Stability analysis

Stability analysis

A homogeneous mixture can satisfy the equation of state and still lower its Gibbs energy by separating into phases of different compositions. Stability analysis asks whether creating a small amount of a trial phase would help. It is the appropriate first question when you do not know whether to run a flash, and can supply a useful initial composition for that flash.

This is not the same as selecting a "stable" volume root at a fixed composition. Nor does a negative result tell you the final amount of each phase: a flash solves that material balance.

yaeos provides two routines:

  • tm: the reduced tangent-plane distance of a trial phase with a given composition.
  • min_tpd: search from several starting compositions and return the lowest trial-phase distance found.

Tangent-plane distance

The tangent plane describes the chemical potentials of the homogeneous feed. A trial phase below that plane can reduce the total Gibbs energy. Let be normalized feed mole fractions, pressure in bar, temperature in K, and positive trial-phase mole amounts. tm evaluates the modified, dimensionless distance:

where

Here is a fugacity coefficient and is the feed’s reduced chemical-potential term. When , this reduces to the ordinary normalized tangent-plane distance .

A significantly negative value establishes instability of the homogeneous reference within the model. A nonnegative numerical search result means no destabilizing phase was found, not proof of a global minimum. The trivial trial w=z has zero distance even far from a phase boundary.

Fragment for an initialized model (place declarations before executable code):

use yaeos, only: pr, tm

real(pr) :: z(2)
real(pr) :: w(2)
real(pr) :: P
real(pr) :: T
real(pr) :: tpd

z = [0.4_pr, 0.6_pr]
w = [0.8_pr, 0.2_pr]
P = 50.0_pr
T = 300.0_pr

tpd = tm(model, z, w, P, T)

if (tpd < 0.0_pr) then
   print *, "The feed is unstable"
end if

tm accepts GeModel objects as well as ArModel objects. For a GeModel the activity coefficients replace the fugacity coefficients, and pressure has no effect on the model.

The optional d argument allows the caller to provide precomputed values of . This avoids recalculating the feed fugacity coefficients:

Fragment for an ArModel; also declare Vz, lnphi_z(size(z)), and d(size(z)):

call model%lnphi_pt(z, T=T, P=P, V=Vz, root_type="stable", lnPhi=lnphi_z)
d = log(z) + lnphi_z

tpd = tm(model, z, w, P, T, d=d)

Finding the minimum

min_tpd minimizes with respect to the trial-phase mole numbers:

subroutine min_tpd(model, z, P, T, mintpd, w, all_minima)
   class(BaseModel), target :: model
   real(pr), intent(in) :: z(:)
   real(pr), intent(in) :: P
   real(pr), intent(in) :: T
   real(pr), intent(out) :: w(:)
   real(pr), intent(out) :: mintpd
   real(pr), optional, intent(out) :: all_minima(:, :)
end subroutine min_tpd

The routine returns:

Output Meaning
mintpd Minimum tangent-plane distance
w Trial-phase composition at the minimum
all_minima Array of shape (nc,nc+1): each row holds nc composition entries followed by its distance

min_tpd starts once per component with that component enriched in trial mole amounts, then uses successive substitution. It normalizes each resulting trial and reports the lowest normalized distance. all_minima exposes those trials, including possible duplicates; they are not guaranteed distinct or fully converged minima. There is no returned convergence flag. An (nc,nc) array is too small: the implementation writes the distance in column nc+1.

program stability_example
   use yaeos, only: pr, CubicEoS, PengRobinson76, min_tpd
   use, intrinsic :: ieee_arithmetic, only: ieee_is_finite
   implicit none

   type(CubicEoS) :: model
   real(pr) :: tc(2)
   real(pr) :: pc(2)
   real(pr) :: w(2)
   real(pr) :: z(2)
   real(pr) :: P
   real(pr) :: T
   real(pr) :: mintpd
   real(pr) :: wtrial(2)
   real(pr) :: minima(2, 3)

   tc = [190.564_pr, 425.12_pr]
   pc = [45.99_pr, 37.96_pr]
   w = [0.0115478_pr, 0.200164_pr]

   model = PengRobinson76(tc, pc, w)

   z = [0.4_pr, 0.6_pr]
   P = 50.0_pr
   T = 300.0_pr

   call min_tpd(model, z, P=P, T=T, mintpd=mintpd, w=wtrial, all_minima=minima)
   if (.not. ieee_is_finite(mintpd)) error stop "Nonfinite stability result"
   if (.not. all(ieee_is_finite(wtrial))) error stop "Nonfinite trial phase"

   print *, "Minimum TPD: ", mintpd
   print *, "Most unstable trial phase: ", wtrial

   if (mintpd < -1.0e-6_pr) then
      print *, "The feed is unstable: a flash should be performed"
   else
      print *, "No significant negative distance found in this search"
   end if
end program stability_example

Using stability analysis to initialize a flash

The normalized w is a trial phase, not necessarily a final equilibrium composition. The following fragment uses it as the initial y estimate relative to feed z as x. It requires strictly positive z; reverse the phase roles (k0=z/w) when necessary for the desired heavy/light assignment. Add imports for flash and EquilibriumState, declare result and integer iters first:

call min_tpd(model, z, P=P, T=T, mintpd=mintpd, w=w)

if (mintpd < -1.0e-6_pr) then
   result = flash(model, z, T, p_spec=P, k0=w/z, iters=iters)
end if

This is especially useful for:

  • Liquid-liquid equilibria, where the Wilson correlation may be a poor initial guess.
  • Systems near a critical point, where the two phases have similar compositions.
  • Activity-coefficient models, where the initial K-factors are mandatory.

Multiphase calculations

pt_mp_flash uses stability analysis internally: it tests the feed, solves a two-phase flash, and then tests the resulting phase compositions, adding one more phase if the reference phase is still unstable. It does not iterate until global stability is proved, and its no-split path is not usable. The process, and its limitations, are described on the multiphase flash page.

Practical notes

  • A truly globally stable feed is not metastable. A feed reported as stable by a finite search may be metastable if a lower-energy split was missed. Inspect the trials and compare independent initializations/model data when phase behavior is complicated.
  • For an ArModel, the trial phase is evaluated with the stable volume root. For liquid-liquid calculations, a separate check with the liquid root may be necessary.
  • The distance is dimensionless. The example’s 1e-6 threshold is an illustrative numerical screen, not an adjustable solver tolerance or a universal physical threshold. The routine’s own stopping tests can limit resolution near zero.
  • Near-zero values may mean a boundary, criticality, convergence to the trivial feed trial, or numerical noise. Check trial/feed differences and neighboring states before interpreting them.
  • Feed and trial logarithms require positive fractions. Remove absent components consistently from the model rather than passing exact zeros to this implementation.
  • min_tpd is also available for GeModel objects. In that case the pressure is only used for the calling convention and does not enter the activity coefficient model.

Next steps and reference

Use an unstable trial to initialize a flash, then verify the split’s fugacity and balance residuals. A negative stability test on either returned phase can signal a missing phase; see the limitations on the multiphase page.

Michelsen and Mollerup, Thermodynamic Models: Fundamental and Computational Aspects, Tie-Line Publications (2004), is the formulation referenced by the implementation (review DOI).