TECHNICAL DOCUMENTATION

Copyright (c), H.A. Kooijman, R. Taylor

October 1998

Preface

This book accompanies the ChemSep program, which was developed to allow students to do separation calculations on ordinary personal computers. This book is not a guide where we show how to use ChemSep (see the ChemSep user manual for that) but it is intended to supply technical background to help the user in his selection of models and correlations. It is hoped that sensible selections can be made by providing information on, descriptions of, and references to the models and correlations that are employed in ChemSep. Although we have tried to be as extensive as possible, it is impossible to describe all models and their underlaying theory, so references are given for further reading. There are probably many more literature models and correlations than are available in ChemSep, but we have tried to be as comprehensive as we could. Sometimes a choice had to be made in which models to implement without having any criteria to discriminate between models. Furthermore, not all models are applicable to a particular regime of operation. We try to adapt ChemSep as much as possible to comply with all model limitations and user requirements. This book serves as a replacement for the ``manual" information files that we used to distribute with ChemSep. Therefore, some parts of this document might still be incomplete or unorganized and any suggestions or remarks are welcome. Of course, any remarks on the ChemSep program are welcome as well.

This book is written in LaTeX, a complete typesetting language, and set in the standard Times-Roman 11 point font. It is also provided with ChemSep in ASCII text form (file CHEMSEP.TXT) for online reference which was generated with a LaTeX to ASCII converter. The conversion is limited, with the result that the ASCII text file contains some unconverted LaTeX formatting. A PostScript file (BOOK.PS) also generated by LaTeX can be downloaded from our ftp site. Harry Kooijman
Ross Taylor

Acknowledgements

Many people have helped to shape ChemSep. The project was started in 1988 at Delft University (The Netherlands) by Harry Kooijman, Arno Haket and Ross Taylor. The purpose was to make an interactive interface for doing equilibrium stage calculations on the PC platform. It had to be easy enough for use by students with little computer exposure and yet sufficiently comprehensive to solve the various problems encountered in a course on separation processes.

We would like to express our appreciation to Professor Hans Wesselingh (now at the University of Groningen, the Netherlands) who initially promoted the project and made various resources available and encouraged us by letting students use the program for their course work. This has been an indispensable source of feedback that has helped us to improve the program. We also like to thank Peter Verheijen for his enthusiasm and contributions in the early years of the project. Also, various students have worked on projects to check and improve the programs and documentation, which was very helpful. Finally, ChemSep owes its very existence to the Internet which enabled the authors to keep in touch and continue development while living on different continents.

CONTENTS

Chapter 1. Solving Nonlinear Equations

In this chapter we discuss the methods employed in ChemSep to solve the separation problems at hand. Side-issues such as how to start an iterative method or how ChemSep solves the resulting linear system of equations also pass the revue.

1.1 Newton's method

ChemSep uses Newton's method to solve the system of (MESH) equations derived from the flash or column problems. Newton's method is a Simultaneous Correction (SC) method that each time corrects all the variables. To use it, the equations to be solved are written in the form
  F(x) = 0
where F is a vector consisting of all the equations to be solved and x is, again, the vector of variables. A Taylor series expansion of the function vector around the point x_o which the functions are evaluated gives (ignoring second and higher order terms):
  F(x) = F(x ) + J(x - x )
            o           o

where J is the Jacobian matrix of partial derivatives of F with respect to the independent variables x:
  J   = (d F  / d x )
   ij       i      j

If x is the actual solution to the system of equations, then F(x)=0 and we can rewrite the above equation as:
  J(x - x ) = - F(x )
         o         o

This linear system of equations may be solved for a new estimate of the vector x. If the new vector, x, obtained in this way does not actually satisfy the set of equations, F, then the procedure can be repeated using the calculated x as a new x. The entire procedure is summarized below.
  1. Set iteration counter, k, to zero, estimate x_o
  2. Solve linearized equations for x_k+1
  3. Check for convergence; if not obtained, increment k and return to step 2.
Solving the linear system does not require a full matrix inversion of the Jacobian and is normaly done with Gaussian elimination or some type of decomposition technique. If the Jacobian has a lot of zero entries (i.e. it is sparse) then the linear system can be much more efficiently solved by using a sparse linear solver. For Jacobians with specific structures special solvers can be employed which are more efficient than a complete elimination or decomposition. One very important property of the Newton's method is that the convergence is scale invariant and independent of the ordering of the equations. This means that the same convergence is obtained if one of the equations is multiplied by some number or if the equations are reordered in a different manner. This is very important, because this means we are free to order the equations to obtain a special Jacobian which might enable the use of a special solver. It also makes the method applicable to a wider range of problems and without requiring the user to scale equations or variables.

An important drawback of the Newton's method can be its sensitivity to the initial guess, x_o, since quadratic convergence is only achieved ``close" to the solution. In order to obtain convergence, Newton's method requires that reasonable initial estimates be provided for all independent variables. It is obviously impractical to expect the user of a SC method to guess this number of quantities. Thus, the designer of a computer code implementing a SC method must provide one or more methods of generating initial estimates of all the unknown variables. Several techniques have been developed to improve the convergence away from the solution and to prevent the method from taking a step in a wrong direction. The simplest and most common technique is to ``damp" the step of each variable to some range or fraction of the Newton step. However, this damping also reduces the methods effectiveness.

Simultaneous correction procedures have shown themselves to be generally fast and reliable, having a locally quadratic convergence rate in the case of Newton's method, and these methods are much less sensitive to difficulties associated with nonideal problems than are tearing methods. They also lend themselves to be easier extended with optimization, parametric sensitivity, or continuation methods.

1.2 Continuation method

A simple implementation of a continuation method is incorporated in ChemSep for more difficult problems. Continuation methods use a parameter to make a path from a known solution for a simplified model to the desired solution of the complete model. For example the Newton homotopy starts with the initial guess as model and follows the path to the solution of the real problem by solving
  0 = (1-t) (F(X ) - F(X)) + t F(X)
                o

where t varies from 0 (where X=X_o) to 1 (where F(X)=0). Better continuation methods can be formulated while using a parameter which has some physical significance. In separations problems the most appropriate choice would be the degree to which mass transfer (between the present phases) prevails. In the equilibrium model the stage efficiency and in the nonequilibrium model the mass transfer rates represent this degree. Thus, they will be multiplied with a parameter t which will vary from 0 (no separation at all) to 1 (actual separation).

Chapter 2. Property Models

This chapter discusses the thermodynamic and physical property models available in ChemSep. The selection of these models can be quite important for the results produced by ChemSep. Most formulae are repeated here but additional reading is available in two main sources: References are in between parentheses, by combining the letter A or B with the page number, for example (A43). The model "types" are grouped by ChemSep menu.

2.1 Thermodynamic Properties

2.1.1 K-value models

2.1.2 Activity coefficient models

Here we discuss the activity coefficient models available in ChemSep. For an in depth discussion of these models see the standard references. For the calculation of activity coefficients and their derivatives (for diffusion calculations) see also Kooijman and Taylor (1991).

2.1.3 Vapour pressure models

2.1.4 Equations of State (EOS)

Three types of equations of state may be selected in ChemSep; Ideal Gas, Virial, and Cubic EOS. The fugacity coefficient of an ideal gas mixture (B3) is unity (since the fugacity represents the deviation from an ideal gas, and we use the natural logarithm of the fugacity as the fugacity coefficient). The pressure relation for an ideal gas is:
  P = (R T / V)
The Virial and cubic EOS are discussed in the sections below.

2.1.5 Virial EOS

2.1.6 Cubic EOS

2.1.7 Enthalpy

2.2 Physical Properties

A number of different polynomials is implemented in ChemSep to evaluate physical properties over a certain temperature range. These temperature correlations are assigned a unique number in the range of 0-255 (see Table 2.1). For each up to 5 parameters (A-E) are available. Table 2.2 shows which pure component properties can be modeled with temperature correlations and their typical correlation number.

Table 2.1. Temperature correlations

Equation Parameter(s) Formula
number
2 A,B A + B T
3 A-C A + B T + C T^2
4 A-D A + B T + C T^2 + D T^3
10 A-C exp ( A - B / C+T )
100 A-E A + B T + C T^2 + D T^3 + E T^4
101 A-E exp ( A + B / T + C ln T + D T^E )
102 A-D A T^B / 1 + C/T + D/T^2
103 A-D A + B exp ( -C / T^D )
104 A-E A + B / T + C / T^3 + D / T^8 + E / T^9
105 A-D A / B^(1 + (1 - T / C)^D)
106 A-E A (1-T_r)^(B + C T_r + D T_r^2 + E T_r^3 )
107 A-E A + B ( C / T/sinh (C / T)^2 + D ( D / T/cosh (D / T)^2
All types of equations may be used for any of the physical properties but, of course, some formulas were specifically developed for prediction of particular properties. Besides the parameters A-E the temperature limits of the correlation must also be present. If the temperature specified falls out of the temperature range of a correlation (or the temperature limits are missing/incomplete) normally an alternative (default) method will be used automatically.

Table 2.2. Component properties with the typical correlation number

Liquid density 105
Vapour pressure 101
Heat of vaporisation 106
Liquid heat capacity 100
Ideal gas heat capacity 107
Second virial coefficient 104
Liquid viscosity 101
Vapour viscosity 102
Liquid thermal conductivity 100
Vapour thermal conductivity 102
Surface tension 106
Ideal gas heat capacity (Reid Prausnitz and Poling) 4
Antoine 10
Liquid viscosity (Reid, Prausnitz and Sherwood) 2
Physical properties models can be selected manually or the automatic selection can be used (which is the default). Below we discuss the models for calculating physical properties for pure components and mixtures, for vapour or liquid phases. ChemSep uses an automatic selection when no model is selected at all and the selection is left as *'s. Depending on range, phase, conditions, data availability, and required property ChemSep will make a guess of the best model to use. ChemSep does allow you to pick default models, and will use them if the model's range is valid. In case a property cannot be computed with a specific model it will use an estimation method or a fixed estimate (it is a good habit to check predicted physical properties when possible).

Table 2.3. Default physical property correlations

Mixture liquid density Rackett
Component liquid density Polynomial
Vapour density Cubic EOS
Mixture liquid viscosity Molar averaging
Component vapour viscosity Polynomial/Letsou-Stiel
Mixture vapour viscosity Brokaw
Component vapour viscosity Polynomial
Mixture liquid thermal conductivity Molar average
Component liquid thermal conductivity Polynomial
Mixture vapour thermal conductivity Molar average
Component vapour thermal conductivity Polynomial/9B-3
Liquid diffusivity Kooijman-Taylor/Wilke-Chang
Vapour diffusivity Fuller et al.
Mixture surface tension Molar average
Component surface tension Polynomial
Liquid-liquid interfacial tension Jufu et al.
Certain methods require mixture (critical) properties, commonly used mixing rules are:

             c
  T    = sum      x  T   
   c,m        i=1  i  c,i


             c
  V    = sum      x  V   
   c,m        i=1  i  c,i


             c
  Z    = sum      x  Z   
   c,m        i=1  i  c,i

  P    = Z    R T    / V   
   c,m    c,m    c,m    c,m


           c
  M  = sum      x  M 
   m        i=1  i  i

which will be referred as the "normal" mixing rules. Reduced properties will be calculated by:
  T  = T / T 
   r        c

  P  = P / P 
   r        c

  V  = V / V 
   r        c

unless specified otherwise.

2.2.1 Liquid density

Mixture liquid densities (in kmol/m^3) are calculated with: Pure component liquid densities are computed from the Peng-Robinson EOS for temperatures above a components critical temperature, otherwise with one of the folowing methods:

The pure component liquid densities are corrected for pressure effects with the correction of Thomson et al. (1982) as described for the Hankinson and Thompson method for mixtures.

2.2.2 Vapour density

Vapour densities are computed with the equation of state selected for the thermodynamic properties (possible selections are Ideal gas EOS, Virial EOS, and Cubic EOS).

2.2.3 Liquid Heat Capacity

The mixture liquid heat capacity is the molar average of the component liquid heat capacities, which are generally computed from a temperature correlation. Alternatively the liquid heat capacity could be computed from a corresponding states method and the ideal gas capacity. Rowlinson (1969, see A140) proposed a Lee-Kesler heat capacity departure function which was later modified to:

   L       ig                      -1                               1/3   -1               -1
  C     - C    = 1.45 + 0.45 (1-T )   + 0.25 w [ 17.11 + 25.2 (1-T )    T    + 1.742 (1-T )   ]
    p,i      p                   r                                r      r               r

However, in ChemSep the temperature correlation is used for all temperatures to prevent problems arrising from using different liquid heat capacity methods in the same column (which especially trouble nonequilibrium models). Liquid heat capacities could also be computed from the selected thermodynamic models to circumvent this problem.

2.2.4 Vapour Heat Capacity

The mixture vapour heat capacity is the molar average of the component vapour heat capacities, which are computed from the ideal gas heat capacity (RPP) 4 parameter temperature correlation. If no parameters for this correlation are present, the vapour heat capacity temperature correlation is used (if within the temperature range).

2.2.5 Liquid Viscosity

Mixture liquid viscosity are computed from DIPPR procedure 8H from the pure component liquid viscosities from:

        L        c             L
  ln eta   = sum      z  ln eta  
         m        i=1  i        i

where z_i are either the mole fractions (for molar averaging, the default) or alternatively the weight fractions for mass averaging. A better method is from Teja and Rice (1981, A479). However, this method requires interaction parameters. Here a different mixing rule (for T_cijV_cij) is used which improves the model predictions with unity interaction coefficients:

              c
  w  = sum      x  w 
   m       i=1   i  i


              c
  M  = sum      x  M 
   m       i=1   i  i


               c        c
  V   = sum      sum      x  x  V   
   cm       i=1      j=1   i  j  cij


               1/3      1/3  3
  V    = (( V      + V      )  / 8)
   cij       ci       cj


                c        c
  T   = (sum      sum      x  x  T   V    / V  )
   cm        i=1      j=1   i  j  cij cij    cm

  T   V    = psi   (T   V   + T   V   / 2)
   cij cij      ij   ci  ci    cj  cj

where psi_ij is set to unity for all components. The liquid viscosity of the mixture is computed from two reference components
  ln (epsilon  eta ) = ln (epsilon  eta ) + [ ln (epsilon  eta ) - ln (epsilon  eta ) ] (( w  -w  / w  - w  ))
             m    m               1    1                 2    2               1    1        m   1    2    1

with epsilon defined as

                 2/3
  epsilon  = (V      / sqrt(T   M ))
         i     ci            ci  i

and the reference component vioscosities are evaluated at T T_ci / T_cm. Component liquid viscosities are calculated from the liquid viscosity temperature correlation if the temperature is within the valid range. Otherwise the component viscosity is computed with DIPPR procedure 8G, the Letsou-Stiel method (1973, see A471):

                 1/6                2/3
  xi = 2173.424 T       / sqrt(M ) P      
                    c,i         i      c,i


    (0)                               2      -5
  xi    = ( 1.5174 - 2.135 T  + 0.75 T   ) 10  
                            r          r


    (1)                              2      -5
  xi    = ( 4.2552 - 7.674 T  + 3.4 T   ) 10  
                            r         r


     L        (0)       (1)
  eta   = ( xi    + w xi    ) / xi
      i

Alternatively the simple temperature correlation given in Reid et al. (RPS liquid viscosity, see A439) can be used:
  log eta = A + B / T
A high pressure correction by Lucas (A436) is used to correct the influence of the pressure on the liquid viscosity:

                                 A
  eta = (1 + D (Delta P  / 2.118)  / 1 +C w  Delta P ) eta  
                       r                   i        r     SL

where eta_SL is the viscosity of the saturated liquid at P_vp, and
  Delta P  = (P - P  ) / P  
         r         vp     ci


                        -4           -0.03877
  A = 0.9991 - [4.674 10  /(1.0523 T          - 1.0513)]
                                    r


                                        2             3
  C = - 0.07921 + 2.1616 T  - 13.4040 T   + 44.1706 T  
                          r            r             r


              4             5             6             7
  - 84.8291 T   + 96.1209 T   - 59.8127 T   + 15.6719 T  
             r             r             r             r


                        2.573 0.2906
  D = [0.3257/(1.0039-T      )      ]-0.2086
                       r

2.2.6 Vapour Viscosity

Mixture vapour viscosities are computed using DIPPR procedure 8D-1 from component viscosites as follows:

     L        c           L
  eta   = sum      (x  eta   / sum x  phi  )
      m        i=1   i     i        i    ij

where the interaction parameters phi_ij can be calculated by Wilke's (1950) method:

                                               1/4 2
  phi   = ((1 + sqrt( \eta  / \eta  ) (M  / M )   )  / sqrt( 8 (1 + M  / M ) ) )
     ij                   i       j     i    j                       i    j

or by Brokaw's method:
  phi   = S A sqrt( \eta  / \eta  )
     ij                 i       j


                                  1/4
  sm = ( 4 / (1+M /M ) (1+M /M ) )   
                 j  i      i  j


                                                0.45                                  0.45
  A = (sm / sqrt(M /M )) ( 1 + ((M /M  - (M /M )    ) / 2 (1 + M /M )) + ((1 + (M /M )    ) / sqrt(sm) (1 + M /M )) )
                  i  j            i  j     i  j                 i  j             i  j                        i  j

If the Lennard-Jones energy parameter, epsilon (in Kelvin), and the Stockmayers polar parameter, delta, are known, S is calculated from:

                                                                                                  2                                    2
  S = (1 + sqrt( (T/\epsilon ) (T/\epsilon ) ) + (delta  delta /4) / sqrt(1 + T/\epsilon  + \delta   / 4) sqrt(1 + T/\epsilon  + \delta   / 4))
                            i             j            i      j                         i          i                         j          j

otherwise it is approximated by S=1. epsilon and delta can be estimated from:

                       3.6
  epsilon = 65.3 T    Z      
                  c,i     c,i


                  5     2
  delta = 1.744 10 9 (mu  / V    T   )
                             b,i  b,i

Where mu is the dipole moment in Debye. Vapour viscosities are a function of pressure and a correction is normally applied. Mixture properties are computed with the "normal" mixing rules. DIPPR procedure 8E can be used to compute the high pressure viscosity:
  rho  = 1 / V   
     c        c,m

  rho  = rho / rho 
     r            c


                  1/6                2/3
  xi = 2173.4241 T       / sqrt(M ) P      
                     c,m         m      c,m


                                         1.85
  A = exp (1.4439 rho ) - exp (-1.111 rho     )
                     r                       r


             -7
  B = 1.08 10   A / xi
  eta   = eta + B
     hp

where rho is the vapour mixture molar density.

Both Wilke's and Brokaw's method require pure component viscosities. These are normally obtained from the vapour viscosity temperature correlations, as long as the temperature is within the valid temperature range. If not, then the viscosity can be computed with the Chapman-Enskog kinetic theory (see Hirschfelder et al. 1954 and A391-393):


   *
  T  = T / epsilon

               * -b               *                *
  Omega  = a (T )   + c / exp (d T ) + e / exp (f T )
       v


     V           -             2                    2    *
  eta  = 26.69 10 7 M T / sigma  (Omega  + 0.2 delta  / T )
                                       v

where the collision integral constants are a=1.16145, b=0.14874, c=0.52487, d=0.77320, e=2.16178, and f=2.43787. The viscosity may also be computed with the Yoon and Thodos method (DIPPR procedure 8B):

                   1/6                2/3
  xi  = 2173.4241 T       / sqrt(M ) P      
    i                 c,i         i      c,i


     V            b                                     8
  eta   = (1 + a T   - c exp (d T-r) + e exp (f T ) / 10  xi)
      i            r                             r

where the constants a-f are given in Table 2.4.

Table 2.4. Constants for the Yoon-Thodos method

Hydrogen Helium Others
a=47.65 a=52.57 a=46.1
b=0.657 b=0.656 b=0.618
c=20.0 c=18.9 c=20.4
d=-0.858 d=-1.144 d=-0.449
e=19.0 e=17.9 e=19.4
f=-3.995 f=-5.182 f=-4.058
Another method for calculating the vapor viscosity is the Lucas (A397) method:

          -7          0.618
  eta = 10   [0.807 T       - 0.357 exp (-0.449T ) +
                     r                          r


                                   o   o
  0.340 exp (-4.058 T ) + 0.018] F   F   / xi
                     r            p   q


                      3    -5    4   1/6
  xi = 0.176 (( T  / M  (10   P )  ))   
                 c             c

where F_p^o and F_q^o are polarity and quantum correction factors. The polarity correction depends on the reduced dipole moment:

                             -30 2    -5         2
  mu  = 52.46 ((mu / 3.336 10   )  (10   P ) / T  )
    r                                     c     c

If mu_r is smaller than 0.022 then the correction factor is unity, else if it is smaller than 0.075 it is given by

    o                         1.72
  F   = 1 + 30.55 (0.292 - Z )    
   p                        c

else by

    o                         1.72
  F   = 1 + 30.55 (0.292 - Z )     [0.96 + 0.1 (T  - 0.7)]
   p                        c                    r

The quantum correction is only used for quantum gases He, H_2, and D_2,

    o         0.15                     2 1/M
  F   = 1.22 Q     (1 + 0.00385[(T -12) ]    sign(T -12) )
   q                              r                r

where Q=1.38 (He), Q=0.76 (H_2), Q=0.52 (D_2). There is also a specific correction for high pressures (A421) by Lucas.

                   o
  eta = Y F  F  eta 
           p  q


               e       f          d -1
  Y = 1 + (a P   / b P   + (1+c P  )  )
              r       r          r


               o       -3     o
  F  = (1 + (F   - 1) Y   / F  )
   p          p              p


               o        -1               4      o
  F  = (1 + (F   - 1) [Y   - 0.007 (ln Y) ] / F  )
   q          q                                q

where eta^o refers to the low-pressure viscosity (note that the original Lucas method has a different rule for Y if T_r is below unity, however, this introduces a discontinuity which is avoided here). The parameters a through f are evaluated with:

               -3                    -0.3286
  a = (1.245 10   / T ) exp 5.1726 T        
                     r              r

  b = a (1.6553 T  - 1.2723)
                 r


                                 -37.7332
  c = (0.4489 / T ) exp 3.0578 T         
                 r              r


                                 -7.6351
  d = (1.7368 / T ) exp 2.2310 T        
                 r              r

  e = 1.3088

                           0.4489
  f = 0.9425 exp -0.1853 T       
                          r

where, in case T_r is below unity, T_r is taken to be unity. For mixtures the Lucas model uses the following mixing rules:

               c
  T   = sum      y  T  
   cm       i=1   i  ci


               c
  V   = sum      y  V  
   cm       i=1   i  ci


               c
  Z   = sum      y  Z  
   cm       i=1   i  ci

  P   = R T   Z   / V  
   cm      cm  cm    cm


              c
  M  = sum      y  M 
   m       i=1   i  i


     o          c       o
  F    = sum      y  F   
   pm        i=1   i  pi


     o            c       o
  F    = A sum      y  F   
   qm          i=1   i  qi

where A is a correction factor depending on the molecular weights of the components in the mixture. Let H denote the component of highest molecular weight and L of lowest, then if M_H/M_L > 9:

                        0.57
  A = 1 - 0.01 (M  / M )    
                 H    L

else A=1. The mixture vapor viscosity is computed with the Lucas method as for a component which has the mixture properties T_cm, P_cm, M_m, F_pm^o, and F_qm^o. Therefore, the method is not interpolative in the same way as the techniques of Wilke and Brokaw (that is, the method does not necessarily lead to pure component viscosity eta_i when all y_j=0 except y_i=1).

2.2.7 Liquid Thermal Conductivity

The mixture liquid thermal conductivity, lambda^L_m (W/m K), can be computed using the following methods from the component liquid thermal conductivities: A correction is applied when the pressure is larger than 3.5 bar:

                    1.2                                     1.4
  lambda   = (0.63 T     P  / (30 + P ) + 0.98 + 0.0079 P  T     ) lambda
        hp             r  r          r                   r     r

This is DIPPR procedure 9G-1 where the mixture parameters are computed by the "normal" mixing rules. Component liquid thermal conductivities are calculated from one of the following methods:

2.2.8 Vapour Thermal Conductivity

If the system pressure is larger than 1 atmosphere a corection is applied according to DIPPR procedure 9C-1. Mixture parameters are computed using the "normal" mixing rules. Critical and reduced densities are computed from:
  rho  = (1 / V   )
     c         c,m

  rho  = (rho / rho )
     r             c

If the reduced density is below 0.5 then a=2.702, b=0.535, and c=-1; if the reduced density is witin [0.5,2] then a=2.528, b=0.67, and c=-1.069; otherwise a=0.574, b=1.155, and c=2.016. The high pressure thermal conductivity correction is then calculated from:

                      -8                                   1/6       2/3         5
  Delta lambda = (a 10   (exp (b rho ) + c) / (( sqrt(M ) T       / P       ) ) Z     )
                                    r                  m      c,m       c,m       c,m

which must be added to the calculated thermal conductivity for low pressure.

Pure component vapour thermal conductivities are estimated from the following methods:

2.2.9 Liquid Diffusivity

Generalized Maxwell-Stefan binary diffusion coefficients DD_ij are computed from the Kooijman-Taylor (1990) correlation where


    k      o
  DD    = D   , k=i
     ij     ij



    k      o
  DD    = D   , k=j
     ij     ji



    k           o    o
  DD    = sqrt(D    D   ), k != i, k != j
     ij          ik   jk




             c       k    x_k
  DD   = sum      (DD    )   
    ij        i=1     ij

Liquid binary infinitive diffusion coefficients (D^o_ij) are normally computed by the Wilke-Chang method unless selected otherwise. The following models are available:

2.2.10 Vapour Diffusivity

Generalized Maxwell-Stefan binary diffusion coefficients D_ij are equal to the normal binary diffusion coefficients (since the gas is considered an ideal system for which the thermodynamic matrix is the identity matrix). Normally these are computed with the Fuller-Schettler-Giddings method (see A587) but if Fuller volume parameters are missing the Wilke-Lee modification of the Chapman-Enskog kinetic theory is used.

2.2.11 Surface Tension

Mixture: Component surface tensions are only determined for temperatures below the component's critical temperature, otherwise it is assumed that the component does not contribute to the mixture surface tension (i.e. sigma_i=0). The following methods are available:

2.2.12 Liquid-Liquid Interfacial Tension

This property is only required for simulating Liquid-Liquid extractors with the nonequilibrium model. API method 10B3 uses the calculated surface tensions for both liquid phases and the interfacial tension, sigma^,, is computed from

       ,
  sigma  = sigma  + sigma  - 1.1 sqrt( \sigma  \sigma  )
                1        2                   1       2

This method generally overpredicts the interfacial tension for aqueous systems. We use a general method from Jufu et al. (1986):

              ,,     ,
  X = - ln ( x    + x   + x   )
                1     2    3r


                                     ,,        ,
  sigma , = (K R T X / A   exp (X) (x    q  + x   q  + x   q ))
                        w0             1  1     2  2    3r  3

with A_w0=2.5 10^5 (m^2/mol), R=8.3144 (J/mol/K), K=0.9414 (-), and q_i is the UNIQUAC surface area parameters of the components i. The components are ordered in such a manner that component 1 and 2 are the dominating components in the two liquid phases. Then the rest of the components are lumped into one mole fraction, x_3. This lumped mole fraction is taken for the phase which has the largest x_3 (the richest). q_3 is the molar averaged q for that phase for all components except 1 and 2.

Symbol List

a, b    Cubic EOS parameters
B         Second virial coefficient (m3/kmol)
c         Number of components
C_p       Mass heat capacity (J/kg.K)
D_ij    Binary diffusion coefficient (m^2/s)

DD_ij  Binary Maxwell-Stafan diffusion coefficient (m^2/s)
D^o_ij  Infinite dilution binary diffusion coefficient (m^2/s)
K_i       K-value of component i, equilibrium ratio (K_i=y_i/x_i)
k_ij    Binary interaction coefficient (for EOS)
M         Molecular mass (kg/kmol)
R         Gas constant = 8134 (J/kmol K)
r         Radius of gyration (Angstrom)
P         Pressure (Pa)
P^*,P_vap   Vapour pressure (Pa)
P_i       Parachor (m^3 kg^1/4 / s^1/2) of component i
PF        Poynting correction
q         UNIQUAC surface area parameter
T         Temperature (K)
T_r       Reduced temperature (T_r=T/T_c)
T_b       Normal boiling temperature (K)
V         Molar volume (m^3/kmol)
V_b       Molar volume at normal boiling point (m^3/kmol)
V_s       Saturated molar volume (m^3/kmol)
w         Weight fraction (of component)
x         Liquid mole fraction (of component)
y         Vapour mole fraction (of component)
Z         Compressibility
Z_R       Racket parameter
Greek:
alpha    Attractive parameter in EOS
w    Acentric factor
Omega_a, Omega_b  EOS parameters
Omega_v  Collision integral for viscosity
Omega_D  Collision integral for diffusion
gamma    Activity coefficient
delta    Stockmayer parameter
epsilon  Molecular energy parameter (K)
lambda   Thermal conductivity (W/m.K)
rho      Molar density (kmol/m^3)
eta      Viscosity (Pa.s)
phi_i    Fugacity coefficient of component i
phi_s    Association factor for solvent s (Hayduk-Laudie)
phi_ij   Interaction parameter for viscosities
phi_i    Fugacity coefficient of component i
phi^*  Pure fugacity coefficient at saturation
sigma    Surface tension (N/m)
            Collision diameter (Angstrom)
sigma_b  Surface tension at T_b (N/m)
sigma ,  Liquid-liquid interfacial tension (N/m)
mu       Dipole moment (Debye)
xi       Inverse viscosity (defined in text)
Superscripts:
L    Liquid
V, G  Vapour, gas
*    Saturated liquid,
       T/epsilon
Subscripts:
b    at normal boiling point
c    critical
i    of component i
j    of component j
m    mixture
r    reduced
s    saturated liquid
Abbreviations:
EOS    Equation of State
RK     Redlich-Kwong
SRK    Soave Redlich-Kwong
PR     Peng-Robinson

References

Digulio, Teja, Chem. Eng. J., Vol. 38 (1988) pp. 205.

E.N. Fuller, K. Ensley, J.C. Giddings, ``A New Method for Prediction of Binary Gas-Phase Diffusion Coefficients'', Ind. Eng. Chem., Vol. 58, (1966) pp. 19--27.

E.N. Fuller, P.D. Schettler, J.C. Giddings, ``Diffusion of Halogonated Hydrocarbons in Helium. The Effect of Structure on Collision Cross sections'', J. Phys. Chem., Vol. 73 (1969) pp. 3679--3685.

W. Hayduk, H. Laudie, ``Prediction of Diffusion Coefficients for Nonelectrolytes in Dilute Aqueous Solutions'', AIChE J., Vol. 20, (1974) pp. 611--615.

W. Hayduk, B.S. Minhas, ``Correlations for Prediction of Molecular Diffusivities in Liquids'', Can. J. Chem. Eng., 60, 295-299 (1982); Correction, Vol. 61, (1983) pp. 132.

J.O. Hirschfelder, C.F. Curtis, R.B. Bird, Molecular Theory of Gases and Liquids, Wiley, New York (1954).

F. Jufu, L. Buqiang, W. Zihao, ``Estimation of Fluid-Fluid Interfacial Tensions of Multicomponent Mixtures'', Chem. Eng. Sci., Vol. 41, No. 10 (1986) pp. 2673--2679.

R. Taylor, H.A. Kooijman, ``Composition Derivatives of Activity Coefficient Models (For the estimation of Thermodynamic Factors in Diffusion)'', Chem. Eng. Comm., Vol. 102, (1991) pp. 87--106.

A. Letsou, L.I. Stiel, AIChE J., Vol. 19 (1973) pp. 409.

Lielmezs, Herrick, Chem. Eng. J., Vol. 32 (1986) pp. 165.

J.M. Prausnitz, T. Anderson, E. Grens, C. Eckert, R. Hsieh, J. O'Connell, Computer Calculations for Multicomponent Vapor-Liquid and Liquid-Liquid Equilibria, Prentice-Hall (1980).

J.S. Rowlinson, Liquids and Liquid Mixtures, 2nd Ed., Butterworth, London (1969).

R.C. Reid, J.M. Prausnitz, T.K. Sherwood, Properties of Gases and Liquids, 3rd Ed., McGraw-Hill, New York (1977).

R.C. Reid, J.M. Prausnitz and B.E. Poling, The Properties of Gases and Liquids, 4th Ed., McGraw-Hill, New York (1988).

M.A. Siddiqi, K. Lucas, ``Correlations for Prediction of Diffusion in Liquids'', Can. J. Chem. Eng., Vol. 64 (1986) pp. 839--843.

A.S. Teja, P. Rice, ``Generalized Corresponding States Method for the Viscosities of Liquid Mixtures'', Ind. Eng. Chem. Fundam., Vol. 20 (1981) pp. 77-81.

S.M. Walas, Phase Equilibria in Chemical Engineering, Butterworth Publishers, London (1985).

C.R. Wilke, J. Chem. Phys., Vol. 18 (1950) pp. 617.

C.R. Wilke, P. Chang, ``Correlation of Diffusion Coefficients in Dilute Solutions'', AIChE J., Vol. 1 (1955) pp. 264-270.

C.R. Wilke, C.Y. Lee, Ind. Eng. Chem., Vol. 47 (1955) pp. 1253.

P.H. Winterfeld, L.E. Scriven, H.T. Davis, AIChE J., Vol. 24 (1978) pp. 1010.

Chapter 3. Flash Calculations

A flash is a one stage operation where a (multiple phase) feed is "flashed" to a certain temperature and/or pressure and the resulting phases are separated. The flash in ChemSep deals only with two different phases leaving, a vapour and a liquid. Liquid-Liquid or multiphase Vapour-Liquid-Liquid flashes are currently not yet supported in ChemSep. For more information see also the general references given at the end of this chapter.

3.1 Equations

The vapour and liquid streams leaving the flash are assumed to be in equilibrium with each other. The equations that model equilibrium flashes are summarized below:

where F is the molar feedrate with component mole fractions z_i. V and L are the leaving vapour and liquid flows with mole fractions y_i and x_i, respectively. Equilibrium ratios K_i and enthalpies H are computed from property models as discussed in chapter 2. Q is defined as the heat added to the feed before the flash. If we count the equations listed, we will find that there are 2c+3 equations, where c is the number of components. As flash variables we have (depending on the type of flash):

Since we have 2c+3 equations, two of the 2c+5 variables above need not be specified. ChemSep allows the following nine flash specifications:

3.2 Solution of the Flash Equations

FLASH uses Newton's method for solving flash problems as well as simpler bubble and dew point calculations. The vector of variables used in the PQ-FLASH is:


     T
  (X)  = (V, y  , y  ... y  , T, x  , x  ... x  , L)
              1    2      c       1    2      c

the vector of functions, (F), is:


     T
  (F)  = (TMB ,CMB , CMB  ... CMB , H, EQM , EQM  ... E , SUM),
                  1     2        c        1     2      c

The structure of the Jacobian matrix [J] is shown below:

References

Henley, E.J., J.D. Seader, Equilibrium-Stage Separation Operations in Chemical Engineering, Wiley (1981).

King, C.J., Separation Processes, Second Edition, McGraw Hill (1980).

Chapter 4. Equilibrium Columns

This chapter describes the equilibrium stage model for column operations such as distillation, absorption and extraction. The equations that ChemSep solves are discussed as well as other issues.

4.1 Introduction

Multicomponent separation processes like distillation, absorption and extraction have been modelled using the equilibrium stage concept for a century. The equilibrium stage model was first used by Sorel in 1893 to describe the rectification of alcohol. Since that time it has been applied with ever increasing frequency to all manner of separation processes: distillation (including rectification, stripping, simple (single feed, two product columns), complex (multiple feed, multiple product columns), extractive, azeotropic and petroleum refinery distillation), absorption, stripping, liquid-liquid and supercritical extraction.

The equations that model equilibrium stages are called the MESH equations. The MESH equations for the interior stages of a column together with equations for the reboiler and condenser (if they are needed) are solved together with any specification equations to yield, for each stage, the vapour mole fractions; the liquid mole fractions; the stage temperature and the vapour and liquid flowrates.

Since the late 1950's, hardly a year has gone by without the publication of at least one (and usually more than one) new algorithm for solving the equilibrium stage model equations. One of the incentives for the continued activity has always been (and remains) a desire to solve problems with which existing methods have trouble. The evolution of algorithms for solving the MESH equations has been influenced by, among other things: the availability (or lack) of sufficient computer storage and power, the development of mathematical techniques that can be exploited, the complexity of physical property (K-value and enthalpy) correlations and the form of the model equations being solved.

It is not completely clear who first implemented a simultaneous correction method for solving multicomponent distillation and absorption problems. As is so often the case, it would appear that the problem was being tackled by a number of people independently. Simultaneous solution of all the MESH equations was suggested as a method of last resort by Friday and Smith (1964) in a classic paper analysing the reasons why other algorithms fail. They did not, however, implement such a technique. The two best known and most frequently cited papers are those of Goldstein and Stanfield (1970) and Naphtali and Sandholm (1971), the latter providing more details of an application of Newton's method described by Naphtali at an AIChE meeting in May 1965.

To the best of our knowledge, a method to solve all the MESH equations for all stages at once using Newton's method was first implemented by Whitehouse (1964) (see, also, Stainthorp and Whitehouse, 1967). Among other things, Whitehouse's code allowed for specifications of purity, T, V, L or Q on any stage. Interlinked systems of columns and nonideal solutions could be dealt with even though no examples of the latter type were solved by Whitehouse. Since the pioneering work of Whitehouse, Naphtali and Sandholm and Goldstein and Stanfield, many others have employed Newton's method or one of its relatives to solve the MESH equations.

Simultaneous correction procedures have shown themselves to be generally fast and reliable. Extensions to the basic method to include complex column configurations, interlinked columns, nonstandard specifications and applications to column design result in only minor changes in the algorithm. In addition, simultaneous correction procedures can easily incorporate stage efficiencies within the calculations (something that is not always possible with other algorithms). Developments to about 1980 have been described in a number of textbooks (see, for example, Holland, 1963, 1975, 1981; King, 1980; Henley and Seader, 1981) and a recent review by Wang and Wang (1980). Seader (1985) has written an interesting history of equilibrium stage simulation.

Seader (1986) lists a number of things to be taken into consideration when designing a simultaneous correction method; a revised and extended list follows and is discussed in more detail below.

  1. What equations should be used?
  2. What variables should be used?
  3. How should the equations be ordered?
  4. How should the variables be ordered?
  5. How should the linearized equations be solved?
  6. Should the Jacobian be updated on each iteration or should it be held constant for a number of iterations or should it be approximated using quasi-Newton methods. Should derivatives of physical properties be retained in the calculation of the jacobian (J)?
  7. Should flexibility in specifications be provided and, if so, how?
  8. What criterion should we use to determine convergence?
  9. How should the initial guess be obtained?
  10. What techniques should we use to improve reliability?

4.2 Equations

Each equilibrium stage in the column has a vapour entering from the stage below and liquid from a stage above. They are brought into contact on the stage together with any fresh or recycle feeds. The vapour and liquid streams leaving the stage are assumed to be in equilibrium with each other. A complete separation process is modeled as a sequence of s of these equilibrium stages. Each stage can have optional sidedraws where part of the vapour or liquid stream leaving the stage is leaving the column.

The equations that model equilibrium stages are termed the MESH equations, MESH being an acronym referring to the different types of equations that form the mathematical model. The M equations are the Material balance equations, the E equations are the Equilibrium relations, the S equations are the Summation equations and the H equations are the entHalpy balances:

where we have vapour and liquid leaving flows from stage j, V_j and L_j, with mole fractions y_ij and x_ij, vapour and liquid sidedraws, W_j and U_j, Feeds F_j with mole fractions z_ij, equilibrium K-values K_ij, enthalpies H_j, and stage heat duty Q_j.

If we count the equations listed, we will find that there are 2c+4 equations per stage. However, only 2c+3 of these equations are independent. These independent equations are generally taken to be the c component mass balance equations, the c equilibrium relations, the enthalpy balance and two more equations. These two equations can be the two summation equations or the total mass balance and one of the summation equations (or an equivalent form). The 2c+3 unknown variables determined by the equations are the c vapour mole fractions, y; the c liquid mole fractions, x; the stage temperature, T and the vapour and liquid flowrates, V and L. For a column of s stages, we must solve s(2c+3) equations. The table below shows how we may easily end up having to solve hundreds or even thousands of equations.

c s s(2c+3)
2 10 70
3 20 180
5 50 650
10 30 690
40 100 8300

The first entry in this table corresponds to a simple binary problem that could easily be solved graphically. The second and third are fairly typical of the size of problem encountered in azeotropic and extractive distillation processes. The last two entries are typical of problems encountered in simulating hydrocarbon and petroleum mixture separation operations.

4.3 Condenser and Reboiler

The MESH equations can be applied as written to any of the interior stages of a column. In addition to these stages, the reboiler and condenser (if they are included) for the column must be considered. The MESH equations shown above may be used to model these stages exactly as you would any other stage in the column. For these special stages it is common to use some specification equation instead of the enthalpy balance. Typical specifications include:

  1. the flowrate of the distillate/bottoms product stream,
  2. the mole fraction of a given component in either the distillate or bottoms product stream,
  3. a component flow rate in either the distillate or bottoms product stream,
  4. a reflux/reboil ratio or rate,
  5. the temperature of the condenser or reboiler,
  6. a heat duty to the condenser or reboiler.
ChemSep includes all of these specifications as well as a few others that have not been listed.

In the case of a total condenser, the vapour phase compositions used in the calculation of the equilibrium relations and the summation equations are those that would be in equilibrium with the liquid stream that actually exists. That is for the total condenser, the vapour composition used in the equilibrium relations is the vapour composition determined during a bubble point calculation based on the actual pressure and liquid compositions found in the condenser. At the same time, these compositions are not used in the component mass balances since there is no vapour stream from a total condenser.

4.4 "Nonequilibrium" Stages

In actual operation the trays of a distillation column rarely, if ever, operate at equilibrium despite attempts to approach this condition by proper design and choice of operating conditions. The degree of separation is, in fact, determined as much by mass and energy transfer between the phases being contacted on a tray or within sections of a packed column as it is by thermodynamic equilibrium considerations. The usual way of dealing with departures from equilibrium in multistage towers is through the use of stage and/or overall efficiencies. The Murphree stage efficiency is most often used in separation process calculations because it is easily combined with the equilibrium relations:


         MV                        MV
  E   = E    K   x   - y   - (1 - E   ) y      = 0
   ij      j  ij  ij    ij           j   i,j+1

where E^MV_j is the Murphree vapour efficiency for stage j. If the efficiency is unity we obtain the original equilibrium relation from above. The Murphree efficiency must be specified for all stages except the condenser and reboiler which are assumed to operate at equilibrium.

4.5 Solution of the MESH Equations

Almost every one of the many numerical methods that have been devised for solving systems of nonlinear equations has been used to solve the MESH equations. However, as mentioned earlier, ChemSep uses mostly the Newton's method to solve the nonlinear algebraic MESH equations. Here we will discuss how ChemSep uses the Newton's method.

4.5.1 How to Order the Equations and Variables?

Separation problems in ChemSep result in stages with each a set of various types of equations. There are basically two ways to group the equations and variables: by type or by stage. Grouping the equations and variables by stage is preferred for problems with more stages than components (practically all distillation and many absorption and extraction problems) while grouping by type is preferred for systems with more components than stages (some gas absorption problems). ChemSep employs a by-stage grouping of the equations and variables. We define a vector of variables for the j-th stage as:


      T
  (X )  = (V , y  , y   ... y  , x  , x   ... x  , L )
    j       j   1j   2j      cj   1j   2j      cj   j

and a vector of functions for the j-th stage (F_j):


      T     T                                            L-V
  (F )  = (M  , M  , M   ... M  , H , E  , E   ... E  , S    )
    j        j   1j   2j      cj   j   1j   2j      cj      j

where


   L-V       c
  S    = sum      (x   - y  ) = 0
              i=1   ij    ij

If the equations and variables are grouped by stage we have


     T        T      T         T
  (F)  = ((F ) , (F )  ... (F )  )
            1      2         s


     T        T      T         T
  (X)  = ((X ) , (X )  ... (X )  )
            1      2         s

4.5.2 The Jacobian

To evaluate the Jacobian matrix, one must obtain the partial derivative of each function with respect to every variable. Part of the appeal of the grouping by stage approach is that for single columns at least the Jacobian matrix is block tridiagonal in structure:

[B1][C1]
[A2][B2][C2]
[A3][B3][C3]
. . .
. . .
[Am-1][Bm-1][Cm-1]
[Am] [Bm]

in which each entry [A], [B], [C] is a matrix in its own right. The [A] submatrices contain partial derivatives of the equations for the j-th stage with respect to the variables for stage j-1. The [B] submatrices contain partial derivatives of the equations for the j-th stage with respect to the variables for the j-th stage. Finally, the [C] submatrices contain partial derivatives of the equations for the j-th stage with respect to the variables for stage j+1. The structure of the submatrices [A], [B] and [C] is shown below.

The elements of [B] include partial derivatives of K-values with respect to temperature and composition. Since it is rather a painful experience to differentiate, for example, the UNIQUAC equations with respect to temperature and composition, in some SC codes this differentiation is done numerically. This can be an extremely time consuming step. However, neglect of these derivatives is not recommended unless one is dealing with nearly ideal solutions, since, to do so, will almost certainly lead to an increase in the required number of iterations or even to failure.

Almost all of the partial derivatives needed in ChemSep are computed from analytical expressions. The exception is the temperature derivatives of the excess enthalpy which requires a second differentiation with respect to temperature of the activity and fugacity coefficient models. Coding just the first derivatives was bad enough.

If the column has pumparounds extra matrices will be present which are not on the diagonal and the use of block tridiagonal methods becomes less straight forward. Similar problems arise with non-standard specifications that are not on the variables of the condenser and reboiler. When we solve multiple interlinked columns (currently not supported by ChemSep) a special ordering is required to maintain the diagonal structure of the jacobian. Therefore, ChemSep now uses a sparse solver to solve the system of equations involved. In the case of non-standard specifications (which allow "wild" specification such as the temperature on a stage in the middle of a column, instead at one of its ends) the specification equation must be calculated using finite differencing.

4.5.3 How Should the Linearized Equations be Solved?

It is absolutely essential to take account of the sparsity of the Jacobian when solving the linearized equations; straightforward matrix inversion is totally impractical (and probably numerically impossible). Linear systems with a block tridiagonal structure may reasonably efficiently be solved using a generalized form of the Thomas algorithm. The steps of this algorithm are given by Henley and Seader (1981).

Still further improvements in the block elimination algorithm for solution of separation process problems can be effected if we take advantage of the special structure of the submatrices [A], [B] and [C]. In fact, [A] and [C] are nearly empty. ChemSep uses the sparse matrix solver NSPIV for solving the sparse linear system effectively.

As ChemSep calculates the Jacobian mostly analytically, the Jacobian is computed for each iteration of the Newton's method. This has shown to be the most effective method until now. Alternatively, the Jacobian could be held constant for a number of iterations or approximated using quasi-Newton methods. ChemSep currently does not implement these options. Another possibility is to selectively retain the derivatives of physical properties in the calculation of the Jacobian. For dynamic simulations this has shown to be counterproductive (Kooijman, 1995) and therefore ChemSep doesn"t support such a technique.

4.5.4 The Convergence Criterion

Convergence of the Newton's method in ChemSep is determined from the residuals of the equations which make up the model. These function-values are squared and summed to give one convergence parameter: the sum of squares (SS). In ChemSep we compare the square root of this sum of squares with a user specified convergence criterion. Typical values for this convergence criterion vary between 10^-6 for equilibrium column and flash calculations to 10^-3 for nonequilibrium column calculations.

As ChemSep does not scale the equations it is possible that this sum of squares remains too large for such small convergence criterions. As this is not always forseeable, we extended the convergence test in ChemSep with a check on the change in the variables. If the absolute relative value of the change of each variable is smaller than the convergence criterion, ChemSep also considers the problem solved.

Finally, in order to prevent calculations to continue endlessly, there is a maximum number of iterations after which the Newton's method exits. ChemSep will then ask the user whether to continue the calculations, and for how many additional iterations. Typically, we found that 30 Newton iterations suffices to solve most separation problems at hand. Often, when more iterations are needed, there are other problems which prevent the method to converge. For example, specifications can be impossible to attain, or the specifications define a problem with multiple solutions, where the Newton's method oscillates between two or more solutions (such problems have only been found and documented recently).

An exception to this observation of convergence form the simulations of packed columns using the nonequilibrium with plug flow model approximations, where convergence tends to be slow.

4.5.5 The Initial Guess

In order to obtain convergence, Newton's method requires that reasonable initial estimates be provided for all s(2c+3) independent variables. ChemSep uses an automatic initialization procedure where the user does not need to make any guesses. Flow rates are estimated assuming constant molar flows from stage to stage. If the bottoms flow rate and reflux ratio are NOT specified and cannot be estimated from the specifications that are supplied then the bottoms flow rate is arbitrarily assigned a value of half the total feed flow and the reflux ratio is given a default value of 2. This, of course, could cause serious convergence problems. In the future, optional guesses might be added to the specifications to circumvent this problem.

The next step is to estimate the compositions and temperatures. This is done iteratively. We start by estimating the K-values assuming ideal solution behavior at the column average pressure and an estimate of the boiling point of the combined feeds. (This eliminates the normal requirement of estimates of end stage temperatures).

Mole fractions of both phases are estimated by solving the material balance and equilibrium equations for one component at a time.

We define a column matrix of discrepancy functions


     T      c           c           c
  (F)  = (ME    , ... ME    , ... ME     )
             i,1         i,j         i,s

where ME^c_ij is the component material balance equation combined with the equilibrium equations to eliminate the vapour phase mole fractions. Each equation depends on only three mole fractions. Thus, if we define a column matrix of mole fractions (X) by

     T
  (X)  = (x   , ... x     , x  , x     , x    )
           i,1       i,j-1   ij   i,j+1   i,s

we may write
  (F) = [ABC](X) - (R) = (0)
With the equations and variables ordered in this way, the coefficient matrix [ABC] has three adjacent diagonals with coefficients:
  A  = L   
   j    j-1

  B  = - (V  K j + L  )
   j       j  i     j

  C  = V  K j
   j    j  i

The right hand side matrix (R) has elements
  R  = - F  z  
   j      j  ij

This linear system of equations can be solved for the mole fractions very easily using Gaussian elimination. Temperatures and vapour compositions are computed from a bubble point calculation for each stage. The bubble point computation provides K-values for all components on all stages. So we solve the tridiagonal system of equations again using the old flow rates and the new K-values. Temperatures are recomputed as before. This procedure is repeated a third time before proceeding with the main simulation.

For columns without condenser and reboiler a different initialization is used where the compositions of the liquid are set equal to the top liquid feed compositions and the compositions of the vapour equal to the bottom vapour feed compositions. Temperatures in the whole column are set equal to the temperature of the first feed specified.

For columns with either a condenser or a reboiler the compositions are initialized to the overall compositions of all feeds combined, and the temperatures to the bubble point at the column pressure and the overall feed compositions.

Currently there is no special initialization routine that will handle pumparounds if they are present. The result is that columns with large pumparoundflows will require flow initialization by the user, or, to repetitively solve the problem using the old results as the initialization and increasing pumparoundflow (see below).

4.5.6 Reliability

SC methods are far more reliable and versatile than most other methods. The same method will solve distillation, gas absorption and liquid extraction problems. It must be admitted though, that although the probability that Newton's method will converge from the automatic initial estimates is quite high, there is no guarantee of convergence. The difficulty of supplying good initial estimates is particularly severe for problems involving strongly nonideal mixtures, interlinked systems of columns and nonstandard specifications.

Several methods have been used to improve the reliability of Newton's method; damping the Newton step, use of steepest descent (ascent) formulations for some of the iterations, and combination with relaxation procedures; none of which has proven to be completely satisfactory. The methods most recently proposed for assisting convergence of Newton's method are continuation methods.

In default mode, ChemSep does NOT use any of these techniques, other than a check to make sure that all quantities remain positive. Mole fractions, for example, are not permitted to take on negative values. The user does have the option of supplying damping factors.

4.5.7 Damping factors

The Newton's method in ChemSep has some extra features that will enhance the convergence to the solution. The Newton's method computes a new solution vector based on the current Jacobian and function vector. However, the new solution vector might be physically meaningless, for example if a composition becomes smaller than zero or larger than unity. Also, the new solution vector might represent a too large a change in stage temperatures or flows for the method to be stable. To eleviate these problems ChemSep uses damping of the newton's iteration changes. Temperature changes are limited to a maximum (default to 10 K) and flow changes up to a maximum fraction of the old flows (default set to 0.5). The compositions require a special type of damping. If a composition is becoming negative or larger than unity, the change is limited to half the distance to the extreme. Also, if a damping factor is specified, the maximum change in composition equals the factor (the default factor is 1 allowing a change over the whole mol fraction range). This type of damping has turned out to be very effective. The damping factors can be found under the Options - Solve options menu.

If for some reason your column simulation does not converge, changing the damping factors might help. If you know the iteration history (by either limiting the number of iterations or by printing out the intermediate answers. Both can be done in the Options - Solve options menu) you can adapt the factors so the column simula