involve third derivatives of a r which are not given in the
corresponding tables here. While these derivatives are still simple for pure polynomial and exponential terms, they become rather complex for hard sphere terms,
Gaussian bell shaped terms and nonanalytical terms, see Tables 3.6, 3.7, and 3.8,
respectively. Although analytic derivatives are advantageous in principle, the use
of numerical derivatives is justified in this case to avoid these complex expressions. When using data for enthalpy differences, the contribution fromp(T,p,n) has
2

Since reduced properties are used in the corresponding residua, z(T,p) has to correspond to a
reduced property in Eqs. 4.69 and 4.71, and in Table 4.2, too.

82

4 Setting up Multiparameter Equations of State for Pure Substances

Table 4.2. Derivatives of implicit properties with respect to density

Property, reduced

Derivative (azjap)T,n

Enthalpy

I
(a(h/RT)J = -'(Ta~,
+a~ +oa~o)
ap
Pr

T,n

Isobaric heat capacity

a( CaP'P/ R)

T,n

- ' -T

a orr +

2(a~ +oa~o -Ta~, -aT abo, )

r

1+20a o +o aM
(2 ab +40abo +0 2 abOo)
'(I+oa~-ora~,)2
(1+20a~ +0 2 a~o)
Pr

'(l+oa~ -arab, )2)

Speed of sound

a(w 2 /RT)]
I
r
2r
r
r
= -'(2a
+0 aoM -(I+oa o -oTa o,)
[
or +40a oo
ap

Pr

T,n

2(ab +oa~o -ra~, -oTabo,)

T2(a~ +a~)

Joule-Thomson coeff.

( a(p-a'PRp)

J -T,n

a~rr(l+oad -oTad, )2]

T2(a~ +a~)

1 (aro+3ua
.. oo
r +u.. 2 aMo+Ta<lT
r
r
-1 ' ( - -

1/J

Pr

r ,)
+ma oo

+2~(ab+oabo-Ta~,-OTabOr )'(l+oab-OTab,)
1/J

:2 a~(1+20ab+02a~o)
_T 2 :2 (a~+a~)(2a~+40adO+02a~M)J
_T 2

with

1/J = (I +oa:i-oTa:i,) -T2(a~+a~ )(1+ 20a:i+ 02a:io )

<p = oa~+02a:io+OTad'

to be considered both for PI andpz, For data on the phase boundary, Eq, 4,70 has
to be replaced by

4.3 Describing the Residual Helmholtz Energy

83

and the other derivatives involved in Eq. 4.69 have to be evaluated either for p'(T)
or for p"(T).
Nonlinear data can be used directly in combination with nonlinear algorithms,
no linearisation is required. For the speed of sound the corresponding residuum
was already given in Eq. 4.55. The derivative of Eq. 4.55 with respect to nj reads

(4.73)
where the list of fixed parameters is omitted for the derivatives of a with respect to
nj. The residuum of the isobaric heat capacity,

(4.74)
yields the derivative

1'2 (

aa r

------I!...

an;

[20(~::;)+02( a::o))'(1 +oa:; -orab,)2

I

(1 +20ab +0 2abO)2

2[0(~ )-O1'(~ )}(l+oab-Orab,)

l+20ab +0 2abo

(4.75)

Data for the Joule-Thomson coefficient can be considered using the residuum
(4.76)

and its derivative

84

4 Setting up Multiparameter Equations of State for Pure Substances

((1 + oa~-ora~r)2 _r2( a~+a~r )(1 + 2oa~+o2a~o)

Ho(:: }or(~n7 )}(l +oa;-Jra&l-r'(":n7 )

(4.77)

The vapour-liquid phase equilibrium condition can be considered in exactly the

same way as described for linear algorithms, see Eqs. 4.56-4.58. However,
Ahrendts and Baehr (1979b/ 1981b) proposed a direct nonlinear approach for the
thermal properties at phase equilibrium, which does not require a complete set of
equilibrium properties at the same temperature. For this approach the required
residua and their derivatives become

= Ps -

p,

Ps,ealc(T,n)

P r RTr

'

(4.78)

(4.79)

_ P' - p~alc(T,n)

sp' -

Pr

(4.80)

(4.81)

r " = P" -

~p

" ( T,n)
Peale

Pr

(4.82)

4.3 Describing the Residual Helmholtz Energy

85

(4.83)

and

This procedure requires considerably more computation time than the linearised
solution, since the phase equilibrium has to be calculated from the equation of
state for every single residuum. However, it is particularly advantageous for substances where the data situation for one of the properties is scarce. This is because
it does not constrain the equation of state to a questionable ancillary equation (see
Marx et aI., 1992; Wagner et aI., 1993). An advantage for substances with very
accurate data sets is the improved consideration of phase equilibria very close to
the critical point, where the decoupled residua of the nonlinear approach allow a
more precise weighting of the data. Thus, nonlinear direct fits to phase equilibrium
data are state-of-the-art for group 1 reference equations of state (see Chap. 5) and
for equations of state which are designed to describe substances with a poor data
situation (see Sect. 6.2).
4.3.4 Assigning Weights to Experimental Data

The definition of the likelihood function, Eq. 4.1, by Fischer introduces the variance ~ of the experimental data. This variance can be found as well in the definition of the sum of squares of the multiproperty approach, Eq. 4.51, where a7"m
takes into account the difference between the variances calculated for different
properties p, The quotient 1 / a7"m is referred to as the "weight" of a data point.
A problem arises from the fact that usually only a single result is reported for
each measured point - thus, its variance cannot be determined according to the
definitions given by the statistical theory. And even if repeated measurements are
since the systematic
reported, the resulting variance should not be used for
errors are often larger than the normally distributed random errors in experimental
set-ups. When working on reference equations of state, it is useful to identify a7"m
with an estimated experimental uncertainty of the corresponding point or, more
precisely, with the experimentally caused uncertainty of the corresponding residuum. In this case, the weighted contribution of a data point to the sum of squares
becomes smaller than 1 if the point is represented within its experimental uncertainty and the assessment of results becomes comparably easy,
The problems which are related to the estimation of an experimental uncertainty
were already mentioned in the introduction to this chapter. Estimates published by
the experimentalists themselves are very helpful but unfortunately are often overly
optimistic, especially for older data sets. When setting up a reference equation of
state, the framework of weighting has to be based on a detailed analysis of the
available data sets. This analysis has to consider both the experimental techniques
applied and comparisons between different data sets, where possible even between
data sets for different properties. Since Helmholtz equations yield consistent results for all properties by definition, preliminary equations can be powerful tools to
detect inconsistencies. However, this kind of data analysis requires a profound
knowledge regarding restrictions of the functional forms used (especially when

a;,m

86

4 Setting up Multiparameter Equations of State for Pure Substances

assessing data in the critical region) and a lot of experience both with the data set
and with setting up equations of state. This is the reason why highly accurate reference equations are usually the result of long-term projects which comprise dozens
or even hundreds of preliminary equations.
When calculating weights, it is important to realise that a;,.m does not correspond just to the uncertainty of the dependent variable y; the uncertainties of the
independent variables x must be considered as well. A rigorous way to do so is to
rewrite Eq. 4.51 in a more general way as
(4.84)

where X and Y are the "true" values of the variables again, see Sect. 4.1. Since X
and Yare not known, Eq. 4.84 cannot be evaluated directly. Methods to estimate
~p(X,y,n) and the (Xk-Xk) from the observed ~p (x,y,n) were discussed in the literature and software is available, e.g., for weighted orthogonal distance regressions,
but an easier solution is used in general practice today. With the assumption that
errors in the different variables are independent of each other, that (Y - y) and all
(Xk-Xk) are small enough to neglect the curvature in ~p, and that ~p is a continuously differentiatab1e function, Eq. 4.51 becomes equivalent to Eq. 4.84 with
(4.85)
This relation is known as "law of error propagation" and results from the combined uncertainty of y and x when breaking a Taylor expansion after the linear
terms (see Ahrendts and Baehr, 1979bI1981b). The assumptions which lead to
Eq. 4.85 are usually fulfilled for experimental data of thermodynamic properties;
there are only two kinds of problems which make different approaches necessary.
Especially close to the critical point, the curvature of ~p becomes too large to be
neglected for some properties. Figure 4.2 shows a plot of the isochoric heat capacity of carbon dioxide in the immediate vicinity of the critical point as calculated
from the Helmholtz equation of Span and Wagner (1996). When calculating acv,m
for the plotted point with the given value of aT, which is overstated here for better
illustration, the resulting contribution from the uncertainty in temperature, acV'T, is
overestimated for temperatures above the measured one (T> Tm) and underestimated for temperatures below the measured one (T < Tm). This effect could be
avoided if the curvature of C v were considered. But when taking into account the
curvature, different variances would be obtained for T> Tm and T < Tm and the
linear procedures discussed here are not able to consider different variances for a
single data point. Thus, the value of aT has to be reduced to use the data point in a
reasonable way. This is possible in certain situations where the distance from the
critical temperature is known more accurately than the absolute temperature. In
this case, the given temperatures have to be adjusted by the difference between the
critical temperature which was selected for the equation of state and the value

4.3 Describing the Residual Helmholtz Energy

87

.-.., 5.0 r------.---------..,

~
Ucv.T

for T< Tm

304.10

304.12 Te 304.14

304.16

304.18

Temperature TIK
Fig. 4.2 Plot of the isochoric heat capacity of carbon dioxide in the immediate vicinity of the
critical point. The strong curvature leads to incorrect results when using the law of error propagation, Eq. 4.85, to detennine UcvT.

which results from the respective data set, !J.T = Teeq.- Te.m, where !J.T has to be
smaller than the original value of aT. The adjusted data can be used with a smaller
value of aT,adj. since only the uncertainty of the difference from the critical temperature has to be considered here.
Close to the phase boundary implicit data may cause problems, since the phase
boundary results in discontinuous plots of '(,p(T,p). Using the speed of sound in
carbon dioxide as an example, Fig. 4.3 shows a data point in the liquid phase. The
linearisation implied by Eq. 4.85 is justified, the curvature of the isotherms can be
neglected when calculating the contribution of the pressure uncertainty to aw,m'
However, the pressure difference Pm- Ps(Tm) is smaller than the uncertainty of the
vapour pressure calculated from the equation of state, which is overstated in the
figure again. Thus, the data point may be regarded as a vapour point with a much
smaller calculated speed of sound in a nonlinear fit, see solution (B). In this case,
the contribution to the sum of squares becomes very large and the whole fit may
become unstable, oscillating between two solutions with the corresponding data
point in the liquid and in the vapour phase, respectively. To avoid such problems,
it is useful to start the iteration of the density belonging to Tm and Pm with the precorrelated value for p(Tm , Pm) from the linearised data set when calculating the
nonlinear residuum '(,,.,2. If the precorrelated value corresponds to a liquid density,
the result of the iteration will also correspond to a liquid state even though this
state may be metastable, see solution (A). In this way problems with implicit data
close to the phase boundary can be avoided, except for data close to the critical
point. For temperatures close to Te, the Maxwell loop becomes very flat (see Sect.
3.3.5) and a metastable solution on the liquid side may become impossible; the

88

4 Setting up Multiparameter Equations of State for Pure Substances

500

,.'""'

450

--

400

e'"

'-'

=
0

"0

0p.eq.

"I

w'7

op

{J-I- -- .I
(A)

350

;::I

'"

300

a,)
a,)

250

to-.
0
"0

Q..
c:Il

200
1505.20

.l

~j

. +0
P.eq
.

. -

metast. .

~liquid' T = 290 K
\

wvapour, T = 2:0

L w"

(B)

5.25

K:
:
:

5.30 Ps

5.35

5.40

Pressure p I MPa
Fig. 4.3 Implicit data close to the phase boundary may cause problems in nonlinear fits. The
discontinuity of ~(T,p) violates the conditions for use ofEq. 4.85.

iteration results into a vapour density independent of the starting value. In this
case, the corresponding data point has to be removed from the data set in order to
guarantee stable nonlinear fits.
In Eq. 4.85 the derivative (a~play)x is easy to evaluate, but a preliminary equation of state is required to calculate the derivatives of l;p with respect to the Xk. For
an initial weighting, the contributions from the uncertainty of the independent
variables could be neglected if no suitable equation is available from the literature,
but unfortunately this approach fails for ppT data. The main contribution to appT
does not result from the pressure, which corresponds to y in Eq. 4.45, but from the
independent variable p. An initial weighting which is based only on a p would lead
to an overfitting of liquid data and may not result in an equation which can be used
for an improved weighting in the next step. Thus, it is better to use a simple cubic
equation or an equation of state adopted from another substance by a simple corresponding states approach for an initial weighting. Data in regions where problems
are expected and data for implicit properties should not be used before a reasonable preliminary equation has been established. After this is accomplished, the
weighting procedure becomes an iterative process which soon converges - small
changes in the preliminary equations result in small changes of the weights, but
they have little influence on the whole fit.
To gain additional flexibility, it is advantageous to introduce an additional
weighting factor which allows some subjective influence on the weights of the data
used. To do so, the expression (1/ap,m)2 has to be replaced by ifwt/ap,m)2 in Eq.
4.51. The weighting factor fwt has to be equal to 1 in general, but it can be increased for accurate data in regions where experimental results are scarce and it

4.3 Describing the Residual Helmholtz Energy

89

can be reduced in regions with high data density. In this way, overfitting is avoided
on the one hand and on the other hand regions which are experimentally investigated less good are considered sufficiently by the fitting algorithms.
Even though the different accuracy of data is considered by the weighting procedure described above, it is not useful to include all data in the used data set when
setting up reference equations for substances with extensive data sets. In regions
where accurate data from different apparatuses are consistent with each other,
there is no need to use less accurate data to set up the equation. To exclude less
accurate data avoids negative influences which may result from systematic errors
of those data; the procedures described above consider different accuracies formally correct for data with random errors, but negative influences from systematic
errors of inaccurate data cannot be excluded. Furthermore, computing times are
reduced and comparisons become easier and unequivocal. However, it is important
to keep in mind that this step does not mean a disparagement of the excluded data.
The whole ranking is influenced more by the data situation in a certain region than
by the absolute uncertainty of the data and data sets which are excluded in one
region could be very helpful in other regions or for other substances.
Corrections of published data sets are justified in two different situations. The
first reason for corrections is that the data analysis reveals an obvious error in the
evaluation of the original experimental results. This is usually very difficult, if not
impossible, since it requires very detailed information on the experimental procedure. A common exception are data which are based on calibrations with some
kind of published primary data. If these data are known to be in error from comparisons with more accurate data, the secondary data can be corrected for this
error. Such a procedure is a true correction of the published data and its justification is unquestionable. The second reason for corrections is that systematic deviations of a data set from some kind of reference become obvious. Since theoretical
models yield highly accurate results only for the low density gas phase (see Sects.
3.1.1.1 and Eq. 4.2), corrections which are based on a theoretically founded reference can usually be applied only to gas phase data. Where such data do not meet
well known properties of the ideal gas in the limit of vanishing pressure, the observed differences can be used for corrections; see for instance the correction of
isobaric heat capacities of Bender (1982) applied by Smukala et al. (1999). When
a data set shows systematic deviations from significantly more accurate data in a
sufficiently broad region of overlap, these deviations may be used for corrections
as well. This kind of correction is only useful of course if the less accurate data set
covers regions which are not covered by the more accurate data; otherwise the less
accurate data could simply be omitted from the data set. As an example see the
correction which was applied by Tegeler et al. (1997/1999) to the data published
by Barreiros et al. (1982). Corrections which are just based on comparisons with
other data sets without any experimental background are theoretically questionable
and in general the term "adjustments" describes this procedure better than the term
"corrections': However, even pure adjustments are often justified due to the unsatisfactory way in which systematic errors are taken into account by the applied
weighting and fitting algorithms. Uncorrected data sets with significant systematic

90

4 Setting up Multiparameter Equations of State for Pure Substances

errors may affect the whole fit even if the data are met within the assigned uncertainty.
When working on equations of state for technical applications (see Sect. 6.2)
instead of reference equations of state, the applied weighting strategy has to be
changed. A characteristic feature of such equations is that they do not need to
represent certain data within their experimental uncertainty. Instead, they should
represent thermodynamic properties within uncertainties which are required to
fulfil some level of technical demands. Of course, the definition of the required
uncertainties is difficult, and to a certain degree subjective, but it is clear that they
do not depend on the accuracy which can be achieved in state-of-the-art experimental set-ups. If an uncertainty of t1.p / p ~ 0.2% is regarded as sufficient for example, a state-of-the-art ppT point which is weighted with its experimental uncertainty of appT/P = 0.02% could contribute (~ppT / appT)2 = 100 to the sum of
squares while it is still met within the required uncertainty. In a different region
where the best available data are accurate to appT/P = 0.1 % a point which is met
just within the required uncertainty would contribute only (~ppT / a pp T)2 = 4. Since
equations of state with simple functional forms, so desired for many technical
applications, are usually not able to represent all data within their experimental
accuracy the fitting algorithm would concentrate on an improved representation of
regions where highly accurate data are available while neglecting regions with less
accurate data. To avoid this effect it is better to replace a p by the required !J..yP
when working on such equations. Only where data with a p > !J..yP have to be considered due to the data situation, should these data be weighted with their experimental uncertainty. In this way data with different uncertainties have the same
influence on the fit as long as they are accurate enough to fulfil the demands on the
accuracy of the corresponding property.

4.4 Optimising the Functional Form

Until now it has been assumed that the functional form of the equation (the
mathematical form of every single term Ai in Eq. 4.43) and the length of the equation (the number of terms [ in Eq. 4.43) is known; only the determination of the
coefficients ni was discussed. In fact, the Ai contain additional "internal" parameters. For simple polynomial terms these are the exponents d i and ti in Eq. 3.25 and
for exponential terms, the density exponents in the exponential function, Pi, have to
be considered , too. When using more complex functional forms for the description of properties in the critical region, additional parameters have to be determined, see Sects. 4.5.3 and 4.5.4. Furthermore, not only the total number of terms,
but also the number of terms in each group of terms, thus the [Pol and [Exp in Eq.
3.25, have to be determined before the ni can be fitted. This whole procedure is
referred to as "optimisation" of the functional form.
In the past, the functional form of equations of state was determined by trial and
error strategies which were based only on the experience of the correlator. In 1974
the stepwise regression analysis proposed by Wagner (1974) was the first procedure which allowed an objective optimisation of functional forms. Wagner used

4.4 Optimising the Functional Form

91

his procedure to establish a new class of vapour pressure equations (see Wagner,
1974; Wagner and Pollak, 1974) but soon it was applied to equations of state as
well. This was the beginning of a continuing development of optimisation algorithms. Although these algorithms have become rather complex, their basic idea is
still the same: A regression matrix is set up based on the weighted data set and on
an extensive set of terms with different values for the respective internal parameters which is called a "bank of tenns': Then a mathematical algorithm which is
based on objective statistical criteria is applied to select the best combination of
terms out of the bank of terms - the terms in the final equation of state are always
a subset of the terms in the bank of terms. To guarantee a successful optimisation,
the bank of terms has to contain all terms which are regarded as promising candidates for use in the equation of state, or at least a sufficiently extensive set of such
terms. The number of terms which are considered in the bank of terms and in the
resulting equation is a measure for the complexity of the optimisation process.
While Wagner made his first step with the selection of 4 terms from 27 terms,
recent reference equations require a selection of about 40 terms out of up to 1000
terms.
Section 4.4.1 gives some criteria for a reasonable set-up of a bank of terms, assuming that only simple polynomial and exponential terms are used in the equation
of state; for special critical region terms, the corresponding information is given in
Sects. 4.5.3 and 4.5.4. The construction of the required regression matrix is discussed in Sect. 4.4.2 and Sects. 4.4.3-4.4.8 summarise the ongoing efforts to
improve the selection algorithms. A very recent algorithm which allows an optimisation based on data for different substances is discussed in Sect. 6.1.
4.4.1 Defining a Bank of Terms

Although the set-up of the bank of terms is crucial when developing an equation of
state, little information on reasonable set-ups has been published. Different authors
have published banks of terms which they have used to set up a certain equation of
state, but these banks of terms can hardly be regarded as universal since different
restrictions apply to the bank of terms which are based mainly on the optimisation
algorithm used, on the available computing power and on the requirements of the
considered substance.
For use with simple optimisation algorithms (see Wagner, 1974; de Reuck and
Armstrong, 1979) banks of terms with no more than 100 terms are recommended.
Some information on such banks of terms was given by Jacobsen et al. (1986b/c).
With the functional forms introduced in Eq. 3.25 such banks of terms can be written as

2: n

a f ('f, a) =

IPol

i=1

t
i 'f ,

ad, +

2: ni

lPoI +IExp

'f t,

ad, exp( -Yi aP')

(4.86)

i=IPoI+I

for set-ups which contain only simple polynomial and exponential terms. The only
difference between the general formulation of a Helmholtz equation, Eq. 3.25, and
Eq. 4.86 is that Eq. 3.25 does not contain all the terms from Eq. 4.86. Thus, [Pol

92

4 Setting up Multiparameter Equations of State for Pure Substances

and I Exp are smaller in Eq. 3.25. However, when establishing a Helmholtz equation
of state, the restriction of the bank of terms to 100 terms implies a significant preselection of terms which is very likely to affect the optimisation result. Although
the optimisation algorithm is guided by objective statistical criteria it cannot compensate for subjective criteria which were used when setting up the bank of terms.
Recent optimisation techniques and enhanced computing powers allow much
larger set-ups which gain an increased flexibility and a reduction of subjective
influences. For use with such algorithms, Eq. 4.86 can be rewritten as

(4.87)
For polynomial terms, i I D P needs to be an integer value larger than 0; otherwise
the resulting equation would not be conform to the virial expansion at low densities and higher density derivatives a r would not become 0 in the limit of vanishing
densities. Thus, Il:nn = 1 and D P == 1 are mandatory for a reasonable description of
the gas region. Common values for the upper limit of i depend on the application.
Values up to I!:u,x = 10 are common and improve the representation of properties in
the liquid region, especially at low temperatures. However, recent investigations of
the extrapolation behaviour of multiparameter equations of state (see Sect. 4.6.2 or
Span and Wagner, 1997) show that values of I~ax > 4 affect the representation of
properties at very high temperatures and pressures. Thus I~ax =4 should be used
for reference equations of state where the extrapolation behaviour is considered as
crucial. Values of I~ > 4 are useful for equations of state which are designed for
limited technical applications where the extrapolation behaviour to extremely high
temperatures and pressures is considered as arbitrary.
The temperature exponent j I Tf does not need to be an integer value. Sufficiently large values of Tf are important for accurate results since the representation
of the thermodynamic surface is very sensitive for changes in the temperature
exponents. Values used today range from Tf = 2 to Tf = 16 with Tf = 8 as a good
compromise. Negative values for jlTf are useful especially for equations with
limited range of validity again; Jy'min =- Tf is often used for such applications.
However, for the corresponding virial coefficient, negative temperature exponents
result in increasing absolute values in the limit of very high temperatures where the
virial coefficients are expected to approach 0 or a small finite value. Thus,
Jy'min = 0 should be used for reference equations of state. The upper value Jy'max is
arbitrary in most cases. While large temperature exponents were used in older
equations of state recent banks of terms are restricted to Jy'max =4 Tf in most cases.
Experience shows that usually only polynomial terms with small values of j I Tf are
selected by the optimisation algorithms. Especially in combination with restrictions on the number of polynomial terms as discussed in Sect. 4.6.2, set-ups with
Jy'max = Tf and thus with j I Tf ~ 1 can be useful , too. In this case the required
stronger temperature dependencies are shifted to the exponential terms completely.

4.4 Optimising the Functional Form

93

For exponential tenns, the same restrictions as discussed above are valid for the
parameters De, n,min, and J!,i,min since exponential terms contribute to the viral
expansion of an equation of state too, where the exponential function has to be
considered as
exp( -Y r}kl p' )

=I

Yk r)

kl p'

(Y k r}kl p.)2

"'--"--+---I!
2!

(4.88)

for the same reasons p~ must be an integer value. The upper limits of i / D% and
j / Tr.k are usually much higher for exponential terms. Values for i / D% may reach up
to 15 for equations where I~x is small; otherwise smaller limitations are reasonable. Special limitations apply only to exponential terms with k/ pe = 1 where
> '" 8 may affect the extrapolation to very high temperatures and
terms with i /
pressures. Larger values are useful especially in combination with k/ pe ~ 4. In this
case 0.5 can be used for D% in order to limit the size of the bank of terms.
Common values for k/ pe reach from 1 to 6, where pe is equal to 1 in all stateof-the-art Helmholtz equations. Values above k/ pe = 4 are useful especially if a
good representation of properties in the critical region is attempted without special
critical region terms. If such terms are used or if the representation of data in the
critical region is considered as less important, K~x can be restricted correspondingly. A possible variation of Yk has never been considered in banks of terms.
When the density is reduced with the critical density Yk = 1 is generally accepted as
the best choice. When different values are used for Pr, reasonable values for Yk can
be determined from Yk '" (Pr/ Pel for pe = 1.
The stepwidth which is used for the temperature exponents is still an important
parameter especially for exponential terms with k/ pe = 1; values of 2 - 8 are useful
for T~ 1 while .n.l.max = 4 Tr. 1 is a reasonable upper limit for the temperature exponents. For higher values of k / p~ both the upper limit and the stepwidth has to be
increased. Values of J~k.max '" 30 Tr.k are usual for the highest values of k/ pe and
even higher values have been used for single substances. The inverse stepwidth
varies usually from T12 = 2 to T1k = 0.5 for the exponential terms with high upper
limits for j / Tr.k' Again, the high temperature exponents improve the representation of properties in the critical region but they have to be used with extreme care.
When combined with small values for i / D'k these high temperature exponents
result in plots of relevant virial coefficients which are clearly too steep in the low
temperature limit and which may lead to unreasonable results for properties in the
gas phase at low temperatures. In this region experimental data are either scarce or
inaccurate in most cases and an erroneous behaviour of an equation of state is
difficult to detect.

4.4.1.1 Hard Sphere Terms

Based on the formulation proposed by Saager et al. (1992), the use of equations of
state which contain hard sphere terms has been described in detail in Sect. 3.1.3.1.
These and similar terms have been used in empirical multiparameter equations of
state, but they have never been included in an optimisation algorithm. Numerically

94

4 Setting up Multiparameter Equations of State for Pure Substances

stable nonlinear algorithms and good starting solutions are required to detennine
the adjustable parameters of Eq. 3.37, Pn Tn and rp - these parameters cannot be
detennined by optimisation procedures which are based on linear algorithms.
Thus, equations of state which use hard sphere terms have to be optimised using
the set-up which was formulated in Eq. 3.35. Like the ideal gas contribution, a~ the
contribution from the hard sphere term, a~ has to be detennined independently and
only the formulation for the remaining residual contribution, ar~ can be optimised;
the residua given in Sect. 4.3.2 have to be adapted to the changed definition of the
residual Helmholtz energy.
Theoretically, this approach is promising since it involves additional knowledge
from statistical thermodynamics in the establishment of empirical equations of
state - corresponding requests are formulated frequently, especially from scientists
who are engaged in statistical methods. This point of view is supported by graphs
like those given in Fig. 4.4. Figure 4.4a shows the Helmholtz energy surface of
methane as calculated from the reference equation of state published by Setzmann
and Wagner (1991). States in the two phase region were calculated according to
Eq. 3.84 and correspond to the stable phase equilibrium. Figure 4.4b shows the
corresponding plot for the residual Helmholtz energy
af(T,p)

=a(T,p)-aO(T,p)

(4.89)

and Fig. 4.4c shows the residual Helmholtz energy

ar*(T,p)

=a(T,p)-aO(T,p)-ah(T,p),

(4.90)

of a combined approach which introduces a hard sphere contribution calculated

from the corresponding equation of Muller et al. (1996). When comparing Figs.
4.4b and 4.4c, it seems as if the subtraction of the repulsive forces results in a
Helmholtz surface which is much easier to fit; correspondingly simpler functional
forms could be expected for empirical formulations which describe the residual
Helmholtz energy a'~
However, the main problem is not to model the Helmholtz surface itself but to
model the derivatives which are responsible for the representation of the relevant
thermodynamic properties. These derivatives can hardly be seen from absolute
plots like the one given in Fig. 4.4. Therefore, Fig. 4.5 shows the corresponding
plot for the isobaric heat capacity, cp , of methane, for the residual isobaric heat
capacity, cp-c~, and for the residual isobaric heat capacity after introduction of the
hard sphere term, cp-c~-c;. From this figure it becomes obvious that those features of the thermodynamic surface which are difficult to model depend on attractive forces and critical effects, but not on the repulsive forces which are described
by the hard sphere term. It is not easier to represent the residual heat capacity
given in Fig. 4.4c than it is to represent the one given in Fig. 4.4b. The fact that the
residual contribution vanishes in the limit of very high densities when using hard
sphere terms implies additional constraints which make the search for a suitable
functional form even more difficult.

4.4 Optimising the Functional Form

.::::I

95

100
0

00

..:.:: - 100
.....,
..:.::
'-'

--

....

- 100
..:.:: -200
~
- 300
'-'
-- - 400
*....
- 500
(::!
- 600
100

180

270
o

0'>

360

90

...:' 1

o~

~t?

Fig. 4.4. The Helmholtz free energy surface of methane as calculated from the reference equation of state by Setzmann and Wagner (1991). Figure a shows the Helmholtz energy a, b the
residual Helmholtz energy, a-ao, and c the residual Helmholtz energy, a-ao-a h , for an approach which introduces an additional hard sphere term.

Without any doubt, hard sphere terms are useful in theoretical models which try
to describe thermodynamic properties based on molecular parameters of the considered substance; such approaches need to distinguish between the effects of
attractive and repulsive forces. But they are of little use for empirical equations of
state which are macroscopical approaches, since they make the resulting equation

96

4 Setting up Multiparameter Equations of State for Pure Substances

;-..,

I
~

Oll

.....
~

'-"

-...

~~

;-..,

~
I

Oll

.....

'-"

-...

o~~

I
~~

;-..,

I
~

Oll

-,
~

'-"

-...

..c~~

3
400

320

240 /"l

$'
~C

160

o~~

u~
N

TIl(

t-

Fig. 4.5. The isobaric heat capacity of methane as calculated from the reference equation of
state by Setzmann and Wagner (199\). Figure a shows the isobaric heat capacity, cp, b the
residual isobaric heat capacity, cp-c~, and c the residual isobaric heat capacity, cr c~-c~, for
an approach which introduces an additional hard sphere term.

4.4 Optimising the Functional Form

97

more complicated and the empirical description of the remaining residual contribution is not simplified with respect to its development or with respect to its application. This example shows a common misunderstanding with regard to the use of
theoretical approaches in empirical formulations. The question which has to be
answered is not whether an approach is useful or elegant on a theoretical, molecular level - the question is whether it is able to simplify the empirical contribution
or whether it is able to improve the results on a macroscopical level. Until now, no
theoretical results have been found for describing the residual Helmholtz energy
which could be incorporated into empirical equations of state directly. In general,
the merit of theoretical approaches is still restricted to an improved knowledge on
the qualitative behaviour of certain properties in regions which are difficult to
access by experimental means, see for instance Sect. 4.5.1.

4.4.2 Setting Up a Regression Matrix

In general, optimisation procedures rely on linear algorithms for the determination
of the required quality criterion, the sum of squares of an equation of state which
uses the combination of terms which is investigated in the current optimisation
step. In Sect. 4.1, it was shown that the information which is necessary to determine the coefficients of an equation can be written in matrix notation as

AN=Q

(4.91)

for linear problems without constraints and as

(4.92)
for linear problems with constraints. The same formulations with exactly the same
set-up of the different elements (see Eqs. 4.10-4.11 and 4.17 -4.18) can be used
for optimisation procedures as well, with the only difference being that the parameter I which determines the size of the matrices and vectors corresponds to the
number of terms in the bank of terms. The residua which are required to evaluate
Eqs. 4.90 or 4.91 correspond to those defined in Sect. 4.3.2.
However, Eqs. 4.90 and 4.91 hold only the information which is necessary to
determine the coefficients ni of the current equation, and not the information on the
sum of squares. As can be shown by combination of Eqs. 4.5 and 4.8 (see Wagner,
1974) this information is given by

X'

=~( a,(:;~ YmJ

J-t ,~(
n,

"0 (xm,

Y;{"' ('mJ )

(4.93)

or in matrix notation by

X'=S-NQ'

with

s=~("o(:;~YmJJ

(4.94)

98

4 Setting up Multiparameter Equations of State for Pure Substances

and with QT according to Eq. 4.11 if the parameter vector N of an equation is

determined by a linear fit.
Thus, for problems without constraints, the regression matrix

B==[~T ~]

(4.95)

holds all the information which is required to determine both the current parameter
vector N and the current sum of squares X7 For problems with constraints,
Eq. 4.95 has to be extended correspondingly to

Q~

Q]

Qc

(4.96)

Since the regression matrix B is a symmetric (/ + C + 1) matrix it is sufficient, to

supply only half of it.
However, it is useless to solve Eqs. 4.90 or 4.91; for large banks of terms this
step would fail due to numerical problems and an equation of state which contains
all terms out of the bank of terms makes no sense. To determine both the coefficients and the sum of squares of the current equation of state, which contains only
a subset of the I terms in the bank of terms, the regression matrix B has to be transformed by the optimisation algorithm. The steps which are necessary to do so were
explained in detail by Wagner (1974) and are summarised in the following section.
4.4.3 The Stepwise Regression Analysis (SEEQ)

Regression analysises have been used in numerical mathematics for a long time.
However, the common procedures were unsuitable even for the solution of simple
thermodynamic problems such as the development of vapour pressure equations.
Backward regressions fail due to numerical restrictions since they require a solution of Eqs. 4.90 or 4.91 for all terms out of the bank of terms as a starting point
and simple forward regressions fail since they are not flexible enough to find the
best solution if intercorrelations between different terms in the equation occur. To
overcome these problems Wagner (1974) developed the stepwise regression
analysis, a forward regression which is able to perform backward steps if the current combination of terms fails with respect to certain advanced statistical criteria.
Wagner applied his method to vapour pressure equations first, but soon it was
adapted to equations of state (Pollak, 1974/1975). Based on a report to the
IUPAC Thermodynamic Tables Centre (Wagner, 1977) de Reuck reprogrammed
the algorithm (de Reuck, 1979), applied it to an equation of state (de Reuck and
Armstrong, 1979), and introduced the name SEEQ under which the algorithm
became known internationally. This algorithm is still in use today in different research groups.
Figure 4.6 shows a flow diagram of the original procedure proposed by Wagner. The algorithm starts with the selection of the most efficient term (1), that is the
term which yields the largest reduction to the sum of squares. To do so, the sum of

4.4 Optimising the Functional Form

99

Regression matrix

Eliminate term with

lowest significance

Fig.4.6. Flow diagram of the stepwise regression analysis developed by Wagner (1974).

squares which results from the addition of every single term in the bank of terms
which is not in the current formulation 3 (in the first loop no terms are in the formulation) has to be predicted. This can be done using the relation
b2

-~
Xn2 -b
- n,lL -b
D,lL
b

(4.97)

kk

In Sects. 4.4.3 -4.4.6 the expression "formulation" refers to the combination of terms which is
considered in the current stadium of the optimisation process. The best formulation becomes
the functional form of an equation of state when the results are written to a file after the optimisation procedure is finished.

100

4 Setting up Multiparameter Equations of State for Pure Substances

where 0 refers to the old formulation and n to the formulation which is extended by
the k-th term out of the bank of terms. The bij are elements of the regression matrix
B, where i refers to the i-th row and j to the j-th column of the matrix. With
L =I + C + 1, the element bLL refers to the element S which always holds the sum of
squares of the current formulation. For a deduction of Eq. 4.97 and the following
relations, see Wagner (1974).
After testing all terms which are not in the current formulation, the term k which
yields the smallest sum of squares is added to the current formulation and the regression matrix has to be transformed correspondingly. Table 4.3 summarises the
relations which are necessary for this transformation as given by Setzmann and
Wagner (1989a), where IN denotes a vector which is used to identify the terms
which are in the current formulation. The element IN(i) is equal to 1 if the i-th term
from the bank of terms is in the current formulation and it is 0 if not. The given
prescriptions assume that L :2: i :2: j, thus that the lower part of the symmetric matrix
B is used. For the addition of a term to the current formulation, the parameter c
becomes equal to 1.
In step 2, the significance of every single term in the current formulation is
tested proving whether a coefficient is significantly different from 0 or not by a
Student t test. The standard deviation of a coefficient ni which is in the current
formulation is given by
a

ni

=~bLLbii
V

with

v= M-N.

(4.98)

In Eq. 4.98, M is the number of data points which was used when setting up Band
N is the number of terms in the current formulation. When testing the assumption
that a certain coefficient ni does not deviate from 0 significantly, the student t
statistic compares this standard deviation with the value of the coefficient,
(4.99)
The value of the coefficient ni corresponds to the element bLi of the regression
matrix. The probability that the coefficient ni is different from 0 is given by

Si

((V+l)j r (v))
2

1
=~
r

-2-

I 1+--;)-2
v+l

t; [

t2

dt

(4.100)

where r is the gamma function. This relation can be evaluated using standard
software packages. Wagner proposed rejecting terms if their coefficients do not
deviate from 0 with a probability of at least 99.99 %. However, this value is typical
only for short correlation equations and smaller values may yield better results for
equations of state. To gain more flexibility, it is advantageous to use the corresponding limit as an adjustable parameter of the optimisation procedure (see also
Sect. 4.4.5).
If one or more of the terms in the current formulation fail the t test, the term k
with the lowest t value is eliminated from the current formulation. In step 3 IN(k)
is set equal to zero and the regression matrix is transformed according to the pre-

4.4 Optimising the Functional Form

101

Table 4.3. Prescriptions for the transfonnation of the regression matrix B when adding (c = 1) or
eliminating (c = 2) the k-th tenn from the bank of tenns to or from the current fonnulation
Transfonnation

Position of the element

i=j=k
i =k,j < k

or j

=k, i> k

i=j>k"
i=j<k
i>k,j>k,i~j

i>k, j<k
i < k, j < k, i
a

~j

the element bu is treated as a nonnal element with IN(L) = 0, see Eq. 4.97

scriptions given in Tab. 4.3 with c=2; the procedure goes back to the selection of
the next term. However, terms which have been rejected immediately before due to
a failed t test must not be selected again since this could lead into an endless cyclic
exchange of terms.
If all of the terms are significant according to the t test, the significance of the
whole formulation is tested (step 4) using the Fisher F statistic. The variance of the
current formulation is given by
a2

eq

= bvu

with

(4.101)

v=M-N.

From a combination of Eqs. 4.97 and 4.101, the variance of the formulation which
is truncated by the term k with the smallest t value can be calculated according to

a eq_1 =
2

bLL +v+l
blc /bkk

(4.102)

According to the Fisher F statistic, the probability that the difference between
and ifeq.l is not accidental can be calculated as
F

ifeq

/V-l)/2

fo (v+(v+l)ff+o.

df

(4.103)

where the F value is defined as F = c1.q.l /ifeq. Again, this relation can be evaluated
using common statistical program packages. However, the F statistic introduced by

102

4 Setting up Multiparameter Equations of State for Pure Substances

Wagner when working on vapour pressure equations may yield misleading results
for v > "" 500 due to rounding errors. Since more degrees of freedom are common
for the establishment of equations of state, where the number of used data, M, may
be as large as several thousand, the F statistic has to be replaced by an approximation in this case. For large values of v, the F distribution becomes equal to the
single sided normal distribution (see for instance Sachs, 1973). Thus, Eq. 4.103
can be replaced by
Fmin =exp[ z

2(2V+l)]
v(v+l)

(4.104)

for v>"" 100 where z is the standard-normal value which corresponds to the chosen
value of Smin for the single sided normal distribution. Wagner proposed a value of
Smin=99.99 % for the F test which corresponds to z = 3.719. But again different
values may be useful depending on the length of the correlation equation. The
corresponding standard-normal values can be obtained from tables or statistic
packages.
If the current formulation proves to be significant, the next term is added to the
formulation (step 1). If the formulation fails the f test, intercorrelation between
terms in the formulation is expected. In this case, exchanges of each term in the
formulation against each term in the bank of terms (step 5) which is not in the
current formulation are tried. The sum of squares which results from an exchange
of term n against term m can be predicted according to the prescriptions given in
Table 4.4. If one of the tested exchanges reduces the sum of squares, the corresponding exchange is carried out by removing term n (Table 4.3 with c =2) and
adding term m (Table 4.3 with c= 1). Subsequently, the significance of every single term and of the formulation as a whole is tested again. If one of the terms in the
formulation fails the t test, it is eliminated (step 3) and the procedure continues
with the addition of the next term. If the formulation as a whole is significantly
better than the formulation which is truncated by the term with the smallest value
of Si (F test), the next term is added as well.
However, if the formulation could not be improved by an exchange of terms or
if the improvement was not significant in terms of the F test, the current combination of terms is regarded as too long. In this case, the term with the smallest value
of Si is eliminated from the formulation. Finally, it is again tested whether the
truncated formulation can be improved by an exchange of terms; if improvements
are possible the corresponding exchanges are carried out.
The equation which results from the stepwise regression analysis is optimised
both with respect to its functional form and its length, where the length of the
equation is strongly influenced by the value of Smin which is used in the F test. The
coefficients ni of the final equation can be determined from the bLi of the regression matrix.
One characteristic feature of the stepwise regression analysis is that the optimisation algorithm relies only on the regression matrix. Thus, the algorithm can be
used for very different problems, as long as a regression matrix of the form which
was discussed in Sect. 4.4.2 can be supplied.

4.4 Optimising the Functional Form

103

Table 4.4. Prescriptions for the prediction of sum of squares which result from an exchange of
term n out of the current formulation against term m from the bank of terms as summarised by
Setzmann and Wagner (1989a).
Prescriptions for m > n

Prescriptions for m < n

c =b
m

Lm

+ bLnb bnm

nn
b2
d =b + nm
m
mm
bnn
2

2
bLn
X-I =bLL +bnn
2

2
Cm

Xmn =X-I-T
m

4.4.3.1 Introduction of a Pairwise Exchange

When using the stepwise regression analysis in the way described above, problems
arise from the fact that the exchange of terms cannot handle intercorrelated pairs of
terms which are important for the description of the thermodynamic surface. If an
exchange of one of these terms is tested according to the prescriptions given in
Table 4.4, the sum of squares becomes much larger since the second term becomes
inefficient in this way as well and no exchange will be carried out. Nevertheless,
another pair of terms which is contained in the bank of terms but not in the current
formulation may yield better results - the stepwise regression analysis fails at a
point which is important especially for longer correlation equations where such
intercorre1ated terms are common. Theoretically, it is clear that this problem can
be avoided if not just an exchange of single terms is tested, but also an exchange of
each pair of terms in the formulation against each pair of terms not in the current
formulation. However, such an exchange of pairs of terms has been regarded as
impossible due to restrictions of the available computing power.
Bischoff (1988) developed a modified stepwise regression analysis where an
exchange of pairs of terms is tested whenever the exchange of single terms was not
successful. Liiddecke (1991) compared this modified stepwise regression analysis
with a modified stepwise regression analysis without pairwise exchange and with
the more complex optimisation algorithm of Setzmann and Wagner (1989a), see
Sect. 4.4.5. He found that the stepwise regression analysis with double exchange
yields results which are as good as those of the Setzmann and Wagner algorithm
but with considerably lower computing times. However, this conclusion is true
only for relatively short correlation equations and small banks of terms - Liiddecke tested the selection of up to 12 terms out of banks of terms which contained
up to 100 terms.

104

4 Setting up Multiparameter Equations of State for Pure Substances

For an exchange of single terms, the number of combinations which has to be

tested can be calculated according to
nsingle

= N(I -

(4.105)

N),

where N is the number of terms in the current formulation and 1 the number of
terms in the bank of terms. For an exchange of pairs of terms, the number of possible combinations becomes
ndouble

=-N (N -1)(1 - N)(I -N -1).

4

(4.106)

To test an exchange of single terms, a maximum of 4 nsingle numerical operations

is required while the transformations of the regression matrix, which are necessary
to exchange two terms, require (/+C+ 1) . (I+C+2) numerical operations. Thus, with
regard to computing time the transformation of the regression matrix is the dominant factor for an exchange of single terms if 1 > '" 4 . N holds for the size of the
bank of terms. The procedure of Collmann et al. (1996), which is explained in
more detail below, requires up to 20 ndouble numerical operations to test a pairwise
exchange, while the number of operations which are necessary to transform the
regression matrix is increased only to 2 (/+C+ 1) (I+C+2) since four terms are
now involved. Thus, for pairwise exchanges the prediction of sums of squares for
possible exchanges is the dominant factor in the computing time.
Table 4.5 summarises results from Eqs. 4.105 and 4.106 for some typical optimisation tasks and gives approximate ratios of the maximum numbers of numerical
operations which are required for the prediction of the sums of squares and the
following transformation of the regression matrix. For simple optimisation problems, a modified stepwise regression analysis with pairwise exchange of terms may
be faster and as successful as the optimisation procedure of Setzmann and Wagner
(1989a) with its repeated calls of the regression algorithm (see Sect. 4.4.5). But for
longer equations and larger banks of terms, as they are typical for the development
of multiparameter equations of state, a single call of a pairwise exchange needs
more computing time than the Setzmann and Wagner algorithm. And for such
problems, the results of the Setzmann and Wagner algorithm with its heuristic
elements are still better than those of a modified regression analysis since the

Table 4.5. Number of possible exchanges and ratio of the maximum number of numerical
operations which are required for an exchange of single terms and pairs of terms
Number of terms in the
Formulation Bank of terms

4
12
20
35
40

27
100
400
650
1000

Number of possible exchanges

single exchange double exchange

92
1056
7600
21525
38400

1518
252648
l.368107
1.123.108
3.591.108

Ratio of required
numerical operations

27.1
349.3
1429.8
4402.4
6210.4

4.4 Optimising the Functional Form

105

flexibility of a pairwise exchange is not sufficient to resolve the complex intercorrelations in multiparameter equations completely. However, the enormous increase
of available computing power makes a modified regression analysis with pairwise
exchange interesting again even for the development of highly accurate equations
of state with a large number of terms - namely as an integral part of a modified
Setzmann and Wagner algorithm (see Sect. 4.4.5.2). This was the focus of the
work of Collmann et al. (1996) who developed a procedure similar to the one
described below for the required prediction of the sums of squares which result
from pairwise exchanges.
A pairwise exchange of terms involves four terms ki The corresponding elements of a vector f are equal to -1 for those two terms which are in the current
formulation and equal to 1 for those two terms which are not yet in the formulation. To predict the sum of squares which results from the pairwise exchange of
terms a partial matrix A is used which is determined from the regression matrix B
according to
b kJk

all
a2I
a3I
a4I
a5I

a32

a3 3

a42

a 43

a52

a 53

k2kJ

b k2k2

b k3kJ

b k3k2

b k3k3

aM

b k4k2

b k4k3

b k4k4

a54

b Lk1

bLk2

b Lk3

bLk4

a22

=
a55

k4 k J

(4.107)
b LL

The way in which Eq. 4.107 is written assumes kI < k2 < k3 < k 4 ; where this condition does not hold, the indices of the corresponding bij have to be changed to bji if
only the lower half of the symmetric regression matrix B is supplied to the optimisation program. From the matrix A, the sum of squares of a double exchange can
be predicted according to the prescriptions given in Table 4.3. However, it is not
necessary to transform the complete matrix 4 times. The exchange (addition or
elimination) of the term kl requires information from all elements of the matrix,
but to exchange the second term, k 2, only the matrix elements in the triangle a22a52-a55 are required. Thus, only these 10 elements have to be transformed when
exchanging k I To exchange the third term requires only the 6 elements of the
triangle a33-a53-a55 and for the fourth term only the 3 elements of the triangle
a44-a54-a55 are required. Finally, only the element a55 which is equal to the sum
of squares has to be transformed when exchanging the fourth term - to predict the
sum of squares of a pairwise exchange requires 20 numerical operations with matrix elements. Table 4.6 summarises the prescriptions for the required operations
as they result from an adaptation of Table 4.3 to this simpler problem.
To avoid unnecessary pairwise exchanges this routine is called only if the exchange of single terms was not successful. If the predicted sum of squares shows
possible improvements of the formulation, the double exchange which results in
the smallest sum of squares is carried out. The corresponding two terms are eliminated from the current formulation and the two terms from the bank of terms are
added - thus the whole regression matrix has to be transformed four times according to the prescriptions given in Table 4.3. Afterwards, the optimisation procedure
continues with testing the potential of further pairwise exchanges. If pairwise ex-

106

4 Setting up Multiparameter Equations of State for Pure Substances

Table 4.6. Prescriptions for the transfonnation of the working matrix A

Position of the element

Transformation

i=j>k"
i>k,j>k,i;;ej
Conditions

l::5k::54, (k+l)::5i::55,and (k+l)::5j::5i

f(k)
a

=-I for tenns k which have to be eliminated from the current fonnulation
f(k) = 1 for terms k which have to be added to the current fonnulation

the element a55 is treated as a nonnal element withf(5) = I

changes cannot improve the formulation any further, the regression procedure
continues as shown in Fig. 4.6.
4.4.4 The Evolutionary Optimisation Method (EOM)

When using the stepwise regression analysis for complex optimisation problems,
such as the development of an accurate multiparameter equation of state which
requires the selection of about 20 - 40 terms with two independent variables out of
banks with 300-1000 terms, it became obvious that a purely deterministic regression analysis is not flexible enough to find global optima. For different sophisticated optimisation problems, similar experiences were reported. It was shown
early in the seventies that non-deterministic random search strategies are superior
to deterministic procedures since they are able to leave local optima (see Rechenberg, 1973). However, none of the existing random search strategies could be
applied to the development of equations of state directly.
Therefore, Ewers and Wagner (1982a/b) developed the evolutionary optimisation method (EOM), a new optimisation algorithm which imitates known principles
of biological evolution. The evolutionary principles which formed the theoretical
basis of the algorithm were adapted to the development of equations of state wherever such adaptations could improve or simplify the resulting algorithm - it was
not the goal of Ewers and Wagner to develop a faithful imitation of biological
principles. The characteristic elements of the EOM are
Optimisation of a population of formulations, each represented by a parameter
vector which identifies the corresponding terms out of the bank of terms.
Usually the population contains 20 formulations.
Optimisation in generation cycles. Usually 150 generations are calculated
unless convergence occurs before this number of generations is finished.
Mating of diploid individuals and recombination of descendants from these
"parents':
Change of the "genetic code'; the parameter vector, by mutation. To speed up
the procedure, the probability of mutations is much higher than in biological

4.4 Optimising the Functional Form

107

systems. The "step width" of mutations is controlled by a teachable procedure

which allows different step widths for different terms in the formulation. The
terms which are to be changed are selected by a random algorithm.
Change of the "genetic code" by crossing over. Randomly selected parameter
vectors of descendants are detennined from mixtures of the parameter vectors
of the parents. In this way, large changes in the structure of formulations are
performed which are essential to leave local minima.
Selection of the descendants with the smallest sums of squares to form the
offspring generation.
Avoidance of "inbreeding" either by limitation of the maximum number of
descendants from a single pair of parents or, theoretically more profound, by
division of the population.
Although all these elements sound familiar, their implementation becomes extremely complicated and the whole algorithm requires much more computing time
than detenninistic procedures. Moreover, the whole algorithm is very specialised
in the development of equations of state. Beside the regression matrix, additional
information on predetennined features of the resulting equations of state, such as
the total length and the number of terms of certain forms, has to be supplied to the
EOM - in this sense the optimisation is incomplete.
Compared to the stepwise regression analysis, the results which were obtained
from the evolutionary optimisation method were impressive; the equation of state
of Schmidt and Wagner (1985) is still the accepted standard for thermodynamic
properties of oxygen. However, this algorithm has never been used by others due
to its complicated set-up, its high demands regarding the available computing
power, the incomplete optimisation, and its methodical restriction to the development of equations of state.
For the further development of optimisation strategies, an experience gainedfrom the work with the evolutionary optimisation method was decisive: There are
always certain combinations of terms which are characteristic for all formulations
which yield a good description of a given problem. The stepwise regression
analysis is not flexible enough to find these combinations, but when started with
the characteristic terms, it yields very good results even for complex problems.
This is the basic idea of the optimisation algorithm by Setzmann and Wagner
(1989a) which is described in the following section.
4.4.5 The Optimisation Algorithm by Setzmann and Wagner (OPTIM)

In 1989, Setzmann and Wagner (1989a) published a new optimisation algorithm

which combines a modified stepwise regression analysis with elements of the
evolutionary optimisation method such as mutation and optimisation of a population of correlations. This method was designed to overcome the shortcomings of
the EOM - its set-up is simpler, it yields a complete optimisation which includes
the length of the resulting correlation and the number of terms of certain forms,
and it can be used for very different problems. The name OPTIM was introduced

108

4 Setting up Multiparameter Equations of State for Pure Substances

when the corresponding programs were made available for other users (Setzmann
and Wagner, 1989b; Setzmann et aI., 1990).
Like the stepwise regression analysis, the whole procedure is based only on a
regression matrix (see Sect. 4.4.2) and on control parameters. The optimisation
procedure starts with reading these data, see the flow diagram which is given in
Fig. 4.7. The optimisation starts with the initialisation of the first parental generation (step 01). Each of the NP correlations (typically 6 to 8) in the population is
represented by a parameter vector Pi which holds the position of the corresponding
terms in the bank of terms. The NP parameter vectors are initialised by random
selection of Nmax terms out of the bank of terms, where Nmax is an estimate for the
maximum length of the correlation. The random selection is repeated NS times for
each parameter vector and the combination of terms with the smallest sum of
squares is used as a starting solution for the corresponding parameter vector. To
determine the sums of squares, the systems of normal equations which correspond
to the parameter vectors Pi have to be constructed from the regression matrix (see
for instance Eq. 4.107) and have to be solved as for linear fits, cf. Sect. 4.1. The
NP starting solutions form the O-th parental generation (step 02).
In the next step the formulations in the current generation are modified by
"mutation" (step 03), a simplified version of the mutation algorithm which was
used in the EOM. For each of the NP parameter vectors Pi, a limited number of
components Pn.old is selected randomly, where the number of selected terms should
be less than N12. New values are selected for these terms randomly, where the
procedures "neighbour" mutation and ''free'' mutation are used alternating from
parameter vector to parameter vector. When using the procedure neighbour mutation, the selected terms are replaced according to the prescription
Pn

= Pn.old + z(O,a)

(4.108)

where the z(O,a) are normally distributed random numbers which should cover a
range from -3 to 3 with an expected value of 0. The mutants which are determined
in this way are rather similar to the old formulations. When using free mutation the
new values for the selected Pn are determined by random selection from the bank
of terms with the only limitation that the selected term must not be in the current
formulation; mutants which are determined in this way are very different from the
old formulations. For each mutant the quality is determined in the way described
above. If the mutant has a smaller sum of squares, the original formulation is replaced by the mutant. The mutation procedure is applied NM times to each parameter vector in the population; Setzmann and Wagner (1989a) give an example
to explain the mutation process in more detail.
In the next step (04), the terms are selected which are considered especially suited
for the description of the given problem to use them for a starting solution in a
modified stepwise regression analysis. Thus, the terms are identified which are
used in different formulations out of the current population. Assuming that the best
formulations contain most of the well suited terms, only the NR+ 1 best formulations are considered. From those formulations, the terms are selected which are

4.4 Optimising the Functional Form

109

Initialise the first parental

generation by random selection
02

Start new generation

Test mutants of all formulations,

take over improved mutants
Select the most effective terms
as starting solutions

05

Repeated start of the modified

stepwise regression analysis
with different preselected terms

06

Replace the NR worst formulations

from the parental generation by
the regression results

No

Fig. 4.7. Flow diagram of the optimisation algorithm developed by Setzmann and Wagner
(1989a); for details on the modified stepwise regression analysis see Fig. 4.8.

contained in NR+ 1 formulations, NR formulations, and so on down to 2 formulations. With these terms the modified stepwise regression analysis is started NR
times (step 05).
A detailed flow diagram of the modified regression analysis is given in Fig. 4.8.
The regression analysis starts with the addition of the preselected terms (step Rl).
To do so, step 1 from Fig. 4.6 with the corresponding transformation of the regression matrix, see Table 4.3, is repeated until the starting solution contains all
preselected terms. In the next step (R2) the most important term from the bank of
terms is added. The sum of squares which results from the addition of a certain
term is predicted using Eq. 4.97 again.

110

4 Setting up Multiparameter Equations of State for Pure Substances

------------

,,

Add the preselected terms

Rl

Eliminate term with

lowest significance
If 6 < N < 21 mutation of the recent
R3 formulation; if possible exchange recen '----'--E-Ii-m-i-n-a-te-te-r-m-w-i-th---'
formulation against improved mutant
R4e
lowest significance

No

Yes

Eliminate the NC terms

R7a with lowest significance,
add constraints
Modified Stepwise Regression Analysis:

-------------------------- -------------- -----------------------------------

Regression formulation determined

Fig. 4.8. Flow diagram of the modified stepwise regression analysis which is used as an integra1 part of the optimisation a1gorithm developed by Setzmann and Wagner (1989a).

4.4 Optimising the Functional Form

111

After the most important term has been added, the mutation algorithm described
above is applied to the current formulation NM times (step R3) in order to gain
more flexibility in the deterministic set-up of the regression analysis. However,
experience has shown that this step may ruin a good starting solution when applied
at an early state and that the probability of successful mutations decreases with
increasing length of the formulation. Thus, this step is applied only to formulations
with more than 6 and less than 21 terms. The sums of squares of the mutants are
determined by linear fits based on the corresponding systems of normal equations.
At this point, it is important to use a copy of the regression matrix as the working
matrix for the regression analysis since the required matrix for the fit can only be
extracted from the unchanged regression matrix. If one of the mutations reduces
the sum of squares, the current formulation is exchanged against the mutant and
the working matrix is adapted to the new parameter vector by exchanging the necessary terms according to Table 4.3.
Just like the original stepwise regression analysis, the procedure continues with
statistical tests for the significance of every single term (step R4, Student t test, see
Eqs. 4.98 -4.100) and for the significance of the formulation as a whole (step R5,
Fisher F test, see Eqs. 4.10 1- 4.104). If one of the terms in the current formulation
fails the t test it is tested (step R4a) whether this is the term which was added in
step R2. If this is true, the term must not be again removed without other changes
since this would result in an endless exchange of terms. 4 Thus, it is tested whether
the formulation can be improved by the step "exchange of terms" (R4c) as described in Sect. 4.4.3; the sums of squares are predicted using the relations which
are summarised in Table 4.4. If an improvement by exchange of terms is not
possible, the term with the lowest statistical significance (the probability Sj of the t
test) is removed from the formulation (step R4e) and an exchange is tried again. If
the exchange was successful, the procedure continues either with the step mutation
(R3) if less than 11 terms are in the current formulation or otherwise with a new
Student t test (step R4) for all of the terms in the formulation.
If the formulation as a whole is significant according to the Fisher F test (step
R5) and if a predetermined maximum number of terms is not yet reached, the procedure continues with the addition of the next term (step R2). Thus, like the stepwise regression analysis, OPTIM uses the Fisher F test to optimise the length of
correlation equations but an additional upper limit is introduced to guarantee that
the formulations which result from different regression runs are comparable with
regard to their length. If the formulation fails the F test or if the maximum number
of terms is reached, an exchange of terms is tried again (step R6). If the exchange
was successful, the procedure continues either with the step mutation (R3) if less
than 11 terms are in the current formulation or with the Student t test (step R4). If

The procedure which is described here was used in the original Setzmann and Wagner algorithm. However, this procedure cannot avoid cyclic exchanges of 2 or more terms which resulted in endless loops in infrequent cases. Recent modifications of OPTIM additionally use an
algorithm which checks how many of the last 10 added terms are among the last 10 eliminated
terms too whenever adding a new term. If this ratio exceeds a certain limit (e.g. 7 out of 10)
cyclic exchanges are assumed and the regression analysis stops.

112

4 Setting up Multiparameter Equations of State for Pure Substances

the exchange was not successful and if no constraints were fonnulated, the final
fonnulation is determined.
If the resulting fonnulation has to fulfil NC constraints, the following procedure
is applied. The NC terms with the smallest significance (t test) are removed from
the completely developed regression fonnulation by using the prescriptions given
in Table 4.3. Then the NC constraints are added using the same prescriptions (step
R7a). The elements of the matrix which contain contributions from the constraints
can be treated just like all other elements. However, the constraints are not counted
as "terms" when determining the length of the current fonnulation and they cannot
be removed again. After the constraints are added, the procedure continues with
step R2 until the final length of the fonnulation as determined by the Fisher F test
or by N max is reached.
After the n-th regression run has come to an end, the transfonned working matrix is replaced by the original regression matrix and the procedure continues with
the (n+ 1)-th regression run using the next combination of preselected terms until
all NR regression runs are completed. Now the NR fonnulations of the parental
generation with the highest sums of squares are replaced by the fonnulations which
resulted from the NR regression runs (step 06). To guarantee a sufficient flexibility it is important to replace old fonnulations even if the sum of squares of the new
fonnulation is larger. However, the number of fonnulations in the population, NP,
has to be larger than the number of regression runs, NR, in order to preserve the
best fonnulations which have been found before. The new population fonns the
parental generation of the next generation and the procedure continues with step
02.
The optimisation algorithm ends if either all fonnulations in the population are
identical in step 07 (convergence) or if a predetermined number of generations
has been worked out. Convergence is common only for simple optimisation problems such as the development of ancillary equations for the vapour pressure or the

Table 4.7. Recommended values for the control parameters which are used in the optimisation
procedure by Setzmann and Wagner (OPTIM)
II

III

510
5200

530
5400

560
51000

8
6
250
60
0.9999
0.999

7
5
250
60
0.99
0.9

6
4
250
60
0.95
0.8

Characterisation of the problem

Maximum number of terms in an equation
Number of terms in the bank of terms

N max
I

Control Qarameters
Number of formulations in the population
Number of regression runs in each generation
Number of initial random selections'
Number of mutations
Probability value for the Student t test
Probability value for the Fisher F test

NP
NR
NS
NM
Si,min
Smin

smaller values are reasonable for NS when working with starting solutions

4.4 Optimising the Functional Form

113

orthobaric densities, but experience has shown that significant improvements are
very unlikely after about 10 generations even for complex optimisation problems.
Thus, Setzmann and Wagner (1989a) proposed that the algorithm should be
stopped generally after NG = 10 generations. In the modified OPTIM algorithms
which are in use today, NG became an adjustable control parameter. Values of
NG > 10 are still uncommon but under certain conditions smaller values are very
useful.
When using OPTIM in the way which was described above each optimisation
run starts from scratch. However, there may be preliminary equations of state
which proved to be very suitable for a given problem. To make use of such information the optimisation algorithm can be started with preselected terms. In this
case the first regression run starts not with those terms which are present in NR+ 1
formulations of the initial population but with PS preselected terms, where
PS ~ Nmax has to hold. If the preselected terms correspond to a good preliminary
equation the first regression run will result in a formulation of at least equal quality
and the population will already contain at least one good formulation after the first
generation has been worked out. Since the best formulation in the population is
never exchanged against regression results, a good starting solution5 has a strong
influence on the development of the whole optimisation process. However, formally the preselected terms are used only in the first regression analysis and can be
altered during the optimisation process just like all other terms.
When finalising an equation of state, it may be desirable to restrict the flexibility of the optimisation even further. To do so, recent modifications of OPTIM
support a mode where the regression runs of the first generation follow a completely different procedure. In this case, up to 10 regression runs are carried out in
the first generation starting with the PS preselected terms, with PS - 2 preselected
terms, with PS-4 preselected terms, and so on down to PS-18 or 1 preselected
terms. The best results of these regression runs are taken over into the current
popUlation if they are better than the randomly selected formulations in the parental generation. In this way, the population usually contains only descendants of the
starting solution at the end of the first generation and the preceding random selection looses any influence. The equation which results from the optimisation will
probably be very similar to the starting solution, but nevertheless this kind of "fine
adjustment" of the functional form may result in significantly improved sums of
squares. For such optimisations, the number of worked out generations can be
reduced to NG "" 4 in order to save computing time.
Based on first experiences with the new optimisation algorithm, Setzmann and
Wagner (1989a) published proposals for the control parameters of the optimisation
process. Based on further experiences and on the increased computing power
which has become available in the meantime, some of the values given by Setz5

The use of starting solutions is a crucial point since the quality criterion of the optimisation, the
sum of squares, is not sufficient to describe the quality of an equation of state completely. Important features like the extrapolation behaviour do not exercise an influence on the sum of
squares. Thus, it is sometimes important to preserve results which were found in previous optimisation runs, even if an optimisation without a starting solution could be more successful in
finding completely different optima with slightly smaller sums of squares.

114

4 Setting up Multiparameter Equations of State for Pure Substances

mann and Wagner are out of date. Table 4.7 summarises updated recommendations for the most important control parameters. However, just like the values
given by Setzmann and Wagner (1989a) these recommendations are only initial
guesses and may be altered. The probabilities which are used as limits for the
statistical tests have often to be adapted to the specific problem. The value of Smio
should be harmonised with reasonable values of N max in such a way that formulations which result from the regression analysis have on average I or 2 terms less
than the maximum length which is defined by N max The value for Si.mio needs to be
large enough to initiate exchanges of terms but it must not be too large. Otherwise
the procedure may stop early since none of the terms not present in the current
formulation passes the t test while the F test indicates that further terms have to be
added to the formulation. Thus, good problem specific values for the minimum
probabilities can only be found after evaluation of a number of optimisation runs.
4.4.5.1 Adapting OPTIM to Equations of State
One of the basic advantages of the optimisation algorithm of Setzmann and Wagner (1989a) is that it is not restricted to certain problems - any optimisation problem which allows the formulation of a suitable regression matrix (see Sect. 4.4.2)
can be dealt with. This flexibility resulted in applications which range from simple
equations for the melting pressure (see Wagner et aI., 1994) over different kinds of
multiparameter equations of state (see Setzmann and Wagner, 1991; Kruse, 1997)
up to optimisations of multi variant production processes (see Setzmann and Wagner, 1989b). However, regarding the development of equations of state, the performance of OPTIM can be improved if this versatility is sacrificed; corresponding
modifications were introduced by Span (1993).
Adaptations of OPTIM to certain problems generally require additional, problem dependent information on the set-up of the bank of terms. Such information is
not needed for the deterministic steps of the modified stepwise regression analysis
or for the "free" mutation where all terms in the bank of terms are treated in the
same way. But the step "neighbour" mutation can be formulated in a more efficient way if additional information on the set-up of the bank of terms is available.
U sing a section of a very small bank of terms as an example, Fig. 4.9 compares the
definition of neighbours as it is used in the algorithm of Setzmann and Wagner (a)
with the one used by Span (b). Based only on the position of a term in the bank of
terms, the neighbour mutation defined by Eq. 4.108 cannot consider the similarity
of terms appropriately when dealing with set-ups where 2 or more parameters are
varied. While the similar term No. 40 is not identified as neighbour of term
No. 45, the terms 47 -48 are treated as neighbours although their contribution is
very different from the one of term No. 45. Based on additional information on the
different I, J, and K in Eq. 4.87 the definition of neighbours proposed by Span
considers similarities of different terms appropriately. The normal distribution
which was assumed for z(O, a) in Eq. 4.108 was replaced by an adjustable distribution among neighbours of first, second and third degree and the ratio between
neighbour and free mutations also became adjustable. In this way, the non-

115

4.4 Optimising the Functional Form

37

... ...

33

0 4-c 2e-62

34

0 4-c 3e-6 2

35

04-c 4e- 62

36

0 4-c Se-62

o S-cl e-62

38

o S1'2e-62

39

OS1' 3e _62

40

OSt 4e _ 62

41

oS-c Se-6 2

06-c4 e- ,s2[

146

o6-c Se-6 2

143

If42 06-c le-,si )

147

0 6-c 2e-6'J

144
I 49

1145

0 6-c 3e-6 2

ol-c Se-6 4

o l-c le- 64 1148

01-c 2e-64

o 2-c l e-6 4

53

0 2-c 2e- ,s4

... ...

33

o 4-c2e-,s2

Ib 4

o 4-c 3e-6 2

35

o4-c4e-,s2

36

o 4-c Se-6 2

o S-c l e- 62

38

o S-c 2e-6 2

39

o 5-c 3e-6 2

40

05,4 e -62

41

o S-c Se-6 2

1142 o6-c le-6~ 43

0 6, 2e-6 2

2
44 0 6, 3e -6

45

o 6,4 e -6 2

46

0 6-c Se -6 2

49

o l -c 3e-6 4

50

ol-c 4 e-6 4

51

ol-c Se-6 4

52
b

37

47 o l-c le-,s'

48

o l , 2e- ,s'

52 02-c le-cl4

53

o 2, 2e- 64

Neighbours of _

1st degree

0 1-c 3e-6 4

150

ol-c 4e-64

151

......

......
_

2nd degree

3rddegree

Fig. 4.9. Section out of a small bank of terms as it might be used for the development of an
equation of state. The indicated fields are considered as "neighbours" of term No. 45 in the
algorithm by Setzmann and Wagner (a) and in the modified procedure proposed by Span (b).

deterministic steps in OPTIM have become more efficient and the flexibility of the
procedure can now be adjusted to certain tasks.
Another advancement is related to the fact that the algorithm of Setzmann and
Wagner optimises the length of a correlation equation without consideration of
limitations regarding the number of certain kinds of terms. However, both the
numerical expense which is required to evaluate an equation of state and its numerical stability depend not only on the number of terms in the equation but also
on the mathematical structure of these terms; sub limits regarding the maximum
number of terms of a certain functional form are useful, e.g. , for polynomial terms
(see Sect. 4.6.2) and for special critical region terms (see Sects. 4.5.3 and 4.5.4).
During the optimisation process, the information on the set-up of the current
formulation which is held in the parameter vector Pi is transferred to a vector Vi
with I elements in order to simplify necessary interrogations. The i-th element of
Vi is directly related to the i-th term in the bank of terms and is equal to 1 if this
term is part of the current formulation and 0 if not. Wherever the addition of a
term k from the bank of terms into the current formulation is tested, it is tested first
whether Vk is equal to O. If Vk is equal to 1, the corresponding term is already part
of the current formulation and must not be added a second time. To enable the use
of upper limits regarding the number of terms of a certain form, the vector Vi is
copied to a vector Vi which contains an "imaginary set-up" of the current formu-

116

4 Setting up Multiparameter Equations of State for Pure Substances

lation. If a term k is added to the current formulation by random selection (step 01

in Fig. 4.7), by mutation (step 03, step R3 in Fig. 4.8), by addition of a term (steps
R1, R2), or by exchange of a term (steps R4d, R6a), the values of Vk and vt are
changed from 0 to 1. The functional form6 of term k can be identified using the
different I, J, and Kin Eq. 4.87. If an upper limit for the number of terms of the
corresponding form is reached with the addition of k, all vj which represent terms j
with this functional form are set equal to 1. Instead of elements of Vi, elements of
Vi are used whenever the addition of a term is tested (steps R1, R2, R3, R4c,
R6).7 In this way no further terms of the corresponding functional form will be
added to the formulation. If a term of a functional form which has been blocked
before is removed from the current formulation (steps R4b, R4d, R4e, R6a, R7a)
the corresponding elements from the bank of terms are released again by setting
the vj equal to the Vj. In this way, Span (1993) realised upper limits for terms of
certain functional forms without unnecessary restriction of the flexibility of the
optimisation procedure.

4.4.5.2 Introduction of a Pairwise Exchange

The use of pairwise exchanges in the stepwise regression analysis has already been
explained in Sect. 4.4.3.1. As mentioned, the described procedure was developed
by Collmann et al. (1996) for use in the modified stepwise regression analysis as
an integral part of OPTIM. Pairwise exchanges are tested if no improvement could
be realised by single exchanges in R4c and R6.
However, the use of pairwise exchanges in OPTIM is still a challenge with regard to computing time. For the development of short equations of state for technical applications, where selections of 10 to 12 terms out of banks ofterms with less
than 600 terms are typical (see Sect. 6.2), pairwise exchanges could be used without limitations and proved to be a very effective supplement to the optimisation
algorithm of Setzmann and Wagner. For the development of reference equations of
state, where selections of 35 to 40 terms out of banks of terms with up to 1000
terms are typical, the unrestricted use of pairwise exchanges led to insupportable
computing times. However, the pairwise exchange proved to be effective especially for those functional forms where strong intercorrelations between the corresponding terms are typical, namely for Gaussian bell shaped terms (see Sect. 4.5.3)
and nonanalytical terms (see Sect. 4.5.4). To make use of this advantage, the
pairwise exchange can be modified in a way that an exchange is tested only for
pairs of terms which belong to certain functional forms. This step requires information on the different I, J, and K in Eq. 4.87 again; the functional forms which
have to be considered in a pairwise exchange are determined by corresponding
control parameters. Since the number of terms which have to be considered becomes smaller both with respect to the equation and to the bank of terms, the nu6
7

The realised program distinguishes polynomial, Gaussian bell shaped (Sect. 4.5.2), nonanalytic
(Sect. 4.5.3), and exponential tenns with e- d , e- d2 , , and e-d6 .
The only exception are exchanges (R4d, R6a) of 2 tenns with the same functional fonn since
the number of tenns using the respective functional fonn is not increased by such exchanges.

4.4 Optimising the Functional Form

117

merical expense is reduced significantly (c.f. Eq. 4.106) compared to an unrestricted pairwise exchange.
Restricted pairwise exchanges were used as an integral part of OPTIM for the
first time when the new reference equation of state for nitrogen (Span et aI.,
1998b) was developed. The next generation of workstations will show whether an
unrestricted pairwise exchange can improve the performance of OPTIM for the
development of reference equations of state further.
4.4.6 The Nonlinear Optimisation Algorithms by Tegeler et al.

The modified Setzmann and Wagner algorithm, which was described in the preceding sections, is certainly a very powerful tool for the optimisation of the functional form of an equation of state, but like all other procedures described it has
one major disadvantage: it is restricted to linear optimisation problems. This restriction results from four very fundamental features: the way in which the information on the data set is supplied to the optimisation procedure (as a regression
matrix) is suitable only for explicit data, the procedures which are used to predict
the sum of squares (see Eq. 4.97 and Tables 4.4 and 4.6) and to add and eliminate
terms (see Table 4.3) are applicable only for linear fits, and finally, linear fits are
used directly in the steps "initialisation" and "mutation': However, this restriction
has never been a severe problem, since data for properties which result in implicit
linear or implicit nonlinear relations to the Helmholtz energy and its derivatives
could be used in linearised form as described in Sect. 4.3.2. The influence of the
approximation which is implied by this linearisation could be eliminated by using
a cyclic process: the coefficients ni of the equation which results from the linear
optimisation are fitted nonlinearly and directly to all kinds of data; the necessary
linearisation of nonlinear data (see Sect. 4.3.2) is repeated with the nonlinearly
fitted equation before the linear optimisation algorithm is started again (see Setzmann and Wagner, 1991). This cyclic process is repeated until the difference between the results of linear and nonlinear fit is negligible.
During the development of the recent reference equation for argon (Tegeler et
aI., 1997/1999) the first highly accurate speed of sound data from spherical resonator measurements at high pressures became available (Ewing and Goodwin,
1992; Estrada-Alexanders and Trusler, 1995). These data reduced the uncertainty
of speeds of sounds in the high pressure gas phase by at least one order of magnitude and increased the demands on reference equations of state correspondingly.
With the cyclic process described above these data could no longer be represented
appropriately.
As an example, Fig. 4.10 shows percentage deviations between values calculated from a preliminary equation of state for argon and highly accurate experimentally determined speeds of sound on the 300 K isotherm. The open symbols
show deviations which are calculated between the linearly fitted equation and the
linearised data as they are used in linear optimisation procedures. The filled symbols represent the deviations between the original w(T,p) data and values calculated from a nonlinearly fitted equation. The grey lines indicate the experimental
uncertainty of the data. OPTIM resulted in a functional form which is able to de-

118

4 Setting up Multiparameter Equations of State for Pure Substances

0.02

-~
~

<l

0
0

- 0.02 0

T= 300 K
4

Pressure p l MPa

12

16

b., '"

Ewing and Goodwin (1992)

0 , Estrada-Alexanders and Trusler (1995)

Fig. 4.10. Percentage deviations 100 dw / W = 100(wexp-Wcalc}/Wexp between highly accurate

experimental results for the speed of sound in argon and values calculated from a preliminary
equation by Tegeler et al. (1997/ 1999). The open symbols correspond to deviations between
linearised data and a linearly fitted equation and the full symbols correspond to deviations between the original w(T,p} data and the same equation, fitted nonlinearly directly to the w(T.p}
data. The grey lines indicate the experimental uncertainty of the data.

scribe the linearised data very well, but on this level of accuracy the necessary
linearisation has distorted the experimental information. Fitted nonlinearly directly
to the original data the same equation of state is not able to represent the data
within their experimental uncertainty. This status could not be improved significantly by further loops of linear optimisation and nonlinear fit - the need for a
nonlinear optimisation algorithm became pressing for the first time.
4.4.6.1 The Nonlinear Quality Criterion
Common optimisation algorithms use a linear quality criterion, the sum of squares
which results from a linear fit of an equation which uses the current combination of
terms, to come to decisions in different steps of the procedure. OPTIM, for example, uses linear sums of squares during the initialisation (step 01 in Fig. 4.7) to
determine the quality of random selected combinations of terms, in the step mutation (02), and to select the worst formulations in the population (step 06), where
only previously determined sums of squares are needed in step 06. The modified
stepwise regression analysis which is incorporated in OPTIM uses the linear quality criterion to select the most efficient term from the bank of terms (step R2 in
Fig. 4.8), to determine the quality of mutants (step R3), to make decisions regarding single and pairwise exchanges of terms (steps R4c, R6), and to determine the
significance of single terms (steps R4, R4b, R4e, R7a) and of the formulation as a
whole (step RS). The results of the Student t and Fisher F tests which are used for
some of these decisions are advanced statistical criteria which are finally based on
the linear quality criterion.
A nonlinear optimisation algorithm needs to use a nonlinear quality criterion,
the sum of squares which results from a direct nonlinear fit of an equation which
uses the current combination of terms, for all these decisions. However, to replace

4.4 Optimising the Functional Form

119

the linear quality criterion by a nonlinear one has been regarded as almost impossible due to three reasons:
Iterative nonlinear fits require much more computing time than linear ones,
especially when compared to linear fits which are based on a given regression
matrix since the required system of linear normal equations (see Sect. 4.1) can
be derived from the regression matrix without time-consuming consideration of
the original data.
Nonlinear fits need a good starting solution for the coefficient vector N to
guarantee safe convergence.
The methods which are used in the linear optimisation process to predict sums
of squares without extensive matrix operations (see Eq. 4.97, Table 4.4, and
Table 4.6) cannot be adapted to a nonlinear quality criterion.
Considering the recent development of available computing power, the computing
time argument became less compelling at least for a direct comparison between
linear and nonlinear fits. Nevertheless, the starting value problem still had to be
solved and to replace the fast procedures which are used to predict sums of squares
by nonlinear fits completely is still unrealistic today.
For the steps "initialisation" (01), "mutation" (02, R3), and "elimination of a
term" (R4b, R4e, R7a) the linear quality criterion can be replaced by the nonlinear
one directly. The number of nonlinear fits which is required for these procedures is
proportional to the number of initial random selections (step 01), to the number of
mutations (steps 02, R3), or to the number of terms in the current formulation
(steps R4b, R4e, R7a). All of these parameters can be chosen to be less than about
50 and a corresponding number of nonlinear fits is realistic when using recent
computers.
To solve the problem of appropriate starting values, a regression matrix which
is based on explicit linear data and precorrelated implicit linear and implicit nonlinear data has to be supplied to the optimisation procedure. Based on this regression matrix, starting values for the parameter vector N can be determined by linear
fits when testing results of initial random selections (step 01). However, the preliminary equation which was used to linearise the data in the initial regression
matrix would have a strong influence on the following optimisation steps, if this
matrix is used unchanged for the next process. Thus it is important to repeat the
linearisation with the current formulation and to rebuild the regression matrix (see
Sect. 4.4.6.4) whenever the current formulation is changed during the optimisation
process.
The step "mutation" (02, R3) requires a nonlinear fit for each of the NM tested
mutants and the elimination of terms (steps R4b, R4e, R7a) requires (N-I) nonlinear fits for formulations which are truncated by one of the N terms in the formulation, except for the one added last. 8 The required starting values are again determined by linear fits based on the current regression matrix.
The required computing time for direct nonlinear fits becomes insupportable for
the step "addition of a term" which requires J - N nonlinear fits to identify the term
which yields the best result, where the number of terms in the bank of terms, J, can
8

The elimination procedure is changed significantly in the nonlinear algorithm, see Sect. 4.4.6.2.

120

4 Setting up Multiparameter Equations of State for Pure Substances

be as large as 1000. For the steps "exchange of terms" and "pairwise exchange of
terms'; the number of possibilities which have to be tested is far larger (see
Eq. 4.105 and Eq. 4.106, respectively). To avoid unrealistically large numbers of
nonlinear fits, the principle of linear preselection is applied:
Implicit linear and implicit nonlinear data are precorrelated based on the current formulation.
A linear regression matrix is set up based on the linear and linearised data. All
I terms and if necessary C constraints are considered.
These steps are applied in any case after each change of the current formulation in
order to be able to calculate an appropriate starting solution for the coefficient
vector N, see above.
The regression matrix is transformed N times according to the prescriptions
given in Table 4.3 to add those terms which are in the current formulation.
The linear quality criterion is determined in the same way as explained for the
corresponding step in Sect. 4.4.3 with the only difference that the result is
stored for each addition or exchange.
The possible modifications are arranged in order of increasing linear sum of
squares. Since it would require too much memory to store all results for possible exchanges or pairwise exchanges of terms when working with large banks
of terms and long equations (see Table 4.5) it is necessary to combine this step
with the preceding one.
Based on this linear ranking list, the results of the NA, NE, or NPE best modifications are examined by nonlinear fits. NA, NE, and NPE are control parameters which can be chosen independently for the steps "addition of a term';
"exchange of terms'; and "pairwise exchange of terms': This procedure is based
on the assumption that the modification which results in the best nonlinear sum
of square does not need to be the one with the best linear sum of squares, but
that it will be among the ones with the best linear sums of squares with high
probability .
The modification which results in the smallest nonlinear sum of squares is
carried out if it improves the current formulation.
If NA, NE, and NPE are chosen large enough, this procedure will find the solution
which yields the best nonlinear sum of squares with high probability without the
need to carry out an insupportably large number of nonlinear fits.
The effect of the linear preselection is illustrated in Fig. 4.11 which shows results from the development of a gas phase equation for argon (see Tegeler et aI.,
1997). The tested step is an exchange of terms for a current formulation with 13
terms and a bank of terms with 400 terms. Thus, 5031 possible modifications
(4644 if one of the terms in the formulation has been added in the previous step)
have to be considered in the linear ranking. The modification which yields the
smallest linear sum of squares (position 1 in Fig. 4.11) does not improve the nonlinear sum of squares at all. The second modification improves the nonlinear sum
of squares, but the modification with the best nonlinear sum of squares is found on
position 23 of the linear ranking list. With increasing rank, the scatter of the resulting nonlinear sums of squares increases, but at the same time the average sum of
square increases too - the probability that a better solution can be found some-

4.4 Optimising the Functional Form

121

400.-----------------------------------------,

'"
<1)

1il

g.

'"
o

'"

best linear
sum of squares
c9

to-.

300

200

0
COO

o
o

o
0

0
00

0> 00

o
CD

0
0
0
00

000

0
00 00

I'\.r\

0 -u- 0
0 0
0 0

0
0

00

0<0
00

omoo~~~-~-~--y-----------

~ best exchange

Position in the ranking list

Fig. 4.11. Nonlinear sums of squares which result from exchanges of terms which were placed
at the top of a ranking list established by linear preselection.

where beyond the 100th position of the ranking list is obviously very small. This
assessment has been verified by extensive tests with larger values of NE even for
problems like this, where the influence of implicit nonlinear data (speeds of sound)
on the sum of squares is exceptionally strong. Table 4.8 summarises recommendations for the control parameters NA, NE, and NPE as given by Tegeler et al.
(1997).
4.4.6.2 The Nonlinear Stepwise Regression Analysis (NLREG)

Based on the procedures described above, the modified stepwise regression analysis (see Sect. 4.4.5) can be transfonned into a quasi nonlinear optimisation algorithm. 9 The only steps which have not been adapted to the nonlinear quality criterion so far are the Student t test and the Fisher F test which are used to test the
significance of single coefficients and of the fonnulation as a whole in the stepwise
regression analysis. In fact, it turned out that these tests are unsuitable for a nonlinear optimisation algorithm. To evaluate the t test, the standard deviation of the
considered coefficient is required (see Eq. 4.98) which cannot be determined from
the regression matrix in the nonlinear case. And without infonnation on the results
of the t tests, it would be necessary to select the tenn with the smallest significance
by nonlinearly fitting N fonnulations which are truncated by one of the tenns in the
current fonnulation each to determine the variance a;q-I (see Eq. 4.102). This
9

In scientific literature the term "nonlinear regression" has been used before repeatedly (see e.g.
Kamei et aI., 1995; Tillner-Roth, 1996). However, it refers to procedures which simply incorporate nonlinear fits and not to nonlinear optimisation algorithms in these cases.

122

4 Setting up Multiparameter Equations of State for Pure Substances

Table 4.8. Recommended values for the control parameters NA, NE, and NPE of the linear
preselection
Characterisation of the problem
Maximum number of terms in an equation
Number of terms in the bank of terms

N fDJJX
I

:5;10
:5;200

II

III

:5;20
:5;400

:5;40
:5; 1000

80
100
200

100
200
400

Recommended number of nonlinear fits

Addition of a term
Exchange of single terms
Pairwise exchange of terms

NA
NE
NPE

50
80
150

numerical expense was considered as being unjustified and both the t test and the
F test were replaced by a simpler procedure. A flow diagram of the nonlinear

stepwise regression analysis is given in Fig. 4.12.

Like the modified stepwise regression analysis (see Fig. 4.8) the nonlinear
process starts with the addition of the preselected terms (step Rl in Fig. 4.12).
Based on the supplied regression matrix, a linear fit is carried out to determine
starting values for the coefficient vector N of the preselected formulation and the
nonlinear quality criterion is determined by a subsequent nonlinear fit. The regression matrix is updated using the resulting formulation for the required precorrelations. If the number of terms in the current formulation exceeds a certain limit
(usually N? Nmax-NC) the constraints are added to the regression matrix in step
R2a for the first time. Wherever the regression matrix is updated in one of the
following steps or wherever nonlinear fits are carried out, the constraints are considered , too.
In step R3 the nonlinear stepwise regression analysis continues with the addition of the most efficient term from the bank of terms. To select this term the
method of linear preselection (see Sect. 4.4.6.1) is used with nonlinear fits for the
NA combinations of terms which result in the best linear sums of squares. After
the most efficient term has been added, the significance of the achieved improvement is tested by a simple variance test (step R4). Therefore, the criterion

vnew
X~ew(M-N+l)
---=12
Void
Xold(M-N)

~smin'

(4.109)

with Smin as the minimum improvement of the variance which is expected from a
formulation with one additional term, M the number of data points used in the
nonlinear fit, N the number of terms in the current formulation, X~ew the nonlinear
sum of squares of the current combination of terms, and X~ld the nonlinear sum of
squares of the formulation truncated by the term added last, is tested. If Eq. 4.109
is fulfilled, the procedure continues with the step mutation (R5) as described in
Sect. 4.4.5 with the modifications explained in Sect. 4.4.5.1. If Eq. 4.109 is not
fulfilled the significance of all terms in the formulation except for the one added
last is tested by nonlinear fits of N - 1 formulations which are truncated by one

4.4 Optimising the Functional Form

123

Add constraints

~~N~o+l

R4a

Test all terms except

the one added last;
eliminate the one with
lowest significance

If 6 < N < 21 mutation of the recent

R5 formulation; if possible exchange recentt.-------.J
formulation against improved mutant

Yes

No

Nonlinear Stepwise Regression Analysis

Regression equation determined
Fig. 4.12. Flow diagram of the nonlinear stepwise regression analysis developed by Tegeler et
al. (1997).

term each. The term with the lowest significance is eliminated from the formulation and the procedure continues with the step mutation (RS). The quality of mutants is tested by nonlinear fits with starting solutions determined by preceding

124

4 Setting up Multiparameter Equations of State for Pure Substances

linear fits in step R5. If one of the mutants results in an improved sum of squares,
the current formulation is exchanged against the mutant.
To test possible improvements by single and pairwise exchanges of terms, the
method of linear preselection is used with subsequent nonlinear fits for the NE or
NPE combinations of terms which yield the smallest linear sums of squares (step
R7). A control parameter is used to decide whether pairwise exchanges are tested
or not. If the exchange with the smallest nonlinear sum of squares improves the
current formulation, the exchange is carried out and the current formulation is
replaced by the modified one.
If the current formulation was improved either by mutation or by an exchange
of terms (step R7), the procedure steps back to R5. If not, it continues with step R8
which determines whether a further term should be added or not. The variance test
in step R4 can only assess whether the term which was added last has improved the
formulation significantly; it cannot test the significance of the formulation as a
whole directly since the term which was added last is not necessarily the one with
the lowest significance. Thus, a single variance test yields no secure criterion for
the determination of the best length of the formulation. However, if the current
formulation fails in step R4, the term with the lowest significance (excluding the
term which was added last) is eliminated from the current formulation in step R4a.
Steps R5 and R6 try to improve the truncated formulation both by mutation and by
exchange of terms. If these steps are not successful and if the term which was
eliminated last is added again in step R3, the formulation which is tested in step R4
is identical with the previous one and it will fail again. However, this time another
term will be removed, since the one added last is always excluded from the variance test. If steps R5 and R6 fail again and if the term eliminated last is added in
step R3 again, it is obvious that the regression analysis cannot find a formulation
of the current length (N+ 1 terms) which is significant in terms of Eq. 4.109 - the
procedure would run into an endless cyclic exchange of two terms. In this case,
identical functional forms are found in step R8 repeatedly and the nonlinear regression analysis stops without consideration of the second criterion (N = N max ).
The best truncated formulation with N terms is selected as final regression formulation. In this way, the nonlinear regression analysis is able to optimise the length
of correlation equations as well~o
Unlike the original stepwise regression analysis (see Sect. 4.4.3), the nonlinear
regression analysis depends on preselected terms (step Rl) since the corresponding
formulation is used to reconstruct the regression matrix for the linear preselection
of the first term to be added. As a basically deterministic procedure with only a
single non-deterministic element it is not flexible enough to optimise highly accurate equations of state which are based on large banks of terms starting from
scratch. However, the nonlinear regression analysis is a very powerful tool to finalise highly accurate equations of state which were developed with OPTIM up to a
state where the repeated cycle out of linear optimisation and nonlinear fit cannot
improve the equation any further.

10

N max should not be much larger than the optimum length of the formulation; otherwise the way
in which constraints are considered (R2) becomes questionable.

4.4 Optimising the Functional Form

0.02

--

125

T = 300 K

<l

0
0

- 0.02 0

Pressure p l MPa

12

16

A '"

Ewing and Goodwin (1992)

D Estrada-Alexanders and Trusler (1995)

Fig. 4.13. Percentage deviations IOO.6.w/w = IOO(wexp-wcaJc)/Wexp between highly accurate

experimental results for the speed of sound in argon and values calculated from the best preliminary equation by Tegeler et al. (1997/ 1999) which resulted from linear optimisation algorithms
(open symbols) and from the final equation of Tegeler et al. (filled symbols). The grey lines indicates the experimental uncertainty of the data.

In this sense, the nonlinear regression analysis was used to finalise both the recent reference equations for argon (Tegeler et aI., 1997/1999) and nitrogen (Span
et aI., 1998b). Figure 4.13 illustrates the success of this approach using the speed
of sound data for argon which were already shown in Fig. 4.10 as an example. This
time the open symbols correspond to deviations between the original w(T,p) data
and the best preliminary equation which could be developed by using linear optimisation algorithms in combination with a nonlinear fitting routine. The filled
symbols correspond to deviations between the data and the final equation of
Tegeler et al. (1997/1999). Only the use of nonlinear optimisation algorithms
enabled a representation of these highly accurate speeds of sound within their
experimental uncertainty. Further advantages of the nonlinear optimisation algorithm result from the use of experimental phase equilibrium data directly (see Sect.
4.3.3, Eq. 4.72, Eqs. 4.78-4.83) without precorrelation. For well measured substances, the direct consideration of properties at phase equilibrium improved their
representation mainly close to the critical point. For other substances, information
on the saturated vapour density is very often scarce. In this case the nonlinear
stepwise regression analysis avoids the need to use a questionable ancillary equation for the saturated vapour density when precorrelating the Iinearised residua
(see Sect. 4.3.2, Eqs. 4.56 -4.58).

4.4.6.3 The Nonlinear Optimisation Algorithm (NLOPT)

The nonlinear regression analysis which was described in the preceding section
can be used as an integral part of a nonlinear version (NLOPT) of the modified
form of the optimisation algorithm of Setzmann and Wagner (1 989a). Basically,
there is no need to change the set-up of OPTIM as it was explained in Sect. 4.4.5.

126

4 Setting up Multiparameter Equations of State for Pure Substances

Apart from the steps within the nonlinear regression analysis, the nonlinear quality
criterion is needed to select the best randomly generated formulations of the first
parental generation (01 in Fig. 4.7), to test mutants (03) and to select the worst
formulations of the parental generation (06). All these steps can be performed
based on direct nonlinear fits (see also Sect. 4.4.6.1), no linear preselections are
needed. However, it turned out that two small modifications are useful.
When using NLOPT without preselected terms, the formulations of the first parental generation are usually not good enough to enable a reasonable reconstruction of the regression matrix. Thus, the starting solution for the regression matrix
should not be altered in the step 01 and 02 of the first generation. For the same
reason the regression matrix should not be altered in the nonlinear regression
analysis (05) until the current formulation has reached a predetermined length.
With regard to the selection of the most effective terms (04) it transpired that it is
advantageous to use all terms which are contained at least in NR+ 1 formulations,
NR formulations, and so on down to 2 formulations instead of those terms which
are contained exactly in the corresponding number of formulations as starting
solutions.
While NLREG can be used for reference equations of state on common workstations without insupportable computing times, II NLOPT with its repeated regression runs still places too high demands on the available computing power for
such optimisation tasks. Tegeler et al. (1997) used NLOPT successfully to develop
a highly accurate equation of state for argon which covers the density range up to
0.5 Pc including parts of the extended critical region. This problem required a
selection of 19 terms out of a bank of terms with 400 elements (category II in
Table 4.8) and was extremely influenced by highly accurate nonlinear data (see
Fig. 4.11 which illustrates one step of this optimisation). The routine use of
NLOPT for problems of category III remains for future generations of workstations or users who have privileged access to super computers.
4.4.6.4 Speeding Up Nonlinear Optimisation Algorithms

When working with NLREG or NLOPT, the high demands on computing power
result mainly from two characteristic steps:
The repeated set-up of the regression matrix B which is based on the original
explicit linear data and on linearised data (see Sect. 4.4.2). The regression matrix is needed for the linear steps of the algorithms.
Nonlinear fits which require vectors and matrices which have to be calculated
from the original data.
These steps require a great deal of computing power since they involve summations of all used data for each matrix element.
The contribution of implicit linear and nonlinear data to the required matrices
and vectors depends on the precorrelation of the data or on the current parameter
11

For problems of category III (see Table 4.8) computing times of about 5 hours are common for
NLREG when using the hardware supported quad precision mode (128 bit representation of
matrix elements) of an IBM RS 6000 1260 workstation.

4.4 Optimising the Functional Form

127

vectors in an unresolvable way. However, to avoid unnecessary summations, the

contribution of explicit linear data can be extracted from an initial regression matrix. To do so, it is necessary to rewrite Eqs. 4.10, 4.11 and 4.93 as

(4.110)

2: _.
a~ )
an

impl

m:;;;;i

and

[(

(4.111)

x,y,n /;t.n;

S
impl

M'"

LJ

ao x m' Y m )2 ) + S
2
Om

m=l

(4.112)

expl'

respectively. For implicit data, the more general notation is necessary since nonlinear relations between ~ and N cannot be resolved in the simple way which is
used for explicit linear data, see Sect. 4.3.3. The elements of the matrix C and of
the vector Qc (see Eqs. 4.17-4.18) depend neither on data nor on preliminary
equations of state and do not therefore change during the optimisation process.
Thus, the information on explicit linear data and on the constraints can be supplied
to nonlinear optimisation procedures as

expl

AexPI

C:

Qexpl

QeXPI]
c

Q~

Sexpl

(4.113)

or just as one part of the symmetric matrix Bexpl . To determine the current regression matrix B the contribution of the linearised implicit data has to be added to the
elements of Bexpl by carrying out the summations according to Eqs. 4.110-4.112.
Algorithms for the solution of non-linear systems of normal equations (see
Fletcher, 1971) need a matrix E which contains the derivatives of all equations
with respect to all unknown parameters (N and Ac), a vector F which holds the
current values of the functions which have to be solved and the sum of squares l
which results from the current values of Nand Ac. All these elements depend on N
and Ac and have to be updated in each step of the iterative procedure.

128

4 Setting up Multiparameter Equations of State for Pure Substances

The nonlinear fits in NLREG and NLOPT involve only those terms from the
bank of terms which are part of the current formulation. The contribution of explicit data and constraints can be calculated from the matrix

B *exp1

A exp1

c*

C *T

*T
Qexpl

Q~

(4.114)

where B~xp], A ~xp], C~ and Q~xpl hold only contributions from terms in the current
formulation; B~XPI can be extracted from B exp1 directly (see for instance Eq. 4.107).
Based on B~XPI the contribution of explicit linear data to E, F, and l can be calculated according to
(4.115)

(4.116)

and

=Sexpl-2N

*
T *
T
Qexpl+N Aexp1N+N

*,
T,
Jl.c-QcJl. c '

(4.117)

respectively. Only the information on implicit linear and nonlinear data has to be
added by evaluation of the corresponding summations. In this way the numerical
expense for nonlinear fits can be reduced significantly as well.

4.4.7 Automated Optimisation Algorithms

Basically, optimisation algorithms are designed for use by experienced experts
only. Simple fits are common features of thermodynamic property packages in
process simulators today, but optimisation algorithms are numerically both more
flexible and more sensitive. When using optimisation algorithms, inexperienced
users are likely to establish unreasonable correlations for one of the following
reasons:
Unreasonable set-ups of the bank of terms (see Sect. 4.4.1) result in unphysical
behaviour mostly outside of the region which is covered by accurate data, e.g.,
for the limit of the ideal gas at vanishing pressures.
Inadequate weighting of experimental data or undetected inconsistencies in the
data set result in an overfitting of certain properties in certain regions while
derived properties or properties in other regions are not represented adequately.
Repeated mathematical operations with large, very often ill conditioned matrices result in numerical problems which have to be detected and avoided by the
user.

4.4 Optimising the Functional Form

129

Thus, it is unlikely that mUltipurpose optimisation algorithms like those described

in the preceding sections will become common tools for broad groups of users in
the foreseeable future.
However, for certain tasks, optimisation algorithms can be applied by inexperienced users , too. In view of the problems mentioned above, the corresponding
tasks have to be well defined, allowing experts to define an appropriate bank of
terms in advance. They have to use data sets which cover a predetermined range of
validity with known accuracy and without significant inconsistencies, and which
guarantee an appropriate data distribution. And finally the algorithms have to be
able to detect and to avoid numerical problems automatically. Until now only a
single example for such "automated" optimisation algorithms has been published,
but the underlying problem is quite typical for a variety of engineering applications
and further specialised algorithms may follow.
4.4.7.1 Optimisation of Simplified Equations for Mixtures with
Constant Composition

The so called AGA8-DC92 equation of state by Starling and Savidge (1992) is an

internationally accepted standard for the calculation of natural gas properties in the
gas phase (see also Sect. 8.1.1) which allows calculations for temperatures from
143 K to 673 K and at pressures up to 280 MPa. The set-up of this equation is
rather complex due to the fact that broad composition ranges have to be considered
for a total of 21 typical natural gas components. For common engineering applications this complex set-up is needed only infrequently since natural gases do not
change their composition, e.g., during pipelining processes. Thus, much simpler
equations could be used to describe natural gases as pseudo pure fluids, but these
equations would be specific for a certain natural gas composition. 12
To establish such simplified equations, tests were made to determine whether
empirical correlations for the compression factor can be fitted to data which were
calculated from the AGA8-DC92 equation. For sufficiently broad ranges of validity, simple correlations with given functional form could never represent compression factors calculated from the AGA8-DC92 equation better than within
~/ZsO.I%-O.2%. Since the uncertainty of the AGA8-DC92 equation is
estimated to be ~/ Z S O.1 % in the technically most important regions, the
additional uncertainty of the simplified correlation affected the overall accuracy of
calculated properties significantly.
To overcome this problem, Span et al. (l995b) established an automated optimisation procedure which uses a modified form of the algorithm proposed by
Setzmann and Wagner (l989a) to establish accurate simplified equations in form
of the compression factor. Characteristic features of this procedure are:
Well defined ranges of validity. Four typical ranges of validity (region 1-4)
were defined for the simplified equation which reach from typical pipe lining
needs to the needs of possible future high pressure storage applications.
12

Usually the composition of a natural gas obtained from a certain source changes only very
slowly and simplified equations established once remain valid for months.

130

4 Setting up Multiparameter Equations of State for Pure Substances

Problem specific data grids. For each of the agreed regions, a suitable temperature/pressure grid was established in preceding tests. For the corresponding
grid, data are calculated from the AGA8-DC92 equation when establishing a
simplified equation for a certain composition. The calculated data are weighted
automatically, using the accuracy which is demanded from the simplified
equation instead of the uncertainty of the data.
Predefined banks of terms. For each of the regions, an efficient bank of terms
was determined in preceding tests with very different natural gas compositions.
The used banks of terms could be restricted to pure polynomial terms and to
exponential terms with exp(-02), see Sect. 4.4.1. In this way, small banks of
terms and short computing times could be realised.
Predefined starting solutions. As a result of the preceding tests an equation
was found for each of the four regions which yielded good results when fitted
to data for typical natural gas compositions (t:lZ/Z~0.1%-0.2%, see
above). If these equations are used as starting solutions, the optimisation algorithm finds equations with very accurate functional forms both quickly and reliably.
Automated detection of numerical problems. Numerical problems which may
result from rounding errors during the optimisation process are detected automatically by a comparison of the sum of squares calculated from the regression
matrix with the one calculated from a direct comparison of the resulting equation with the used data. If these sums of squares do not agree with each other,
rounding errors have affected the optimisation process. In this case, the optimisation procedure is repeated with a reduced number of terms in the equation
and, if the second attempt fails as well, the data are scattered randomly to
simulate a small experimental random error (see the conditions for using the
maximum likelihood method in the way it is used, Sect. 4.1 - violations of
these conditions are likely to result in ill conditioned matrices). However, it
turned out that these precautions are necessary only in very rare situations.
Table 4.9 summarises characteristic parameters for the four regions of validity;
allowable gas compositions are defined by the corresponding limits of the AGA8DC92 equation. Within these ranges of validity, the automated optimisation algorithm establishes simplified equations of state which deviate from the AGA8-

Table 4.9. Definition of the four regions of validity considered by Span et al. (1995b) and
characteristic parameters of the corresponding optimisation problems.

Region

Temperature
range

Pressure
limit

1
2
3
4

263 K-338 K
268 K-353 K
273 K-423 K
273K-473K

12MPa
30MPa
60MPa
120MPa

Bank of terms
No. of data
calc. from
polynomial exponential
AGA8-DC92
terms
terms
330
420
576
810

65
35
35
35

63
63
63

Max. No. of
terms in the
equation
6
9
13
14

4.4 Optimising the Functional Form

131

0.05
~

<

N<
.......

-------- --- --------~ ----- ----- -----............

...

,-.,

"

.~

I
<

O.02%

...

N<

'-"

0
0

Data calculated from the AGA8 -DC92 equation:

293 K )1 023 K . 353 K X 393 K . 433 K ... 473 K
X 273 K

-0.05 0

60

40

20

80

100

120

Pressure p / MPa
Fig. 4.14. Percentage deviations between compression factors calculated from the AGA8-DC92
equation and from an optimised simplified equation which is valid for region four of a typical
natural gas.

DC92 equation by t;Z/ Z :s: 0.05 %; for typical natural gas compositions deviations of t;Z/Z:S:0.02% can be realised in general. Figure 4.14 shows deviations
between values calculated from the AGA8-DC92 equation and a simplified equation which describes region four of a typical natural gas using the functional form
10

14

;=\

;=\\

Z(T,p)=~=I+ 2: n ;odir (i + 2:n;odir(ie-02 ;

pRT

(4.118)

for the values of nj, d j, and tj and the composition of the described natural gas, see
Span et al. (1 995b ).
The optimisation algorithm described above is integrated in a program which
contains different direct calculation modes and a composition data base. The user
only has to select a natural gas composition and a range of validity to start the
optimisation process - all other process parameters are predetermined based on
the selected region. Due to some numerical improvements, small banks of terms,
and the use of good starting solutions, the necessary computing time could be
reduced to less than one minute on recent personal computers even for the development of equations which describe region four. Under such circumstances optimisation algorithms can be applied even by completely inexperienced users.
4.4.8 Independent Developments and Future Perspectives

When reading Sects. 4.4.3 -4.4.7 one may think of this description of optimisation
procedures as unbalanced in some way. However, this impression is wrong - the

132

4 Setting up Multiparameter Equations of State for Pure Substances

development of optimisation algorithms for equations of state has been carried out
by the group around W. Wagner almost exclusively during the last two decades.
Besides these developments only one completely independent optimisation algorithm for equations of state has been published, namely the "simulated annealing" algorithm by Kirkpatrick et al. (1983) and Cerny (1985) which was adapted
to equations of state by Shubert and Ely (1995). Shubert and Ely compare equations of state for the refrigerant R134a which were developed with their simulated
annealing algorithm and with the simple stepwise regression analysis (see Sect.
4.4.3). Based on this comparison the simulated annealing algorithm seems to be
slightly inferior even to the stepwise regression analysis. More precise conclusions
cannot be drawn yet, since the algorithm of Shubert and Ely (1995) has not been
published in sufficient detail to carry out additional comparisons. Nevertheless, the
basic idea of simulated annealing, which is a non-deterministic algorithm like the
EOM (see Sect. 4.4.4), is promising and continued work on this topic could
stimulate the whole development of optimisation algorithms.
A possibly trend-setting result was reported by Lemmon and Jacobsen (1998/
1999b). Lemmon and Jacobsen included the temperature and density exponents of
equations of state which were optimised using the stepwise regression analysis into
direct nonlinear fits. When using the results of the optimisation algorithm as a
starting solution, the nonlinear fit could improve the equations further in this way
and the least significant terms in the equations could be eliminated without loss of
quality. In terms of optimisation nomenclature this method could be referred to as
a "manual backward regression': When using recent optimisation algorithms, the
advantage of fitted temperature exponents is anticipated partly since these algorithms allow much narrower graduations of the temperature exponents (see Sect.
4.4.1). However, even a very large bank of terms cannot mimic the results of a
direct fit completely. Thus, it would be very interesting to test whether this kind of
backward regression in combination with a direct nonlinear fit of temperature
exponents could be included, e.g., into the nonlinear optimisation algorithm of
Tegeler et al. (1997), see Sect. 4.4.6.
Another major challenge with regard to optimisation algorithms is the development of a completely nonlinear algorithm. The algorithm of Tegeler et al. uses the
nonlinear quality criterion for all relevant decisions and results in the combination
of terms which yields the smallest nonlinear sum of squares. However, both
NLREG and NLOPT still depend on linear preselections and cannot be used for
problems where important data are in a key position which cannot be linearised in this sense they are only quasi nonlinear algorithms. This restriction becomes a
problem, e.g., when using inequalities as constraints (see Eq. 3.109) which can
only be considered in nonlinear fits. Work on highly accurate equations of state
which avoid unreasonable plots in the instable part of the vapour-liquid two phase
region (see Sect. 3.3.5.1) is one of the ambitious topics still pending and thus, the
need for a completely nonlinear optimisation algorithm may soon become urgent.
Optimisation algorithms for the development of substance specific functional
forms for equations of state will certainly be a matter of ongoing research and the
idea of simultaneous optimisation (see Sect. 6.1) has only been recently formulated.

4.5 Describing Properties in the Critical Region

133

4.5 Describing Properties in the Critical Region

Like all critical points, the gas-liquid critical point of a pure fluid is characterised
by the fact that stable equilibria become indifferent. At the critical point, the critical isotherm, the saturated vapour curve and the saturated liquid curve join each
other. The isothermal compressibility of the fluid becomes infinite; other properties stay finite, become infinite, or become infinitely small, where the diverging
compressibility explains some but not all of these anomalies. Due to gravity and
diverging compressibility, systems with finite characteristic length become nonhomogeneous when approaching the critical point; density stratifications become a
characteristic feature of equilibrium systems. Increasing instability leads to strong
responses on small perturbations of an equilibrium system and the critical enhancement of transport properties such as the thermal diffusivity causes long
equilibration times. On a microscopic level indifferent equilibria result in properties which are determined not only by interactions between adjoining molecules
but by interactions within large clusters of molecules. The characteristic length of
these clusters, the correlation length, is still very small on macroscopic scales, but
it diverges when asymptotically approaching the critical point.
The region where all these effects are essential for the properties of a fluid is
called the "critical region': Its limitations depend on subjective assessment to a
certain degree; common values which can be found in literature are of the order 0.99::;; TITc::;; 1.1 and 0.7 ::;;plPc::;; 1.3. The extended critical region which is
defined by these limits is useful for an assessment of simple equations of state, but
especially with respect to temperature it is far too large to single out the region
which is strongly influenced by critical effects. In this sense and for an assessment
of advanced equations of state, the limits 0.998::;; TITc::;; 1.01 and
0.75 ::;; P I Pc ::;; 1.25 are useful to characterise the critical region. For certain questions, asymptotic regions have to be considered which are even smaller.
The accurate description of thermodynamic properties in the critical region is an
interesting challenge for theoreticians, for experimentalists, and last but not least
for those who are engaged in the development of empirical equations of state. The
theoretical treatment of critical effects resulted in an interesting edifice of thoughts
with a high degree of internal consistency. The resulting models do not depend on
the substance considered and are able to predict the observed critical effects
qualitatively correctly, but an accurate description of properties still requires a
considerable number of substance-specific fitted parameters. Experimental work
on properties in the critical region has resulted in highly complex adaptations of
basically known experimental techniques (see for instance Edwards, 1984; Wagner
et aI., 1992) as well as in experimental techniques which are applicable only under
the conditions found in the critical region (see for instance Straub, 1967). And
finally, sophisticated empirical equations of state have been developed both as a
reaction to increasing technical interest in accurate data for properties in the extended critical region and to keep pace with experimental and theoretical developments.
In accordance with the orientation of the whole book, this section focuses on the
development of empirical equations of state which describe properties in the criti-

134

4 Setting up Multiparameter Equations of State for Pure Substances

cal region accurately. The theoretical background is described in Sect. 4.5.1 only
as far as necessary to understand both the occurring problems and the semiempiric
approaches which are discussed in Sect. 4.5.5. For a more detailed description of
the underlying theory see Sengers and Levelt Sengers (1986), Albright et al.
(1986), Luettmer-Strathmann et al. (1992), and Sengers (1994).
4.5.1 Predictions from Theory
The key idea of theories which describe critical behaviour is that of critical-point
universality. Systems near the critical points are classified in terms of universality
classes. Just like magnetic systems close to their critical point, fluids near ordinary
critical points are assumed to belong to the universality class of three-dimensional
Ising-like systems, i.e., systems with short-range forces and a scalar order parameter, e.g., the density of the fluid. Usually such systems are described by a combination of two independent variables, namely a variable h which measures the distance from the coexistence curve and a variable r which measures the distance
from the critical point along this curve. Unfortunately, hand r are not identical
with any known thermodynamic properties - they have to be determined recursively, e.g., from the temperature T and the chemical potential fl.
The renormalisation group theory introduced by Wilson (1974; for an introduction see Wilson, 1979/ 1986) extended the theoretical understanding of critical
phenomena in Ising-like systems. With this theory, it could be shown that an approach which was originally introduced much earlier by Verschaffelt (1896) on a
purely empirical basis is valid and that it allows a universal description of systems
which belong to the same universality class: along certain paths throughout the
critical region thermodynamic properties can be described by so called "power
Table 4.10. Examples for power laws describing thermodynamic properties along certain paths
throughout the critical region
Power law

(p' - p") - (Tc - T}fJ

KT-(T-Tc)

-y

Ip- Pcl-Ip- pl
Cv

-IT-Tcl- a

Values determined by evaluation of

RG theo." 3-pt. eq.b 5-pt. eq.c

Described path

erit. expo

phase boundary

f3

0.326

0.002

critical isochore

1.239

0.002

critical isotherm

4.80

0.02

critical isochore

0.1l0 0.003

0.5

0.25

0/0.5 d

a according to Sengers and Levelt Sengers (1986)

b at the critical point (ap / ap)T = 0 and (a 2p / ap2)T = 0 with (a 3p / ap3h > 0 hold

c at the critical point (a 3p/ ap3)T =0 and (a 4 p/ap4)T =0 with (a 5p/ ap5)T

0 hold

additiononally to the conditions given in b

d

a2 = 0.5 when approaching the critical point from the two phase region, T - Tc < 0

4.5 Describing Properties in the Critical Region

135

laws': The non integer exponents of these relations, the so called "critical exponents'; are universal within a universality class; only the proportionality factors,
the so called "critical amplitudes'; depend on the considered system. Table 4.10
summarises the power laws which are most important for the description of thermodynamic properties in the critical region of pure fluids. The values for the corresponding critical exponents which are quoted as results of the renonnalisation
group theory were taken from Sengers and Levelt Sengers (1986); theoretical
values given by other authors agree within the given uncertainties in general.
The evaluation of typical empirical equations of state yields different values for
the critical exponents (for a summary, see Table 4.10). When assuming that an
equation of state is analytic at the critical point, the pressure which results from
this equation can be written as a Taylor expansion according to
p(T,p)- Pc

=(:~ l(p-pc)+(:~)p (T-Tc)

+;, -((:; lO(p-p,j' +{;':T }(P-P,jo(T-T,j+(::' ), o(T-T,j'1

+;,{(:no(p-p,j'+o}

(4119)

where the derivatives have to be calculated at the critical point. This Taylor expansion is valid for simple pressure explicit equations such as cubic equations of
state as well as for highly accurate empirical equations of state which are explicit
in the Helmholtz energy. If the formulation for the Helmholtz energy is analytic,
the resulting relation for the pressure (see Eq. 3.30) is analytic too and this is the
only assumption behind Eq. 4.119.
For most analytic equations of state, the first two derivatives of pressure with
respect to density become 0 at the critical point while the third one is larger than 0
(3-point contact, see also Table 4.10, footnote b); quite often accurate equations
are constrained to meet these conditions and p(Tc, Pc) = Pc exactly at Tc and Pc to
reproduce selected critical data exactly. With these conditions, the power series
which describes the shape of the critical isotherm (T - Tc =0) becomes

(a

1 -3p
p(T,;,p)-Pc=_
3 ) .(p-pJ3 + .....
3! ap T

(4.120)

with the leading exponent 15 = 3. Where equations of state are designed to realise a
5-point contact (see Table 4.10, footnote c) the corresponding power series becomes
p(Tc'p)- Pc

=~.(a5~)
5! ap

.(p_ pJ 5 + .....

(4.121)

with 15 = 5. Along the critical isochore the isothermal compressibility becomes

136

4 Setting up Multiparameter Equations of State for Pure Substances

IC T

(T,Pc)=_1 (ap)-l =_1

Pc ap T
Pc

[(~)'(T_Tc)+
....)-l
apaT

(4.122)

and the critical exponent which describes the temperature dependence of ICT becomes y = 1 both for equations with 3- and 5-point contact.
It is slightly more complicated to determine the values for the critical exponent
{3. With Eq. 4.119, the phase equilibrium condition

p(T,p') = p(T,p")

(4.123)

can be rewritten as

a2p ).(P'_P )'(T-T )+~(a3p) '(p'_p)3 =

( apaT
c
c
3! ap3 T
c
a2p )'(P"_P )'(T_T)+~(a3p) '(p"_p )3
( apaT
c
c
3! ap3
c
,

(4.124)

or simplified as

when considering the limit of vanishing distance to the critical point. The Pij in
Eqs. 4.125-4.134 indicate the i-th derivative of P with respect to p and the j-th
derivative with respect to T. When inserting the power law

(p' - pO) = C P. (Tc - T)P

(4.126)
(4.127)

as
Eq. 4.125 yields
Pll

c' 3(T -T )3 P = -p c" (T -T )l+P _ P30 c,,3(T -T )3 P (4128)

Cp, (T -Tc )l+P + P30
6 P
c
11 P
c
6 pc' .

and thus

P11 (c'P+c,,)+P30(c,3+c"3)(T_T)2P-l_0
P
6 P
P
c
-.

(4.129)

This equation can only be fulfilled independently from T if {3 =0.5 holds. For an
equation with 5-point contact, Eq. 4.125 becomes

P11(P' - Pc)(T -Tc)+ ~~~ (p' - Pc)s

= P11(P" - Pc)(T -Tc)+ ~~~ (p" - PC>s,

(4.130)

and thus Eq. 4.129 becomes

(c'S+c"S)(T_T)4 P-l =0
P11 (c'P+C,,)+P30
P 120 P
pc'

(4.131)

4.5 Describing Properties in the Critical Region

137

resulting in {3 = 0.25.
To determine the critical exponent which results from the increase of the isochoric heat capacity in the homogeneous phase, it is not necessary to develop
power series. Since the isochoric heat capacity is linked directly to a derivative of
the reduced Helmholtz energy (see Eq. 3.14), the isochoric heat capacity has to be
analytic for any analytic correlation in terms of the Helmholtz energy. Thus the
isochoric heat capacity stays finite at the critical point and a =0 holds for T> Tc.
For the isochoric heat capacity on the critical isochore of the two phase system at
T < Tc the situation becomes more complicated. The isochoric heat capacity has to
be calculated from the heat capacities of the saturated vapour and liquid and the
contribution
T

[(!i)p -~

(4.132)

dc - - - - . -'=----,,-'-.,,-----=.....
p)
v
Rp,2
ap T

(a

has to be considered additionally when investigating the corresponding critical

exponent a2 (see Eq. 3.80). With Ps(dT) =p(dT, !yJ') =p(dT, !yJ"), the power
series for the vapour pressure can be written as
Ps = Pc + pO\(T-TJ+ Pu c2,8(T-TJ

1+2,8

+.....

(4.133)

The exponent 2{3 results from the fact that the first term in the power series for the
is symmetric for p' and/" and cancoexisting densities which goes with (Tc cels out; the second term which is antisymmetric goes with (Tc - T) and determines the corresponding contribution. The value of c2fJ is arbitrary in this context.
By examining the required limit, the relation Eq. 4.132 can be rewritten with
Eq. 4.133 as

rI

(7;, - T)2,8 .
(7;, _T)4,8

(7;, - T)1+2,8

[POI + P02 (7;, - T) - POI -

(7;, - T) 2P

(7;, - T)
(4.134)

for {3 ~ 0.5 and with C as an arbitrary constant. Thus, the contribution dC v becomes
independent of infinitesimal temperature changes around the critical temperature
for equations with 3-point contact ({3 =0.5) and az remains equal to O. But for
equations with 5-point contact ({3 = 0.25) dc v diverges with a2 = 0.5.
The relations derived above have been known for decades (see Baehr, 1963)
and when both the theoretical understanding of critical effects and the experimental basis became more evident, it was clear that the critical exponents which result
from analytic equations of state do not match the theoretical and experimental

138

4 Setting up Multiparameter Equations of State for Pure Substances

results~3 This was the starting point for the development of equations of state

which specialise in the description of properties in the critical region.

4.5.1.1 Theoretically Founded Equations of State

To represent thermodynamic properties in the critical region, one needs a functional form for an equation which yields the required nonanalytical behaviour of
the considered properties. For this purpose, the Ising model is not very helpful,
because an explicit equation of state, in this case an explicit scaling function (see
Sengers and Levelt Sengers, 1986), has never been formulated for this model. In
practice, empirical closed-form expressions are used therefore as scaled equations
of state which conform with the asymptotic behaviour and the symmetry of the
Ising model. These formulations require transformations to parametric variables;
common transformations use the parametric variables () and r and are defined
implicitly by
h ='it - 'ito (f)

= a r{J~ ()(1_ ()2 )

(4.135)
(4.136)

with

T
T- =_-..
T'

T
T'

!J..T = I--..

'ito the analytical back~round of the chemical potential, a and c as systemdependent parameters, b as a "universal parameter'; and the critical exponents f3
and o. Based on this implicit transformation, the so called "simple scaling" approach leads to a functional form for the equation of state where just two functions
mo()) and ml()) still need to be determined. The simplest approach for these unknown functions is the "linear model';
(4.137)
13

The recent experimental results of Wagner et al. (1992), Kurzeja et al. (1999/2000), and
Kurzeja and Wagner (2000) for the "thermal" critical exponents p, y, and il, derived from ppT
measurements in the immediate vicinity of the critical point show clear differences from the
values predicted by the renormalisation group theory. These surprising results might have been
caused by an extended validity range of the so-called explicit influence of gravity (the implicit
influence of gravity, e.g., the averaging errors based on density stratifications, was taken into
account when evaluating the experimental ppT data); a theoretical solution of the three dimensional Ising model under influence of an outer field like gravity is still pending.
The region where these results were found is by about one order of magnitude smaller than the
region where state-of-the-art reference equations fail with respect to the representation of certain caloric properties (see Sect. 4.5.2) and it is by at least two orders of magnitude smaller
than the region where simple equations of state fail. Thus, these results do not disprove the
thesis that up to now empirical multiparameter equations of state have certain shortcomings
with regard to the representation of caloric properties in the critical region. However, they are
definitely worth considering when discussing further work on properties in the critical region.

4.5 Describing Properties in the Critical Region

139

which was proposed by Schofield et aI. (1969). This very early scaled equation of
state was used by Angus et al. (1976) for their compilation on thermodynamic
properties of carbon dioxide. However, simple scaling approaches assume a high
degree of symmetry of the critical region and this assumption holds only in a very
narrow region around the critical point.
In the following years, "revised and extended"scaled equations of state were
introduced to overcome these limitations, see for instance Albright et aI. (1987a)
for carbon dioxide. These equations have been revised in the sense that they consider the asymmetry of the phase boundary by using a twisted system of coordinates and extended in the sense that they consider first order Wegner corrections (see Wegner, 1972) to the simple power laws and the corresponding scaled
equations. Based on comparisons with experimental data, Sengers and Levelt Sengers (1986) estimate that the range where typical revised and extended scaled
equations of state are valid is 0.9995 ~ TITc ~ 1.03 and 0.75 ~plpc ~ 1.25; more
recent comparisons with highly accurate experimental data support these results
(see for instance Tegeler et aI., 1997/1999).
From an application oriented point of view, the major disadvantages of revised
and extended scaled equations of state are their complex structure with recursively
defined parametric variables, their small range of validity in combination with a
comparatively large number of fitted parameters (usually about 13 system dependent parameters plus a considerable number of "universal" parameters), and the fact
that a direct combination with an accurate wide range equation of state seems to be
impossible - the analytic background contributions were always calculated from
Taylor expansions in terms of critical point parameters and not from common
equations of state. The last two disadvantages were addressed by Fox (1983) who
introduced a different scaling strategy. Fox redefined the variables of an equation
of state by introducing an empirical "damping function" which results in a scaling
of thermodynamic variables when approaching the critical point; Fig. 4.15 illustrates the basic idea of such an approach. When applied to real systems the resulting relations are as complex and implicit as for the equations described above, but
Fox could apply his approach to a Van der Waals like equation of state in the very
first step, and outside of the critical region, the model approached the classical
limits of this simple equation - a major advantage compared to revised and extended equations of state.
Two different lines of development are based on the idea of Fox: The
"transformation" approach which has to be considered as semiempirical and
which is described in Sect. 4.5.5.2 and the "crossover" approach which forms the
basis of most of the recent scaled equations of state.
Albright et aI. (1986) introduced a theoretically more satisfactory but also more
complex "crossover" mechanism to rescale the variables of an analytic background function. In their first crossover article, the group around J. V. Sengers
used a Van der Waals equation to calculate the analytic background. However, the
development later switched back to Taylor expansions at the critical point (see for
instance Albright et aI., 1987b; Chen et al. 1990a/b). The equations published by
Chen et aI. (1990b) for carbon dioxide, water and ethane are valid in regions
which correspond roughly to 0.96 ~ T I Tc ~ 1.23 and 0.4 ~ P I Pc ~1.6; this is a

140

4 Setting up Multiparameter Equations of State for Pure Substances

dy
dx

df(x') dx'
dx'
dx

-=~.-~

JrW
dx

dy

df(x)

dx=~~

f'(x)

Ax'=a
1--""---I
Ax=a
=

f(x')

:
I

:
I

a simple analytic function

(f(x)) can be scaled (f'(x))
by using it with a scaled
variable (f(x'))
~

x'=g(x, xc)
Fig. 4.15. The basic idea of transformation and crossover approaches is to scale an analytic
function by using it with scaled variables.

major advantage compared to revised and extended scaled equations. However,

with 18 system-dependent fitted parameters, such formulations are even more
complex. Outside the region where revised and extended scaled equations are
valid, the achieved accuracy is unsatisfactory as compared to recent reference
equations of state (see Span and Wagner, 1996).
During the 90's a broad variety of crossover equations was published by different groups both for pure fluids and mixtures - see for instance Kiselev et al.
(1991) for water and carbon dioxide, Jin et al. (1992) for methane, LuettmerStrathmann et al. (1992), who presented the first parametric crossover model for
ethane, Kiselev and Sengers (1993) for methane and ethane, Povodyrev et al.
(1996) for mixtures of methane and ethane, Kiselev (1998) for a crossover model
using a cubic equation of state as background function, and Kiselev et al. (1998)
and Edison et al. (1998) for a more general crossover approach which describes
mixtures. To discuss all these models would be inappropriate at this point. For the
context of this section, three general statements are sufficient:
The development of scaled correlations for thermophysical properties of pure
fluids and mixtures is a lively area of scientific research. Recent work has resulted in a number of interesting models and extended theoretical knowledge
on thermodynamic properties of pure fluids to mixtures as well as to transport
properties.
The complexity of the models still hinders practical application. A crucial point
is the unavoidable introduction of scaled variables which require implicit
definitions and therefore iterative procedures to switch between thermodynamic variables and scaled variables. In combination with the required numerical steps, these implicit dependencies make results unreproduceable and the
use of analytic derivatives for the calculation of derived properties becomes

4.5 Describing Properties in the Critical Region

141

extremely complicated if not impossible. A significant simplification of scaled

models seems to be impossible unless physical exactness is sacrificed (see for
instance Sect. 4.5.5.3).
The formal range of validity of scaled models for pure substances may cover
the whole fluid region (see for instance Kiselev, 1998), but outside of the critical region, such complex models compare well only to very simple equations
of state. The region where their accuracy is comparable to the one of empirical
reference equations of state is still very limited. Improvements seem to be possible in combination with more complex background functions but adapting the
tools which are commonly used to establish accurate equations of state to the
development of background functions is difficult.
Based on this assessment of the practical capabilities of theoretically founded
equations of state it becomes interesting to focus on the practical capabilities of
empirical equations of state again.
4.5.2 Capabilities of Empirical Multiparameter Equations of State

From the theoretical results presented in Sect. 4.5.1, it is usually concluded that
analytic equations of state cannot represent the properties of pure fluids within the
critical region since they do not yield correct critical exponents. However, this
conclusion is incorrect for most thermodynamic properties if state-of-the-art reference equations of state are considered. This section is intended to illustrate both
the capabilities and the limitations of empirical multiparameter equations of state
regarding the representation of properties in the extended critical region.
Figure 4.16 shows percentage deviations between experimental data for the vapour pressure and the saturated-liquid and -vapour density of nitrogen and values
calculated from the recent reference equation of Span et al. (1998b). The data were
measured by Nowak et al. (1997b) and cover the temperature range
0.998::; TITc::; 0.9998. Even close to the critical point, the reference equation is
able to represent the most accurate available data within their experimental uncertainty; no reliable experimental data are available closer to the critical point. The
nitrogen equation of Span et al. uses modified Gaussian bell shaped terms (see
Sects. 3.1.3.2 and 4.5.3) but no nonanalytical terms (see Sects. 3.1.3.2 and 4.5.4)
and is thus analytic in the sense discussed in the preceding section. The equation
was constrained to a three point contact at the selected critical point which results
in a critical exponent of fJ = 0.5.
For carbon dioxide, ppT data measured by Straub (1972) are available directly
on the critical isotherm. Measured with an optical method, the total uncertainty of
these data is relatively large but the consistency, which is important to assess the
limiting behaviour of equations of state, is impressive. In Fig. 4.17, these data are
compared with data which were calculated from the recent reference equation for
carbon dioxide, the equation of Span and Wagner (1996); the critical parameters
used by Straub were readjusted to coincide with the critical parameters the equation was constrained to. Data which were calculated from other equations of state
are represented by dashed and dash-dotted lines.

142

4 Setting up Multiparameter Equations of State for Pure Substances

fO:1
-0.05

0.:

"Q.
<I

0
0

-0.5

"~
"Q..

:1

<I

0
0

-0.5
125.9

:<Xp~L!

:f

~ol

1 ! ILl
126.1

126.0

jol

126.2

Temperature T IK

Fig. 4.16. Percentage deviations 100 dy I y = 100 (yexp- Ycale) I Yexp between data for the vapour
pressure and the saturated-liquid and -vapour density of nitrogen measured by Nowak et al.
(1997b) and corresponding values calculated from the reference equation of Span et al.
(1998b).

304.1282 K

0.005

..... ..

<I
0
0

.:

.. _... -

-0.005
440

~I""\~

460

Pc

480

500

Density p I (kg m -3)

CH4-type equation
02-type equation

- - -

Chen et al.(1990b)

Fig. 4.17. Percentage deviations 100.1.plp= 100 (Pexp-Pcalc) Ipexp between ppT data on the
critical isotherm of carbon dioxide measured by Straub (1972) and values calculated from the
reference equation of Span and Wagner (1996). For comparison, values calculated from other
equations of state are plotted as dashed and dash-dotted lines.

4.5 Describing Properties in the Critical Region

143

The equation of Chen et al. (1990b) is a highly complex nonanalytical crossover

equation of state which describes the extended critical region of carbon dioxide.
The equation which is presented as a C~-type equation uses the functional form
with four modified Gaussian bell-shaped terms developed by Setzmann and Wagner (1991) for methane and was fitted to the same data set as the equation of Span
and Wagner, as was the 02-type equation which uses the functional form developed by Schmidt and Wagner (1985) for oxygen which contains only simple polynomial and exponential terms. An equation with this functional form was published by Ely (1986) for carbon dioxide. However, since this equation was based
on different data sets, it would have been unfair to use it in this comparison. All

150.7 K (1.0001 T,)

0.05 ,..----------'------"'---.,-----,
expo uncertainty

o
-0.05

"----'_---'-_--'-_-'---'-...L..-_.l...----'_---'-_-'

153 K (1.015 Tc)

0.05 ,...--------'---""'-----;-----,
' \ Tiesinga et al. (1994)

o
-0.05
0.05

""'----'_---'-_--'-_-'---'-...L..-_.l...----'_---'-_--'

160 K (1.06 Tc)

o_
-0.05 -

~/:/;i;'T<-i~:' ~
'- - '

"

--

~I

180 K (1.19 T,)

0.05 , . . . - - - - - - - - ' - - - - " ' - - - , : - - - - ,
I

/
./

-0.05 L...---"-_--'-_"'-----L__I....J-_""'----'_--'----'
300
400
500 Pc
600
700

Density p / (kg m-3)

Fig. 4.18. Percentage deviations between ppT data in the extended critical region of argon measured by Gilgen et al. (1994) and values calculated from the reference equation of Tegeler et al.
(1997/1999). For comparison, values calculated from the revised and extended scaled equation
of state by Tiesinga et al. (1994) are plotted as dashed lines.

144

4 Setting up Multiparameter Equations of State for Pure Substances

equations were constrained to the same critical point parameters and none of them
were fitted to the data of Straub - these data were used only as a consistency test.
Due to the chosen three point contact, the analytic equations of state 14 result in a
critical exponent of 15 = 3 and should result in a plot of the critical isotherm which
is too steep when approaching the critical point. In the extreme resolution chosen
for Fig. 4.17, this behaviour becomes obvious for the 02-type equation, but both
the CH4-type equation and the equation of Span and Wagner predict the plot of the
critical isotherm as well as the crossover equation.
As a third example for the representation of thermal properties in the critical
region, Fig. 4.18 shows deviations between highly accurate ppT data for argon
which were measured by Gilgen et ai. (1994) and values calculated from the recent
reference equation of Tegeler et ai. (1997/1999). For comparison, the dashed line
represents results of the revised and extended scaled equation of state by Tiesinga
et ai. (1994). The shown data cover both sub- and overcritical densities throughout
the extended critical region. From the fact that the analytic equation of Tegeler et
aI., which uses modified Gaussian bell-shaped terms but no nonanalytical terms,
yields a critical exponent of 15 = 3, one would expect that it cannot represent the
changing curvature of the isotherms around the critical density. However, the analytic equation represents the best available data within their experimental uncertainty. Even very close to the critical temperature, the scaled equation cannot represent the data appropriately and its uncertainty increases with increasing distance
from the critical point.
The surprising capabilities of state-of-the-art analytic equations of state become
intelligible when analysing the power laws for thermal properties, see Table 4.10.
The power laws which describe orthobaric densities and the pressure on the critical
isotherm are formulated in terms of vanishing distances. The values of the critical
exponents f3 and <'l describe the way in which the corresponding difference vanishes, but since these exponents are defined at the critical point, their influence on
the representation of the data is very small - in the immediate vicinity of the critical point, the remaining differences are small compared to the values of the corresponding properties.
To assess the capabilities of an equation of state in the critical region, it is advantageous to adapt the concept of "local exponents" which was introduced by
Wagner et ai. (1992) for the analysis of their experimental data to equations of
state. The local exponents which correspond to the critical exponents f3 and 15 can
be calculated according to
(4.138)

14

Although the equation of Span and Wagner (1996) contains three nonanalytical tenns, it
behaves basically like an analytic equation with respect to thennal properties, see Sect. 4.5.4.

4.5 Describing Properties in the Critical Region

5
o

0 . 5 0 . . - - - - - - - - - - - - - - - - - - - - -......1,.-,
_ _ Span et aI. (1998b)
extrapolation to Tc~. <.':1
- - - Jacobsen et aI. (1986)
.,'
0.45 --_. Jacobsen and Stewart (1973)
._.~'-'

~
~

'Cd

..5u

145

0.40

.~.---

----------

.-.--

_~-.-.-.-.-"

__ --------

-,-

_,,'"

",,'

___ - - - - - - - - 0.35}--_ _
_ _ _ _ _ _ _ _ _ _- - -

Tc

0.30+----+-----+---+----+-----+-------+-l
125.9
126.0
126.1
126.2

Temperature TIK
Fig. 4.19. Plots of the local exponent P, (see Eq. 4.138) as calculated from three multiparameter
equations of state for nitrogen. The plotted temperature range corresponds to 0.998 :0:; T I Te :0:; 1.
The necessary Maxwell iteration fails for the older equations close to Te.

and

(4.139)

respectively. For T ~ Tc or p ~ Pc the local exponents become equivalent to the

critical exponents.
Using nitrogen as an example, Fig. 4.19 shows plots of the local exponent PI
which result from three generations of empirical equations of state. The functional
form of the equation of Jacobsen and Stewart (1973) became known as the
MBWR equation (see Sect. 3.1.2); this equation is formulated in terms of the compression factor and consists of polynomial and exponential terms, where the exponential terms are restricted to exp(---o2). The equation of Jacobsen et al. (1986a) is
formulated in terms of the reduced Helmholtz energy and uses an optimised functional form which consists of polynomial and exponential terms; with regard to its
capabilities in the critical region, this equation is comparable to the 02-type equation discussed above. And finally, the recent equation of Span et al. (1998b) which
is an optimised equation in terms of the reduced Helmholtz energy consisting of
polynomial, exponential, and Gaussian bell shaped terms; with regard to its capabilities in the critical region, this equation is comparable to the Cf4-type equation
discussed above. It can be seen from Fig. 4.19 that all the equations yield the expected critical exponent P= 0.5, but their behaviour throughout the critical region
is very different. State-of-the-art reference equations of state yield local exponents
which are on the order of values which can be derived from experimental data up
to a region very close to the critical point. The shift to the classical limit occurs so
close to the critical temperature that it does not affect the representation of satu-

146

4 Setting up Multiparameter Equations of State for Pure Substances

s.o
"0-

- - Span et al. (1998b)

Jacobsen et al. (1986)
-. _. Jacobsen and Stewart (1973)

4.5

......
s:::
<!)
s::: 4.0

--------

'-'-'-.

3.5

u
0

--...

........

...' .......

...............

...:l 3.0

'

"'

......

"

..... -.:::.....

2.5
7

10

11

12

Density p/(kmol

13

14

15

16

m-3)

Fig. 4.20. Plots of the local exponent 151 (see Eq. 4.139) as calculated from three multiparameter
equations of state for nitrogen. The plotted density range corresponds to 0.63 $ P / pc $ 1.43.

rated liquid- and vapour-densities significantly~5 The older equations depart from
the experimentally found values for PI too quickly and are therefore not able to
accurately describe orthobaric densities in the critical region~6
Similar results are shown for 01 in Fig. 4.20. Again, the region where the local
exponent deviates from typical experimentally found values is small for the equation of Span et ai. (1998b) and is restricted to a region where IP-Pcl is already very
small - the equation is able to represent even the most accurate ppT data on the
critical isotherm (Nowak. et aI., 1997a/b) within their experimental uncertainty.
For the older equations, the region where 01 approaches the classical limit is
broader and the description of ppT data in the critical region is affected as theoretically predicted.
For the isochoric compressibility along the critical isochore, the situation is the
same. Recent equations of state with modified Gaussian bell shaped terms are
obviously able to describe experimentally found local exponents YI even better
than scaled equations of state which are constrained to the theoretically predicted
value ofy by their functional form, see Fig. 4.18.
State-of-the-art analytic multiparameter equations of state describe thermal
properties in the critical and extended critical region at least as well as scaled
equations of state. This capability is not affected by the classical limiting behaviour of the equations.
For caloric properties the situation becomes more complex. When calculated
from accurate equations of state, properties which describe caloric behaviour for
In fact the observed shift agrees well with recent results found for fluids under the influence of
gravity (see footnote 13). This may be one reason why state-of-the-art multiparameter equations usually describe orthobaric densities in the critical region better than scaled equations.
16 When calculating PI from the older equations the maxwell iteration fails close to the critical
point; this is not a characteristic feature of the used functional forms but depends on the data
and techniques which were used to fit these equations.
15

4.5 Describing Properties in the Critical Region

147

final changes of state such as enthalpy differences !1h or differences of the internal
energy !1u are usually hardly influenced by the different description of critical
phenomena. The region where the limiting behaviour of recent equations of state
becomes relevant is so small that the influence of the effects which will be discussed below simply averages out for typical values of !1T, !1p and !1T, !:.p, respectively. The properties which have to be examined in more detail are the speed of
sound and derived properties such as the heat capacities.
The specific isobaric heat capacity can be expressed as

-(::r
,

~
R

= _r2

(a~,

+ a~)

5c
R

(1+ oab - orab,)

1+2oa
,
0 +o2a 00,
f

-(;p)
ap

(4.140)

where the derivatives of the reduced Helmholtz energy a are abbreviated according to Eq. 3.57. Since (dp/dp)~1 grows much faster when approaching the critical
point than cv , the specific isobaric heat capacity is dominated by the fraction,
which is closely related to the representation of ppT data. Thus, an equation which
yields an accurate description of the ppT surface within the critical region should
also yield reliable values of the specific isobaric heat capacity.
The situation is different for the isochoric heat capacity which is given by
(4.141)
If the second derivative of the residual part of the Helmholtz energy with respect
to r is finite at the critical point, the value of the specific isochoric heat capacity is
also finite!7 An equation which is analytic fails with regard to the representation of
the isochoric heat capacity at the critical point. At the same time, analytic formulations will result in finite values of the speed of sound at the critical point since the
speed of sound corresponds to

RT

17

= ~ +2oab +o2abO

-[i~1

(4.142)

One may argue, that this behaviour agrees well with the results found for fluids under the
influence of gravity (see footnote 13). However, the region where even (analytic) state-of-theart reference equations of state fail with respect to the representation of C v is much larger than
the region where classical critical exponents were experimentally found.

148

4 Setting up Multiparameter Equations of State for Pure Substances

P=Pc

1
.,

- - - Chen et al. (1990b)

- - SpanandWagner(1996
- -. - CH4-type equation

I
I

- - - -. 02-type equation

o
o

~
~
'-'

.......

Edwards (1984)
Abdulagatovet al. (1994)

U
2

T----1
300

302

304

306

308

310

Temperature T / K
Fig. 4.21. Experimental results for the isochoric heat capacity on the critical isochore of carbon
dioxide and values calculated from four equations of state.

At the critical point, the expression cap / iJp)r becomes zero and (iJp / iJnp is a finite
value. Thus, only if Cv becomes infinite, the speed of sound becomes zero.
Figure 4.21 illustrates the limitations of analytic multiparameter equations of
state using the representation of experimental results for the isochoric heat capacity on the critical isochore of carbon dioxide as an example. When approaching the
critical temperature from the two phase region, all equations yield qualitatively
correct behaviour up to temperatures which are very close to Te. In this region the
dominant contribution to the residual isochoric heat capacity results from L1c~ and
L1c~ which are described satisfactorily by rather simple analytic equations of state.
When approaching the critical temperature from the single phase region the O2type equation fails to represent the increasing isochoric heat capacity for temperatures below about 1.01 Te (~307 K). Equations which use modified Gaussian bell
shaped terms like the C~-type equation are able to follow the increasing isochoric
heat capacity down to about 1.002 Te (:==304.7 K) but as analytic equations of state
they cannot result in an infinite isochoric heat capacity at the critical point. Just
like the nonanalytical crossover equation by Chen et al. (1990b), equations which
use nonanalytical terms (see Sects. 3.1.3.2 and 4.5.4) like the recent reference
equation by Span and Wagner (1996) are able to follow the experimentally found
strong curvature and result in an isochoric heat capacity which becomes infinite at
the critical point.
The corresponding results for the speed of sound are shown in Fig. 4.22. Just
0.5 K above the critical temperature the C~-type equation, the equation of Span
and Wagner (1996) and the crossover equation by Chen et al. (1990b) still yield
very similar results. But when further approaching the critical temperature, the

4.5 Describing Properties in the Critical Region

149

200r---------------------------~

175

'i

'"
E

--

'-'

~
"0

150

.....'"0

125

"0
II)
II)

r:l.

CI:l

100

400

Pc

500

600

Density pi (kg m-3)

Fig. 4.22. Values for speed of sound in the critical region of carbon dioxide as calculated from
three different equations of state.

speeds of sound calculated from the analytic Cf4-type equation remain finite,
while the crossover equation and the nonanalytical multiparameter equation result
in the theoretically expected vanishing speed of sound at the critical point.
However, for reasons which will be explained in more detail in Sect. 4.5.4,
equations which use nonanalytical terms of the form introduced by Span (1993)
and Span and Wagner (1996) are still not able to fulfil the theoretical demands on
the asymptotic limiting behaviour of the isochoric heat capacity. Figure 4.23 shows
the plot of the isochoric heat capacities calculated from the 02-type equation, from
the equation of Span and Wagner (1996) and from the crossover equation by Chen
et al. (1990b) in a double logarithmic diagram. Only the crossover equation yields
the theoretically expected linear extrapolation of the experimentally found isochoric heat capacities for temperatures asymptotically close to Te.
Summarising the results regarding the representation of properties in the critical
region by multiparameter equations of state, one has to conclude as follows:
Regarding the representation of thermal properties recent reference equations
of state with modified Gaussian bell shaped terms are at least as good as special scaled equations of state. Properly designed equations which use only
polynomial and exponential terms are sufficient for most technical applications
in the critical region, but with respect to data needs for scientific applications,
they tend to yield unsatisfactory results.

150

4 Setting up Multiparameter Equations of State for Pure Substances

20r-~------------------------~

10

0. 10- 6

10-5

10-3

10-4

IT- Tel I Te

11111111

0.001

11111111

0.01

11111111

11111111

0.1

10-2

10- 1

11111111

10

IT- Tel IK
Span and Wagner (1996)
Chen et al. (1990b)
02-type equation

0
0

Edwards (1984)
Abdulagatov et al. (1994)

Fig. 4.23. Double logarithmic representation of experimental results for the isochoric heat capacity on the critical isochore of carbon dioxide and of values calculated from three equations of
state.

Recent reference equations of state with nonanalytical terms are able to represent caloric properties in the critical region within the uncertainty of the available data. Equations with modified Gaussian bell shaped terms fail to describe
isochoric heat capacities and speeds of sound in the critical region for about
1 ~ T I Tc ~ 1.002 but they are still sufficient for all technical and scientific data
needs. Equations which use only polynomial and exponential terms fail in the
critical region for about 0.998 ~ TITc ~ 1.01 and tend to yield unsatisfactory
results for advanced data needs.
Empirical multiparameter equations of state do not fulfil theoretical demands
on the limiting behaviour of thermodynamic properties for the asymptotical
approximation to the critical point. No theoretical conclusions should be drawn
from the limiting behaviour of empirical equations of state. 18
Based on this assessment, the use of nonanalytical terms has been avoided for the
most recent reference equations of state (see Tegeler et aI., 1997 I 1999; Span et
aI., 1998b; Smukala et aI., 1999). The numerical expense which is caused by the
use of nonanalytical terms for both the user and for those who develop the equation, seems to be justified only for substances where properties in the critical region are of special scientific interest.

18

This is in fact almost self-evident; theoretical conclusions which are drawn from empirical
correlations are almost always questionable.

4.5 Describing Properties in the Critical Region

151

4.5.3 Setting Up Equations with Modified Gaussian Bell Shaped

Terms
The modified Gaussian bell shaped terms introduced by Setzmann and Wagner
(1991) in the form
(4.143)
have 6 internal adjustable parameters besides the leading coefficient ni which can
be determined directly in linear or nonlinear fits. The inverse reduced temperature
and the reduced density 0 have to be calculated with the corresponding critical
parameters when using modified Gaussian bell shaped terms. The parameter Bi has
always been chosen to be equal to 1 since any other value would result in an undesired asymmetry of the contribution of the term. The parameter Yi shifts the point
where the derivatives of the term become maximal (p I Pc = B i and TI Tc = Yil) into
the unstable part of the two phase region and has to be larger than 1. The exponents ti and d i are typical temperature and density exponents which have to be
determined by optimisation algorithms just like those for the polynomial and exponential terms.
Setzmann and Wagner (1991) determined reasonable values for TJi' Pi, and Yi by
nonlinear fits, but the nonlinear relations between these parameters and the residua
of the thermodynamic properties resulted in numerically unstable fits. The parameters had to be fitted one by one together with the ni of all other terms in the equation and the whole procedure became both very time-consuming and unsatisfactory. Thus, the most powerful combination of these parameters was determined by
using the optimisation procedure, after a range of reasonable values had been established by nonlinear fits. The extensive derivatives of Ai and its derivatives with
respect to TJi' Pi, and Yi, which are required for nonlinear fits are not needed for the
optimisation approach and are therefore not given here. For equations of state
without additional nonanalytical terms the following range of parameters turned
'
out to be useful: 19
1 ~ di ~ 3

o ~ ti ~ 3
15

TJi ~ 25

200 ~Pi ~ 375

1.11

~ Yi ~

1.25

(4.144)

In combination with nonanalytical terms Wagner and ProB (1997/2000) used

smaller values for TJi (10 ~ TJi~ 20) and for Pi (150 ~Pi~ 250) while Span and
Wagner (1996) used parameter ranges which correspond roughly to the limits
given in Eq. 4.144.
19

Setzmann and Wagner (1991) additionally used values up to 40 for 1,1i, 1.07 for Yi, and 0 for di ;
the recommendations given here are based on later experiences.

152

4 Setting up Multiparameter Equations of State for Pure Substances

While working with modified Gaussian bell shaped terms, it is important to

realise that correlated pairs of these terms are especially effective. Since the capabilities of the original Setzmann and Wagner (1989a) algorithm are restricted
when optimising such intercorrelated pairs of terms, it is advantageous to use it in
its modified form (see Sect. 4.4.5.2) testing double exchanges of the modified
Gaussian bell shaped terms in the equation against those in the bank of terms.
4.5.4 Setting Up Equations with Nonanalytical Terms

The nonanalytical terms which were already discussed above and in Sect. 3.1.3.2
are essentially based on a development by Span (1993) and were used in a modified form by Span and Wagner (1996) and Wagner and ProB (1997/2000). To
establish equations of state which use such nonanalytical terms, it is important to
understand the basic idea behind these terms. In order to cause the isochoric heat
capacity to become infinite and the speed of sound to vanish at the critical point
(see footnote 17), in an equation of state it is necessary to introduce nonanalytical
terms which yield infinite values for the second derivative of a with respect to T at
the critical point. However, reasonable nonanalytical terms have to fulfil three
additional demands:
The values resulting for a~ have to be finite everywhere except at the critical
point.
Singular behaviour of the other second derivatives and all derivatives with
respect to 0 has to be avoided everywhere.
Within the O,T surface of the critical region, the maximum of a~ has to follow
the course of the saturated vapour- and saturated liquid line in order to avoid
unreasonable maxima of the isochoric heat capacity in the single-phase region.
These demands led to the functional form
(4.145)

The nonanalytical behaviour is introduced by the non-integer exponent hi of the

distance function A in combination with the set-up of A itself. The reduced density
in Eq. 4.145 had to be introduced in order to guarantee a completely vanishing
contribution when approaching the ideal gas limit and the exponential term damps
the influence of the nonanalytical terms outside of the critical region. The derivatives ofEq. 4.145 with respect to T and 0 have been summarised in Table 3.8.
When approaching the critical point, the asymptotically leading terms in
(dA;litr)<1 become

4.5 Describing Properties in the Critical Region

and

lim

,=I.Ci.... 1

(aAi)
__ CI.(O_I)2aibi-2ai+;i_C2.(O_I):,cbi-O.5)_ .....
ar Ci

153

(4.146)

where the contribution from the exponential function is considered analogous to

Eq. 4.88 and Co, CJ, and C2 are arbitrary constants which have no influence on the
limiting behaviour. Since the first temperature derivative of has to be finite and
continuous at the critical point, the conditions bi > 0.5 and ai < (2/Hb i-l)t i can
be derived from Eq. 4.146. The second derivative (d 2AJi.tr2)c) is expected to diverge at the critical point. For a course along the critical isochore, the asymptotically leading term becomes

Ai

lim

u=l, , .... 1

(aar2Ai)
2 _C . (1- r )2bi-2 + ..... .

(4.147)

Ci

When comparing Eq. 4.147 with the power law for the isochoric heat capacity (see
Eq. 3.14 for the calculation of C v and Table 4.10 for the power law), the relation
a =2 - 2 b i between b i and the critical exponent a becomes obvious - the asymptotic behaviour of the isochoric heat capacity is determined by the nonanalytical
term with the smallest value of bi. Theoretically, a value of bi = 0.945 would be
expected from a = 0.110. The coefficient ni of the leading nonanalytical term has
to be negative to result in ciTe,Pe) = +00. With the leading term

the condition that (d 2AJ dO 2), has to be finite and continuous everywhere yields the
relations Pi2ai(l-bi)+2ri and bi >0.5+Pi' No further relevant conditions
were found when investigating the limiting behaviour of (d 2AJi.trdO),
(d 3AJi.trd0 2 ), and (d 3AJdO\; (d 3AJi.tr3)c) and (d 3AJi.tr2dO) are expected to become infinite at the critical point. The definition of the distance function A and the
fact that () is used only with positive exponents in the derivatives of (see Table
3.8) guarantees that none of the derivatives is discontinuous anywhere but at the
critical point. Thus, if the conditions for ai, bi, and Pi are fulfilled, Eq. 4.145 fulfils
the first two demands formulated above.
The contour of the curve where the contribution of a nonanalytical term to
(d 2a / i.tr 2)c) becomes a maximum is determined by the form of the distance function
A. The relation

Ai

(l-r)+ci [(O-I)2]2/3i =0

(4.149)

can be interpreted as single sided equivalent of the power law describing the orthobaric densities, see Table 4.10. Using a transformed form ofEq. 4.149,
(4.150)

154

4 Setting up Multiparameter Equations of State for Pure Substances

-(~~ ),1

Fig. 4.24. Qualitative illustration of the contribution of a nonanalytical term to the derivatives
(a 2a / Or2)~ and (a 2a / Or ao) in the critical region.

Ci and fJi can be determined by a nonlinear fit to data for both the saturated-vapour
and -liquid density in the immediate vicinity of the critical point. However, since
the condition for hi which was derived above is very sensitive to large values of fJi
the value for fJi has to be chosen slightly smaller (j3i'" 0.30) than ex~ected from the
power law and only Ci can be fitted to orthobaric densities directl/ The necessary
data can be calculated from the corresponding ancillary equations.
The contribution of a nonanalytical term to (iia / (}r2)o and to (a 2a / (}r ao) is illustrated in Fig. 4.24. For (a 2a / (}r2)o the sharp maximum which corresponds to the
desired increase of the isochoric heat capacity at the critical point becomes visible
while the contributions to (a 2a / (}r 00) and to the other first and second derivatives
stay very small. Terms like this can be used as an integral part of empirical multiparameter equations of state to obtain a nonanalytical behaviour with respect to
(a 2a / (}r2)o.

The parameters ai, hi, di, ei, and /; can be determined from a nonlinear fit theoretically. Reasonable values for hi, ei, and/; were determined by Span and Wagner
(1996) by nonlinear fits of preliminary equations. However, as with the Gaussian
bell shaped terms (see Sect. 4.5.3), this procedure turned out to be very timeconsuming and unsatisfactory since the parameters could only be fitted one at a
time due to numerical restrictions. The final parameter combination was determined by means of an optimisation algorithm by Span and Wagner (1996) and
20

One disadvantage of this kind of nonanalytical terms is that the contour defined by Eq. 4.149
can never match the phase boundary calculated from the equation of state exactly. To avoid
unreasonable maxima of Cv close to the phase boundary, Ci and {Ji have to be chosen in such a
way that the described contour lies within the two phase region, as close to the actual phase
boundary as possible. Scaled approaches which use the parametric variables defined by Eqs.
4.135-4.136 avoid this problem since the orthobaric densities correspond to () = 1 per definition. However, the advantage of this set-up of the parametric variables is questionable since
their definition restricts the flexibility of the equation; scaled formulations cannot describe
orthobaric densities in the critical region as acc urately as multiparameter equations of state.

4.5 Describing Properties in the Critical Region

155

Wagner and ProB (1997/2000). The nonanalytical terms used in the bank of tenns
of Span and Wagner (1996) covered the parameter ranges
3.0::;; ai::;; 4.0
0.875 ::;; hi ::;; 0.925
0.30::;; di ::;; 1.00
10::;; ei::;; 15
225 ::;;1; ::;; 275.

(4.151)

Slightly different combinations were used by Wagner and ProB (199712000). The
extensive derivatives of Ai and its derivatives with respect to ai, hi, d;, ei, and 1;,
which are required for nonlinear fits, are not needed for the optimisation approach
and are therefore not given here.
The introduction of nonanalytical terms has significantly improved the capabilities of empirical equations of state with regard to the representation of caloric
properties in the critical region, but such equations still have certain limits if an
exact fulfilment of the asymptotic power law for the isochoric heat capacity is
required as shown in Fig. 4.23. In multiparameter equations of state, the analytic
terms contribute decisively towards the description of critical effects. Nonanalytical tenns do not replace this contribution close to the critical point but they fill the
increasing gap between analytic and nonanalytical behaviour with regard to
Ciia/ifr2)d. Since the relative growth of this gap is much faster than the relative
growth of C v in the region where experimental data are available, efficient values of
the exponent hi are smaller than theoretically expected and result in a critical exponent a which is too large from an asymptotic point of view. Basically, this is an
unavoidable problem, but recent semiempiric approaches show that further improvements are possible (see Sect. 4.5.5.3).
A second problem is linked to the density derivatives of Eq. 4.145. When using
these kinds of nonanalytical tenns, it was assumed that continuous plots of derivatives which do not diverge are sufficient; possible unphysical contributions have to
be compensated by other terms in the multiparameter equation of state in this case.

-0.340....----------------....,
...
Tc +0.9K
: ' -0.338
~ -0.336
,
-0.334
--...
Co
:'---------'-' -0.332
'Pc
-0.330 '---'-_.L.....--'-_--'----'''-''--'-_'---'-_"'----I
467.55
467.60
467.65

-a

Density p/(kgm-3)
Fig. 4.25. Results for the third density derivative calculated from the equation of Span and
Wagner (1996) close to the critical isochore.

156

4 Setting up Multiparameter Equations of State for Pure Substances

5000

7'
i

2000

~
bI)

..:.d

1000

~
....... 500
u'"

Tc + 0.1 K

'-'

.~

200

u
~

100

Q)

...r::
u

.~

..c
0
til
......

50

20
10~~~~~~--U-~--~~--~~

350

400

450

500

550

600

Density p/(kgm-3)
Fig. 4.26. Close to the critical point the sudden changes in higher density derivatives caused by
nonanalytical tenns result in an unphysical behaviour of the isobaric heat capacity.

However, along the critical isochore, Eq. 4.145 results in changes of higher density
derivatives which are small and continuous but so rapid that even modified Gaussian bell shaped terms cannot compensate for this effect. Figure 4.25 shows such a
change for the third density derivative of the carbon dioxide equation by Span and
Wagner (1996). The observed decrease in (a 3alao 3). is too small to affect the
representation of thermal properties, but it results in an unphysical behaviour of
the isobaric heat capacity, see Fig. 4.26~\ This problem has to be addressed first to
further improve empirical nonanalytical terms - for future applications, Eq. 4.l45
should be reformulated in a way that its higher density derivatives become well
behaved along the critical isochore.
4.5.5 Semiempirical Approaches
Compared to theoretically based descriptions of the critical region, the major advantage of the nonanalytical terms described in the preceding section is that no
2\

The isobaric heat capacity is especially sensitive to such effects since the sum in the denominator of Eq. 4.140 is nearly zero close to the critical point and results in the extreme increase
of the isobaric heat capacity. Even very small oscillations in a:i or add result in visible changes
of Cpo

4.5 Describing Properties in the Critical Region

157

implicit transformations of variables become necessary. With the independent

thennodynamic variables (T and p) the equation and all its derivatives can be
evaluated directly without any iterative steps. For certain purely scientific applications where this feature is regarded as less important, semiempiric approaches can
be used in combination with multiparameter equations of state to achieve a
nonanalyticallimiting behaviour. The following sections summarise these methods
just briefly, since the related problems depend mainly on the theoretical approaches used and not on the applied multiparameter equations of state.
4.5.5.1 The Use of Switching Functions
The main idea of switching approaches is that two equations of state, a scaled
equation of state for the description of the critical region and an empirical multiparameter equation for the description of the whole fluid region, are linked by a
third function, the so called "switching" or "blending junction". This function
combines both equations in such a way that the results of the analytic multiparameter equation are dominant far away from the critical point while the results of the
nonanalytical scaled equation dominate close to the critical point.
The first well known application of switching functions was documented in the
IUPAC monograph on thennodynamic properties of carbon dioxide which was
published by Angus et al. (1976). Angus et al' combined the pressure explicit multiparameter equation of state by Altunin and Gadetski (1971) and the simple scaling equation by Schofield (1969) and Schofield et al. (1969) using the switching
function by Chapela and Rowlinson (1974). This approach reads

P = f(r)' Pan. + (1- f(r) Psc.

(4.152)

with the switching function

(4.153)
The independent parametric variable of the switching function, r, corresponds to a
distance to the critical point and is defined by Eqs. 4.135-4.136. nJ, n2, kJ, and k2
are adjustable parameters of the switching function. For small values of r the
switching function fir) becomes small (almost 0) and Psc. dominates in Eq. 4.152.
For large values of r the switching function fir) becomes equal to 1 and Pan. dominates.
This simple switching approach resulted in major problems with regard to derived properties such as heat capacities. To avoid unphysical results, caloric properties tabulated by Angus et al. were partly obtained from graphical interpolations.
In the switching region, where fir) changes from 0 to 1 or vice versa, the derivative
of Eq. 4.152 with respect to an arbitrary variable x (in practice p or T) becomes

(:~)y =f(r)C~:n.)y +(1- f(r){a:~c. )y + (Pan. - PscJ{:\

(4.154)

158

4 Setting up Multiparameter Equations of State for Pure Substances

where (i)jldx)y has to be calculated as (djldr)(drldx)y in this case. The first two
terms in Eq. 4.154 correspond to the set-up of Eq. 4.152 and are physically expected, but the last term which depends on differences between the equations and
on derivatives of the switching function causes significant unphysical contributions
to derived properties especially for higher derivatives.
Wooley (1983) showed that this problem is theoretically unavoidable, but that
its impact can be minimised by a suitable construction of the switching function.
However, with a total of 15 adjustable parameters in the distance function, the
switching function proposed by Wooley was far too complicated for practical
application.
The breakthrough in switching was achieved by Hill (1990) who combined an
analytic Helmholtz equation of state for water with a revised and extended scaled
equation of state. His switching approach reads
(4.155)
using the switching function
j =1-exp

-1

exp((~/Ql) )-1

(4.156)

with only three adjustable parameters Qj, Q2, and Q3. The success of this comparably simple switching approach is based on two facts: The analytic and the scaled
equation of state are harmonised in the sense that they yield very similar results in
the switching region - the expression (a.n. - a se ) becomes small when adapting
Eq. 4.154 to a Helmholtz approach. And the range where both equations yield
similar results is sufficiently broad due to the use of a revised and extended scaled
equation on the one side and a sophisticated multiparameter equation on the other
side - the switching range can be extended and the expression (iJjl dx)y in
Eq. 4.154 becomes small. The unphysical contribution in Eq. 4.154 cannot be
avoided, but it is for the most part negligible.
The features which make Hills approach successful raises criticism at the same
time. Since results calculated from the analytic equation have to be very similar to
results calculated from the scaled equation throughout the whole switching region,
the analytic equation has to be quite complex and advantages of the combined
model are restricted to a very small region around the critical point. In addition,
since the scaled equation has to have a rather broad range of validity, complex
formulations with a large number of adjustable parameters have to be used. Thus,
when using this kind of switching approach, one has to pay a high price for small
advantages.
4.5.5.2 The Transformation Approach

The transformation approach introduced by Erickson and Leland (1986) and Erickson et al. (1987) is based on the earlier work of Fox (1983), see Sect. 4.5.1.1.

4.5 Describing Properties in the Critical Region

159

Erickson and Leland replaced the "damping function" introduced by Fox by a

more complex formulation with four instead of two adjustable parameters. Using
this improved damping function they applied the concept proposed by Fox to
MBWR type pressure explicit multiparameter equations of state without refitting
the parameters of these equations. The reported results for methane show that the
analytic equation could be improved with respect to some typical critical region
effects, but major shortcomings become visible, too. Obviously a multiparameter
equation which describes critical effects as well as possible with analytic terms is
unsuitable as a background equation for a scaling approach.
Erickson et al. (1987) modified the damping function again and fitted its parameters simultaneously with the 32 coefficients of the MBWR type background
equation to data for pentane and carbon dioxide. The resulting nonanalytical
model avoids in part the shortcomings which became obvious in the preceding
report and is said to be superior to an equation which uses the form developed by
Schmidt and Wagner (1985) for oxygen (Oz-type equation, see Sect. 4.5.2) in the
critical region, while the Oz-type equation is superior outside of the critical region.
However, state-of-the-art multiparameter equations of state are far better than O2type equations even without nonanalytical terms (see Sect. 4.5.2) and the results of
the transformation approach are still not asymptotically correct. It does not seem as
if the limited advantages justify the numerical expense implied by the use of transformed variables.
The transformation approach has not been further pursued over the last few
years. When taking up this idea again today, one would certainly try to scale analytic multiparameter equations of state using recent crossover mechanisms, see
Sect. 4.5.1.1.
4.5.5.3 The Approach of Kiselev and Friend
Recently, Kiselev and Friend (1999) have published a semiempirical approach
which is similar to the use of nonanalytical terms described in Sect. 4.5.4. According to the crossover theory developed by Chen et al. (1990a), the renormalised
Landau expansion of the Helmholtz energy can be written as
a

M(r,~1]) =a12 f Y- 2~ ~1]2 Y

y-2P

y-2P

---u- +a04 ~1]4 Y --~- -K(p,f2)

(4.157)

where K is the "kernel term" which provides the expected scaling behaviour of the
isochoric heat capacity.
When investigating the recent reference equation for water (Wagner and PruB,
1997/2000) which reads

ar(r,o)

51

;=1

;=8

= 2: n;r t, Od, + 2:n;rt' Od, exp(-y;oP,)

+

2: n; rt, Od, exp( -1];(0 - EJ2 - f3;(r - yJ2)

54

;=52

160

4 Setting up Multiparameter Equations of State for Pure Substances

2: n O~bi exp(-e (o-I)2 - flr-l)2)

56

i=55

(4.158)

Kiselev and Friend found that this kind of multiparameter equation yields an excellent description of the thermodynamic properties in the critical region, but that the
nonanalytical terms (i = 55, 56) still do not result in the asymptotically correct
limiting behaviour for the heat capacities:2 Thus, they replaced the nonanalytical
terms in Eq. 4.158 by a modification of the kernel term in Eq. 4.157 to improve
the asymptotic behaviour of the equation. The modified equation reads
af('r,o)

51

i=1

i=8

=2: ni ili Odi + 2: niili odi exp(-Yi OPi )

+

2: n
54

i iIi odi

exp(-1Ji(o-ef -

fJi(i-YJ 2 )

i=52

(4.159)
with

T-Tc

i=--'

Tc

'

the terms 1-54 remain unchanged. The kernel term in Eq. 4.159 reads

(4.160)
where the crossover function Y is implicitly defined by
(4.161)
(4.162)

22

This conclusion is undenied, see Sects. 4.5.2 and 4.5.4 or Span and Wagner (1996).

4.6 Consideration of the Extrapolation Behaviour

and

R=(I+~)2
qo+q

161

(4.163)

a, /3, y, ~, b; and qo are "universal" constants of this approach and a20, a2J, a\2, Gi,
VI, and V2 are parameters which were adjusted to experimental data. To evaluate
Eq. 4.160 for a given temperature and density, the equations Eqs. 4.161-4.163
have to be solved simultaneously using an iterative solver. To calculate the derivatives required in thermodynamic property calculations from Eq. 4.160 analytically
becomes extremely complicated due to this implicit set-up. Thus, this approach
shows typical disadvantages of scaled equations, even though it is still rather simple compared to typical crossover equations~3 Furthermore, the general approach
of Kiselev and Friend, as applied to water, is questionable since both the nonanalytical terms in the original equation and the kernel term in the modified equation
contribute to all derivatives of a. Except for (a 2a / dr 2 )" and higher temperature
derivatives, these contributions may be small, but they are certainly not completely
negligible. Thus, such a kernel term should not be added to an existing equation
but it has to be fitted together with all other coefficients of the equation and, if
possible, it should be considered when optimising the functional form of the equation. These demands make the whole approach increasingly complex.
However, based on the results reported by Kiselev and Friend, the modified
equation seems to be superior to the equation with nonanalytical terms especially
with regard to the asymptotic behaviour of the isochoric heat capacity. Since the
basic concept of adding a nonanalytical contribution to an analytic multiparameter
equation is the same as for the comparably simple nonanalytical terms, this result
gives reason to believe that further improvements of empirical nonanalytical terms
are possible too, even though the principal restrictions discussed in Sect. 4.5.4
remain valid. Hopefully, such improved nonanalytical terms can preserve their
explicit set-up and their relative simplicity.

4.6 Consideration of the Extrapolation Behaviour

Over the years, considerable interest in thermodynamic properties of fluids at very
high pressures and temperatures has resulted mainly from applications in geology,
petrology and geophysics. Several simple equations of state have been developed
especially for these applications. Usually, these equations are valid only in restricted ranges of temperature and pressure and they fail to properly represent
accurate experimental data at lower temperatures and pressures. On the other hand,
empirical multiparameter equations of state have often failed with respect to extrapolation beyond the temperature and pressure range of the data to which the
equations were fitted. Instead of developing special equations of state for very high
23

A major advantage of the kernel term defined by Eqs. 4.160-4.163 is that it depends directly
on the thermodynamic variables T and p and not on some kind of parametric variables r and (),
as recent crossover equations usually do. In this way a second level of implicit dependencies is
avoided.

162

4 Setting up Multiparameter Equations of State for Pure Substances

temperatures and pressures, it would be desirable to improve the extrapolation

behaviour of these accurate equations of state in order to describe the entire range
of thermodynamic properties of a fluid as accurately as possible with a single
equation of state. With this in mind, the extrapolation behaviour of empirical
equations of state was one of the main topics of the Fifth International Workshop
on Equations of State which took place at the Ruhr-Universitiit Bochum in 1990.
De Reuck (1991) summarised the results of this discussion, which focused mainly
on the so called "ideal curves" (see Sect. 4.6.3) of pure fluids. Based on these
discussions and the report of de Reuck, Span and Wagner investigated the extrapolation behaviour of multiparameter equations of state further and published a comprehensive report on the current status of our knowledge of the extrapolation behaviour of empirical equations of state (Span and Wagner, 1997). Essentially, the
set-up of this section follows the set-up ofthe report by Span and Wagner.
It is well known that recent multiparameter equations of state are able to represent experimental data accurately up to very high temperatures and pressures; for
some examples see Chap. 5. However, for the extrapolation discussion it is more
important to point out that state-of-the-art equations of state are not flexible
enough to follow systematically wrong courses of single data sets in the highpressure region. During the optimisation of the mathematical form, strongly correlated pairs of terms are automatically replaced by single terms with similar contributions and steps for the optimisation of the length of the equation are used. Although recent reference equations contain usually 35 or more terms, only a few of
these terms contribute significantly to the behaviour in the high-pressure region
and the flexibility of the equations in this region is therefore very restricted. To
illustrate this concept, Fig. 4.27 shows a comparison between experimental data

Pressure plMPa
Shmonov and Shmulovich (1974)
-.. -.. - Sterner and Pitzer (1994)

III

'f!

Vukalovich and Altunin (1962)

Fig.4.27. Percentage deviations of selected ppT data at high temperatures and pressures from
values calculated from the equation of Span and Wagner (1996). Values calculated from the
equation of Sterner and Pitzer (1994) are plotted for comparison.

4.6 Consideration of the Extrapolation Behaviour

163

for carbon dioxide and results from the recent reference equation by Span and
Wagner (1996) which is used for the baseline, and from a special high-pressure
equation by Sterner and Pitzer (1994, see also Pitzer and Sterner 1995a/b). With
its 42 fitted coefficients, the empirical reference equation represents the accurate
data at pressures up to 60 MPa much better than the semiempiric equation of
Pitzer and Sterner with 28 fitted coefficients, but at pressures above 100 MPa both
equations yield similar results and do not follow the faulty course of the dati4
In contradiction to common teachings, empirical equations with a carefully optimised mathematical structure are not flexible enough to follow incorrect courses
of data sets in the high pressure and temperature region even if these data are the
only experimental information which is available in this region. Thus, it can be
concluded that they will also be stable enough to yield a reasonable extrapolation
behaviour in regions not covered by data. Qualitatively this statement agrees with
experiences made during the establishment of other equations of state (see Setzmann and Wagner, 1991; Span and Wagner, 1996; PruB and Wagner, 1995;
Tegeler et aI., 1997/1999; Span et aI., 1998b/ 1999; Panasiti et al., 1999; Lemmon
et aI., 1999; Wagner and PruB, 2000) but it has not been quantified up to now. It is
clear that the limits of the range in which an extrapolation is useful depend on the
considered property, on the demanded accuracy, on features of the data set used to
establish the equation and finally on features of the equation itself. Thus, a simple
answer cannot be expected; systematic studies on this topic are still lacking. However, more systematic information on the extrapolation behaviour of empirical
equations of state is available from the following three approaches.
4.6.1 Comparisons with Data Beyond the Range of Primary Data

At pressures and temperatures beyond the range covered by reliable experimental

pure fluid data, which are usually used to establish multiparameter equations of
state, additional experimental information on fugacities is available particularly for
substances of geological interest. The origin of these data are measurements of
chemical equilibria and their evaluation depends on sets of thermodynamic data of
the other components involved in the chemical eqUilibrium. The resulting fugacities vary significantly depending on the assumptions made for the other components. Geologists are familiar with the internally consistent sets of thermodynamic
data needed for the evaluation of the measured equilibria, but scientists working on
reference equations of state are usually not. Thus, it would be valuable to set up a
pure-component data base by calculating the corresponding fugacities from the
eqUilibrium data published mainly in geological literature.
For carbon dioxide, which can be used again as an example, only Haselton et aI.
(1978) have published pure-component fugacities which are derived from the
evaluation of their experimental results for the decarbonation of magnesite and
calcite. Figure 4.28 compares these data with results calculated from the equations
of state of Span and Wagner (1996) and of Sterner and Pitzer (1994). While the
24 For

a more detailed discussion on the shortcomings of the data sets of Shmonov and Shmulovich (1974) see Sterner and Pitzer (1994)

164

"I"""

4 Setting up Multiparameter Equations of State for Pure Substances

250

"0

---"""

.....,

'-'

1200K
1400K
1600 K

h
h
A>

200

S-

..s
N

f5,

P
.....
u

150

ro

Span and Wagner (1996)

Sterner and Pitzer (1994)

bIl
;:l

100

1000

2000

4000

3000

Pressure pi MPa
Fig. 4.28. Fugacities of carbon dioxide calculated from the equations of Span and Wagner
(1996) and Sterner and Pitzer (1994) at high temperatures. The corresponding experimental
results of Haselton et al. (1978) are given as symbols indicating the isotherm to which they belong.

semiempiric equation of Sterner and Pitzer was fitted to an extensive set of fugacities calculated from published equilibria data, the data of Haselton et al. were used
only for an assessment of the extrapolation behaviour during the development of
the equation of Span and Wagner. Nevertheless, the empirical reference equation,
which is fitted to data only up to pressures of 800 MPa and temperatures of
1073 K, yields a slightly better representation of these data up to 1600 K and more
than 3600 MPa. This observation clearly supports the claim that the extrapolation
behaviour of state-of-the-art multiparameter equations of state is far better than
usually expected.
At even higher pressures and temperatures shock-wave measurements of the
Hugoniot curve are available for some substances. The evaluation of the Hugoniot
relation

hh or

ho = 0.5 (Ph -

Po) (Po 1 + Ph 1 )

(4.164)
(4.165)

yields data for the enthalpy hh or the internal energy Uh as a function of pressure Ph
and density Ph at shock wave pressures up to tens of GPa and temperatures of
several thousand Kelvins; the index 0 corresponds to the initial state prior to release of the shock wave. Consideration of these data in the development of multiparameter equations of state results in nonlinear relations which require iterative
solutions for temperature. For optimised functional forms, nonlinear fits which
involve Hugoniot data hardly exercise an influence on the representation of these

4.6 Consideration of the Extrapolation Behaviour

30000
25000
c;j

Po.

:::E

--

I:l,

20000

Span et al. (l998b) I

Jacobsen et al. (1986)
Jacobsen and Stewart (1973)

o Nellis and Mitchel (1980)!

o Zubarev and Telegin (1962)
I

15000

;::l

...'"'"
(l)

Po.

165

10000
/

---

5000
0
800

1000

1200

1400

1600

1800

2000

Density p / (kg m-3)

Fig. 4.29. Plot of the Hugoniot curve with To =77.5 K and ho =h'(To) of nitrogen as calculated
from three multiparameter equations of state. The corresponding experimental results are given
as symbols.

data - the extrapolation behaviour of equations of state is closely related to their

functional form (see also Sect. 4.6.2). Any attempt to use Hugoniot data in linear
optimisation algorithms requires a precorrelation of the temperatures belonging to
the data points. Since no ppT data, which could verify the precorrelated temperatures, are available under these extreme conditions, this approach implies the risk
of distorting the experimental information. Thus, data for the Hugoniot curve
could only be used for comparisons until recently, but for state-of-the-art equations
these comparisons showed very reasonable results up to extreme pressures and
temperatures (see for instance Setzmann and Wagner, 1991; Span and Wagner,
1996).
During the development of the new reference equation of state for nitrogen
(Span et aI., 1998b/1999) Hugoniot data were used directly according to Eq.
4.164 as I1h = h(Ph, Ph) - h(po, Po) in combination with the nonlinear optimisation
algorithm of Tegeler et ai. (1997/1999; see Sect. 4.4.6). This procedure resulted
in an exceptionally good representation of the Hugoniot curve although the Hugoniot data were used with only low weights; their contribution to the sum of squares
was less than 1% of the total sum of squares. Figure 4.29 shows the plot of the
Hugoniot curve as calculated from three multiparameter equations of state for
nitrogen. Obviously, state-of-the-art reference equations of state are not necessarily inferior to special high-pressure equations with regard to the representation of
Hugoniot data, but it remains questionable whether the good representation of the
data by the recent equation of Span et ai. (1998b) is a kind of overfitting or not,

166

4 Setting up Multiparameter Equations of State for Pure Substances

since little is known about the accuracy of the data. In addition, when considered
as the only extrapolation criterion, the plot of the Hugoniot curve may be misleading (for an example, see Span and Wagner, 1997) - further criteria are required to
assess or to improve the extrapolation behaviour of a multiparameter equation of
state.

4.6.2 The Influence of the Functional Form

Generally speaking, the terms in empirical equations using the general set-up
shown in Eq. 3.25 25 are highly intercorrelated and it is assumed that the behaviour
of an empirical equation of state cannot be associated with the behaviour of single
terms in the equation. This conclusion is true in most cases, but not for the extrapolation to very high temperatures and very high pressures, which also correspond to
high densities. For high densities, the behaviour of the equation is influenced only
by polynomial terms and by exponential terms with exp( --0 I), if the exponential
terms are combined with high density powers (d i ~ "'5). For high temperatures,
which correspond to small values of T, terms with high temperature exponents ti
fade away as well. Under these conditions, which are typical for the region covered by Hugoniot data, one or a small number of leading polynomial terms can be
identified which determines the behaviour of the whole equation. Based on this, an
attempt was made by Span and Wagner (1997) to formulate demands on the
mathematical form of an empirical equation of state in order to ensure a more
reasonable extrapolation behaviour. These demands can be summarised in the
following way:
The number of polynomial terms in the equation should be small, if possible
less than 10.
Intercorrelations between the polynomial terms affecting the extrapolation to high
densities cannot be avoided completely this way, but they are reduced considerably. At the same time, an increased number of exponential terms with exp( --0 I) has
to be used to guarantee the necessary flexibility of the equation in the range of
intermediate densities. For functional forms like this, usually only one or two terms
determine the behaviour of the equation in the range of very high temperatures and
densities. The term which is dominant at high densities (high values of 0) and high
temperatures (small values of T) is the polynomial term with the smallest temperature exponent ti among the terms with the highest density exponents d i . For this
term, the following requirements can be formulated:
The coefficient ni has to be positive to yield a positive contribution to the residual pressure.
The temperature exponent should fulfil the condition 0 < ti < 1 since the pressure should increase on an isochore with increasing temperature but the compression factor should decrease.

25

Additional critical region terms included in some reference equations of state do not affect the
behaviour in the high pressure, high temperature region.

4.6 Consideration of the Extrapolation Behaviour

167

The density exponent d; has to be an integer value and should be equal to 3 or

4. This is a purely empirical finding which results from a study of the compressibilities calculated both from empirical and semiempirical equations of
state in the high pressure, high density region (see Span and Wagner, 1997).
These conditions were considered during the development of the new reference
equations of state for carbon dioxide (Span and Wagner, 1996), water (PruB and
Wagner, 1995; Wagner and PruB, 1997/2000), argon (Tegeler et aI., 1997 11999),
and nitrogen (Span et aI., 1998b/1999). For the new equation of state for carbon
dioxide, Fig. 4.30 shows the relative contributions of all polynomial terms, of all
exponential terms with exp(-a- 1), and of the leading polynomial term to the residual pressure pC; the plotted lines correspond to isotherms. In the region where the
available Hugoniot data indicate that carbon dioxide is still chemically stable,
roughly for reduced densities 4.5 < p I Pc < 5.5 and reduced temperatures
5 < TITc < 15, a polynomial term with d; = 3 and t; = 0.75 dominates the behaviour
of the equation with a contribution of more than 70% of the total residual pressure.
Since this dominant term fulfils the requirements given above, the whole equation
behaves reasonably in the high density limit. The negative contribution of the
exponential terms in the range of intermediate densities was not desired with respect to extrapolation but unavoidable for the representation of accurate data at
lower temperatures; at least with respect to basic properties like pressure, enthalpy,
and fugacity it does not affect the extrapolation up to the limits where spontaneous
disintegration occurs.
Similar to carbon dioxide, the final mathematical form is a compromise between
requirements for representing the data set and the extrapolation behaviour for the

.'"'"=
ca
.g
.
Q)

200%
150%

0.

.r;;

100%

Q)

-B

50%

0%

.I:J
.~

-50%

Q)

.s

.'S
o
U

exponential ter.!!l~
:.J~

51~- - -::-:-:~~ ..
-~.::- _- - - -1.0 Je- -.............

~,

"' ,-_-:_ \?lo'-

-100%+-~~~--~--r-~--,---r-~--~--~~-2

Reduced density pI Pc
Fig. 4.30. Contribution to the residual pressure of all polynomial terms, of all exponential terms
with exp(-0 1), and of the leading polynomial term of the equation of Span and Wagner (1996) for
carbon dioxide.

168

4 Setting up Multiparameter Equations of State for Pure Substances

other reference equations as well. But nevertheless, the extrapolation behaviour of

an empirical equation of state becomes predictable from its mathematical structure
by such investigations and unreasonable behaviour can be avoided. The new equations of state for carbon dioxide, water, argon, and nitrogen yield reasonable results up to extreme temperatures and pressures.
The results discussed above are applicable to formulations in terms of the compression factor as well, since the exponents d i and ti of the polynomial terms do not
change when the equation for Z is integrated to yield a r , see Sect. 3.1.2.1.

4.6.3 The Representation of Ideal Curves

Ideal curves are curves along which one property of a real fluid is equal to the
corresponding property of the hypothetical ideal gas at the same temperature and
density. Based on this very general definition, ideal curves can be defined for
almost every property, but usually the discussion is focused on the ideal curves of
the compression factor and its first derivatives; these curves are given in Table
4.11 together with their definitions. In the 1960s, there was an intensive investigation of ideal curves in order to specify criteria for a generalised behaviour of pure
fluids. Well known results are those of Brown (1960) which were summarised by
Rowlinson (1965) and the results of Gunn et al. (1966) and Miller (1970) on the
Joule-Thomson inversion curve. Less well known are the very detailed studies of
Morsy (1963), Straub (1964), and Schaber (1965) which have been published only
in German. More recently, Angus (1983) and de Reuck (1991) gave short summaries of the known characteristics of ideal curves. Although various authors have
stated that the representation of ideal curves is a sensitive test for the extrapolation

Table 4.11. The zeroth- and first-order ideal curves of the compression factor and their
definition in terms of the compression factor, Z(T,p), and of the residual part of the reduced
Helmholtz energy, a(T,p).
Designation

(Classical) Ideal curve

Definition in terms of the

Compression factor
Residual Helmholtz energy

Z=l

Boyle curve

(~)
ap

Joule-Thomson inv. curve

(az)
aT

Joule inversion curve

=0

=0

=0
(~)
aT p

a a-)_-0
(aoa.
2

4.6 Consideration of the Extrapolation Behaviour

169

behaviour of equations of state, systematic investigations have always dealt with

results for simple model fluids, with simple equations of state, or with values derived directly from experimental data or from compression factors tabulated for
corresponding states approaches. In order to verify whether ideal curves are really
useful for assessing the extrapolation behaviour of empirical equations of state,
Span and Wagner (1997) compared the ideal curves calculated from equations of
state for argon (Stewart and Jacobsen, 1989; Tegeler et aI., 1997 11999), nitrogen
(Jacobsen and Stewart, 1973; Jacobsen et al., 1986a; Span et al., 1998b/1999),
oxygen (Schmidt and Wagner, 1985), methane (Setzmann and Wagner, 1991),
ethane (Friend et aI., 1991), carbon dioxide (Sterner and Pitzer, 1994; Span and
Wagner, 1996), water (Wagner and PruB, 1997/2000), and helium (Sychev et al.,
1984) with each other and with the "theoretical" predictions.
Figure 4.31 shows a typical plot of the ideal curves discussed here in a reduced
pressure, temperature diagram with logarithmic axes. The broken lines indicate the
limits of the regions where primary data (usually ppT data) are available for the
corresponding substance. For reference substances with low critical temperatures
and pressures like nitrogen and argon, the Boyle, the Ideal, and the Joule-Thomson
inversion curve lie completely within the range covered by primary data; for helium, even the Joule inversion curve lies within this range. The situation changes if
substances with higher values for the critical temperature and the critical pressure
or with a more restricted data set are investigated. For carbon dioxide and methane
the Joule-Thomson inversion curve reaches into the extrapolation region; for wa-

1000
500

_________ ~i!r~l!e~

helium
--------------,------------

200

I::l;

.......

100

dJ

50

I:),

!3
'"'"
~

p..
"0
dJ

;::3

"0
dJ

20
10

.....
0

s.

5
2

0.5
0.5

'"~.
()

~.

~
"

"

10

20

Reduced temperature T/ Tc

50

100

Fig. 4.31. A typical plot of the considered ideal curves in a reduced double-logarithmic pressure,
temperature diagram. The dashed lines indicate the regions where primary data are available for
the substances considered by Span and Wagner (1997).

170

4 Setting up Multiparameter Equations of State for Pure Substances

ter, oxygen, and ethane the Boyle and the Ideal curve exceed the temperature range
and the Joule inversion curve also exceeds the pressure range covered by data.
When considering results of earlier investigations, certain features of the ideal
curves should be universal at least for simple substances (see Gunn et al., 1966;
Miller, 1970; Schaber, 1965). Table 4.12 summarises the results of a comparison
of these "theoretical predictions" with the corresponding values calculated from
the equations of state considered by Span and Wagner (1997). For nitrogen and
carbon dioxide, the results of different equations are given for comparison.
When analysing the results given in Table 4.12 in more detail, it becomes obvious that the numerical values for the characteristic points of the different ideal
curves are useful as criteria for an assessment of the extrapolation behaviour only
for simple substances with limited data sets - and for such substances other corresponding states approaches yield reliable results as well. A comparison between
the different equations of state for nitrogen and carbon dioxide explains this thesis.
Although the accuracy of reference equations of state for nitrogen in the high temperature, high pressure region has been improved substantially since 1973, the

Table 4.12. Characteristic values of the ideal curves as predicted from "theoretical" studies and
as calculated from accurate equations of state for helium, argon, methane, oxygen, nitrogen,
ethane, carbon dioxide, and water.
Subst. Equation
0.00

predicted values
He
Ar
C~

O2
N2

C 2Ht;
CO2
H 2O

Sychev et al. (1984)

-0.39
Tegeler et al. (1997)
0.00
Setzmann & Wagner (1991) 0.01
Schmidt & Wagner (1985) 0.02
Span et al. (1998b)
0.04
Jacobsen et al. (1986a)
0.04
Jacobsen & Stewart (1973) 0.04
Friend et al. (1991)
0.10
0.23 h )
Span & Wagner (1996)
0.23 h)
Sterner & Pitzer (1994)
Wagner & ProS (1997)
0.34

(J/)

OB IT=Te

d)

O( k=Te

e)

Jl rrutx .1T / (J I)

2.66

5.00

19.5

1.435

2.235

11.8/2.25

4.67
2.71
2.67
2.62
2.59
2.60
2.58
2.45
2.37
2.36
2.35

9.03
5.07
5.14
4.90
4.82
4.83
4.76
4.57
4.45
4.29
3.93

29.9
20.3

1.407
1.425
1.428
1.425
1.425
1.423
1.429
1.428
1.441
1.448
1.540

2.201
2.201
2.203
2.201
2.193
2.191
2.195
2.196
2.232
2.242
2.638

17.5/4.48
11.6/2.30
11.8/2.28
11.7/2.25
11.6/2.20
11.6/2.20
11.6/2.21
12.1/2.14

- g)
- g)

17.4
16.6
16.1
-g)

26.9
23.8
7.7

12.5/1.94
13.4 /1.90
18.7/1.96

Reduced temperature (J T / Te at which the Boyle and the Ideal curve end for p 0; at this
temperature the condition B(n 0 holds for the second virial coefficient.
b) Reduced temperature (J = T / Te at which the Joule-Thomson inversion curve ends for p = 0; at
this temperature the condition dB / dT B / T holds for the second virial coefficient.
c) Reduced temperature (J
T / Te at which the Joule inversion curve ends for p 0; at this
temperature the condition dB / dT 0 holds for the second virial coefficient.
d) Reduced density 0
p / pc on the Boyle curve for T Te
e) Reduced density 0 = p / pc on the Ideal curve for T = Te
I) Reduced pressure Jl
P / pc and temperature (J T / Te at the pressure maximum of the JouleThomson inversion curve.
g) Equations without a maximum in B(n yield no intersection between the Joule inversion curve
and the axis p o.
h) Calculated from an extrapolation of the vapor pressure equation given by Span and Wagner
(1996).
a)

Fig. 4.32. Plots of the zeroth- and first-order ideal curves of the compression factor calculated
from the equations of Span and Wagner (1996) and Sterner and Pitzer (1994) for carbon dioxide.

three investigated equations yield very similar results for the characteristic points
of the ideal curves; based on these results, no assessment of the equations is possible. For carbon dioxide, the differences between the equations of Span and Wagner (1996) and Sterner and Pitzer (1994) are larger. The difference observed for
the maximum pressures on the Joule-Thomson curve would be sufficient for conclusions regarding the reliability of the equations, if such differences would occur
for methane or nitrogen. But since carbon dioxide does not belong to the group of
simple substances, the observed differences are still not significant enough for an
assessment.
Nevertheless, the plot of the ideal curves itself contains important information
on the behaviour in the high temperature, high pressure region. To demonstrate the
sensitivity of this graphical criterion, Fig. 4.32 shows the plot of the ideal curves of
carbon dioxide as calculated from the equation of Span and Wagner (1996) and
Sterner and Pitzer (1994). From Fig. 4.27 it can be seen that the equation of
Sterner and Pitzer deviates from the data of Vukalovich and Altunin (1962) by up
to /).p / P z 1% for T / Tc $; 3.5 and P / Pc $; 8; in Fig. 4.32, these deviations result in
visible deformations of the Boyle and the ideal curve. At higher temperatures,
stronger deformations of the Joule-Thomson and the Joule-inversion curve occur.
Thus, in this order of magnitude, oscillations of an equation of state can be detected easily by an investigation of its ideal curves.