Differentiable Phase Envelopes and Inverse Mixture Design
Notebook
34_phase_envelope_peng_robinson.ipynb
traces the full phase envelope of a Peng–Robinson mixture with one
continuation run. It then refines the cricondenbar and cricondentherm as
augmented fold equations, differentiates both with respect to composition
and a binary interaction parameter, and designs a composition that hits a
target maxcondentherm. Every result is checked outside the solve that
produced it.
The reusable implementation is pounce.examples.phase_envelope, and
python/tests/test_phase_envelope_example.py is its regression test. The
numbers on this page come from the notebook’s saved output. That output is
stamped with POUNCE 0.10.0 at source commit fefa6ec3, JAX 0.9.2 on CPU.
Re-execute the notebook to restamp it.
This is the phase-envelope case of the applications epic gh#780 (child issue gh#773, implemented in #777).
The cumbersome baseline
A phase envelope is conventionally built by marching flash or saturation calculations in pressure or temperature, then restarting by hand at each turning point. Its design derivatives come from finite-difference retracing: perturb the composition, trace the whole curve again, and difference the extrema. Both steps fail in a characteristic way, and the notebook keeps both failures in:
| method | outcome on the 4-component gas | points |
|---|---|---|
pressure ladder, jp.solve in a loop | fails loudly at 98 bar; no solution exists past the cricondenbar | 36 |
PathFollower.follow in ln P | reports status="ok" on the trivial branch K_i = 1 | 9 |
PathFollower.trace_arclength | traces the whole envelope through both folds and the critical point | 323 |
The second row is the warning the notebook is built around. K_i = 1
makes the two phases identical, which satisfies the equilibrium equations
exactly at every (T, P). The follow run converges onto that branch to
a residual of 1e-15, and its KKT monitor cannot object because nothing
is wrong with the point. A residual says nothing about whether the answer
is physical. max|ln K| has to be watched explicitly away from the
critical neighbourhood.
Equations
The mixture is 80/10/6/4 methane/ethane/propane/n-butane with all
k_ij = 0 (NATURAL_GAS). The incipient-phase condition of Michelsen
(1980), at fixed vapour fraction beta, is
x_i = z_i / (1 - beta + beta K_i), y_i = K_i x_i
g_i = ln K_i + ln phi_i^V(y, T, P) - ln phi_i^L(x, T, P) = 0, i = 1..N
g_{N+1} = sum_i (y_i - x_i) = 0
The unknowns are X = [ln K_1 .. ln K_N, ln T], with theta = ln P as the
continuation parameter, which makes the system square, N + 1 by N + 1.
Swapping the roles, carrying ln P as a state and continuing in
ln T, gives an independent parameterization of the same curve.
POUNCE has no root() entry point and needs none. As in
Boundary Value Problems, the square system is posed as the
feasibility NLP min 0 s.t. F(X) = 0. For a nonsingular Jacobian the
interior-point step reduces to the Newton step.
The Peng–Robinson cubic is solved in closed (trigonometric Cardano) form
so that the residual stays one differentiable expression. Under
jax.jacobian both root branches are evaluated, and a NaN in the dead
branch would poison the live gradient, so each inactive branch is clamped
before its sqrt/arccos. The vapour takes the largest real root and the
liquid the smallest. That is a local root-selection convention, not a
tangent-plane-distance stability test.
The EOS is checked before anything is built on it:
Z -> 1asP -> 0;- gradients are finite in both the one-root and three-root regions;
- pure n-butane vapour pressure is 1.034 / 2.573 / 9.469 bar at 273.15 / 300 / 350 K, against literature values of about 1.03 / 2.58 / 9.42.
What the POUNCE workflow does
- Trace.
PathFollower(jp).trace_arclengthstarts from a Wilson K-value anchor at 1 bar. It walks the dew branch, passes the cricondenbar (97.59 bar) and the critical point (246.35 K, 92.16 bar, wheremax|ln K|reaches 0.004), and comes out on the bubble branch without being told to. Methane’sKinverts from 1.244 to 0.906 across the critical point. Theln Trun detects the cricondentherm (283.91 K). Each run’s turning point agrees with the plain maximum of the other run to 0.018 K and 0.002 bar. - Read the slope from the factor.
sensitivity_atreturnsd ln T / d ln Palong the boundary from the KKT factor. It agrees with a finite difference along the path to 1% relative and grows 122× toward the cricondenbar, where it diverges. The number the continuation needs and the number that blows up are the same number. This is why parameter continuation cannot pass the fold. - Refine the folds.
PathTrace.turning_pointsonly brackets a fold to the trace spacing.refine_foldsolves the augmented simple-fold system[F, F_X v, v^T v - 1] = 0with POUNCE, seeded from that bracket. Its residual ismax|G| ~ 1e-14. - Differentiate the extrema. The same implicit solve returns
d(extremum)/dqfor theN - 1log-ratio composition coordinateseta_i = ln(z_i / z_N)and onek_ij. - Design.
design_composition_for_extremumsolves for the smallest log-ratio change that moves the maxcondentherm to a target. Its design variables are unconstrained log ratios, so every iterate stays inside the simplex.
Validation independent of the solve
The four-component gas is a worked example, not a reference. The
independent benchmark is Deiters and Bell’s (2019) Peng–Robinson
methane/propane liquid isopleth at x_methane = 0.8, k_12 = 0.042823
(DEITERS_BELL_METHANE_PROPANE):
| check | result |
|---|---|
| published maxcondentherm 282.53 K | POUNCE refined fold 282.5253 K (Δ 0.0047 K) |
trace step size ds = 0.10 vs 0.08, after refinement | all four fold coordinates agree to ≤ 8.4e-12 |
implicit fold derivatives vs four full independent retraces (central differences in eta and k_12) | max relative error 1.07e-8 |
| inverse design to 280 K, then a fresh Wilson anchor, full retrace and independent fold refinement | 280.000000000 K; independent error 3.0e-11 K |
| physical guards at every refined fold (root admissibility, distance from the trivial branch) | pass; branch distance ≥ 0.77 |
retrograde condensation: isotherms between T_c and the cricondentherm | 3 of 3 cross the dew branch twice |
The derivative check deliberately pays for full retraces rather than
differentiating a second expression at the same fold. The design is
accepted only after a retrace that starts from scratch, so what is checked
is the designed mixture itself, not the optimizer’s own constraint
residual. The designed composition is (0.81107, 0.18893), from
(0.8, 0.2).
Reproduce it
From the repository root, with the Python development environment installed:
PYTHONPATH=python python -m pytest -q python/tests/test_phase_envelope_example.py
jupyter nbconvert --to notebook --execute --inplace \
python/notebooks/34_phase_envelope_peng_robinson.ipynb
The notebook’s last code cell asserts each row of the validation table
against a threshold rather than the printed value. Examples are the
published error < 0.02 K, the retrace relative error < 1e-5 and the
inverse-design error < 1e-6 K. A re-execution that drifts past a
threshold therefore fails instead of silently restamping.
Limitations
- This is an inspectable application example, not a production property package. Classical one-fluid Peng–Robinson with these mixing rules may be inadequate for strongly associating or polar mixtures.
- The largest/smallest-
Z, residual, normalization and trivial-branch guards are local checks. They do not replace a tangent-plane-distance phase-stability calculation. trace_arclengthrequires a scalar parameter and equality constraints. The augmented solve assumes a simple fold, and a higher-order singularity needs a different formulation. General branch switching is out of scope.- Both traces end with
status="corrector_failed". This happens at the cold end of the bubble branch, where the heavy components’Kruns past a few hundred and the corrector cannot take a fulldsstep; the envelope is complete well before that point. A terminal rejection is coverage evidence, not proof that every branch was found. - The inverse problem is nonconvex. The fresh retrace verifies the target and the physical branch. It does not certify that the composition change is globally minimal.
References
- D.-Y. Peng and D. B. Robinson, Ind. Eng. Chem. Fundam. 15 (1976) 59–64, doi:10.1021/i160057a011.
- M. L. Michelsen, Fluid Phase Equilibria 4 (1980) 1–10, doi:10.1016/0378-3812(80)80001-X.
- U. K. Deiters and I. H. Bell, AIChE Journal 65 (2019) e16730, doi:10.1002/aic.16730.
See also
- Path Following & Inverse Mapping:
follow,trace_arclength, and fold refinement. - Sensitivity Analysis: where
dx*/dthetacomes from. - Phase-Changing Flash as an MPCC: the same Peng–Robinson constants, solved across a phase change at fixed temperature.