See the deviation, not just the claim

Measured VLE at corpus scale — NIST TRC ThermoML Archive

378 binary pairs / 13,823 measured tie-lines, vendored in this repository. For every measured (T, x) the model is asked what vapour is in equilibrium with that liquid; the number is the average absolute deviation in vapour mole fraction.

Coverage qualifies the deviation and is reported next to it deliberately. NRTL is the most accurate here but only reaches the binaries it has fitted interaction parameters for; the cubics reach nearly every pair in the corpus at roughly twice the deviation. That trade is the result, not a footnote. The gate is on the median rather than per pair — a sweep of arbitrary binaries legitimately contains systems no general-purpose package fits.

Component attribution in the source archive is inferred, not read: the rendering mislabels which component a composition belongs to, so labels are resolved by a pure-component saturation-pressure tie-break that corrected 23% of datasets. The tie-break uses only pure-component data, independent of the mixture models scored here — but it is an inference, and it travels with these numbers.

Pure-component property path — measured against NIST ThermoML

1,490 components, 85,825 measured points. Scores the component databank's own liquid-property path: only 431 of 6,745 components carry a regressed DIPPR block, so for the rest every property is a corresponding-states estimate.

The DIPPR rows are a control, not a result: a regressed correlation scored against the data it was regressed from must come out tight, and that is what makes the estimated rows interpretable rather than merely asserted. Letsou-Stiel viscosity measures at roughly twice its own published error.

Still open: liquid density has no corresponding-states fallback, so 24,343 components carry criticals but cannot produce one. Vapour pressure had the same hole until Lee-Kesler was measured on this corpus and wired in.

Wax appearance temperature — measured against NIST ThermoML

37 liquidus curves and 619 measured points: C16–C44 n-paraffins in 11 solvents. Each point is a measured wax appearance temperature.

With fusion enthalpies alone, as a handbook table gives them, the mean error is 5.8 K (bias -5.8 K), against 1.3 K with every crystal transition.

Estimated solid data, checked by leaving each measured paraffin out of the fit: mean error 2.0 K.

Solid–liquid solubility — measured against NIST ThermoML

65 measured solubility curves, 565 points, over 41 solute/solvent pairs. Both of the solver's solubility models are scored on log10(x), because solubility here spans four decades.

How much each model says the solvent matters

The accuracy columns above are a dead heat, which is why they are not the headline. The ideal line contains no property of the solvent at all, so its spread across a solute's solvents is exactly zero decades — it cannot rank two solvents. Measurement spreads by one to four.

UNIFAC earns that where it counts — ibuprofen in water is 4.3 decades out on the ideal line and 0.9 on UNIFAC — but it is worse in the good solvents, where γ ≈ 1 makes the ideal line nearly exact, and it overshoots the size of the solvent effect (citric acid varies half a decade across its solvents; UNIFAC says five). Screening-grade for ranking a solvent, not for a yield.

Component attribution in the source rendering is inferred, not read: a curve is kept only when exactly one component melts above the measured range and the other below it, and only in the reading whose solubility rises with temperature (ΔH_fus > 0). That tie-break corrected 62% of the curves kept, so the labelling is this corpus's own inference.

CO2 transport — mixture density measured against NIST ThermoML

2,717 measured densities of CO2 with argon, co, ethane, h2s, methane, n2, oxygen, so2, in gas, liquid and supercritical states. Each package's own flash chooses the phase.

Component labels in the archive were settled against CoolProp's multiparameter model, so the gerg-2008 row checks this product's wrapper rather than GERG-2008 independently.

Hydrogen solubility — measured against NIST ThermoML

305 measured hydrogen mole fractions in the liquid, over 19 partners from methane at 100 K to squalane at 870 bar. They set the hydrogen binary parameters of both cubic packages.

The database rows these replace put k_ij = −0.5 on benzene, cyclohexane, n_decane, toluene, which overpredicts measured hydrogen solubility by 108–251%.

Measured partners use values fitted to their own points, so their error is in-sample. The held-out column is not: each hydrocarbon partner is scored with the value fitted to the others, which is what an unmeasured hydrocarbon receives.

MESH solver vs. the exact analytical solution

The production distillation solver run on a synthetic constant-α, constant-molal-overflow binary — the one case where McCabe-Thiele and Fenske are exact. The reference is a different algorithm (sequential stage march with shooting), so agreement tests the solver rather than restating it.

Rigorous column — measured solve time and residual

Real wall-clock timing of the Boston-Sullivan inside-out MESH solver at increasing stage counts, measured on this run.

Timed on the server that generated this report, so absolute times reflect its hardware and current load — the point is the trend with stage count and that every case converges, not the seconds.

Ideal-gas heat capacity vs. independent compilations (TRC, Poling)

5 correlation(s) found wrong beyond the threshold are superseded at runtime by the TRC coefficients and are therefore listed here rather than scored: ethyl_chloride, ethylene_carbonate, ethylene_glycol, nitrogen_trioxide, acrylic_acid.

Matched by CAS against the TRC gas tables and Poling, Prausnitz & O'Connell 5th ed. Appendix A, each evaluated only inside its own stated temperature range. Both populations are shown because a CoolProp cross-check can only see the shadowed ones — which are exactly the correlations the solver never evaluates. This tests transcription, units and equation dispatch against independent compilations of overlapping experimental literature; it is not an independent measurement.

Electrolyte activity coefficients (B-dot vs. Pitzer)

Saturation pressure (Antoine / EoS)

PC-SAFT — pure-component saturation & liquid density

Distillation column — stage count vs. Fenske minimum

Actual converged stages compared against the Fenske minimum-reflux stage count for the same separation; ratio must fall in the expected design range.

Azeotrope location

Predicted azeotrope composition and temperature compared against the cited experimental azeotrope for each system.

Liquid-liquid equilibrium (LLE)

Checks whether the thermo package predicts the same two-phase liquid split the cited data shows, and how much of each component crosses into the other phase.

Thermodynamic consistency (Gibbs-Duhem)

The Herington area-test D and J values for each VLE dataset; |D − J| within 10 flags the underlying data (not the model) as thermodynamically consistent.

APC controller solve-time scaling

Timed benchmark of the real MIMO DMC quadratic-program solver,flowsim.apc.mimo_dmc.dmc_move, at increasing controller sizes — measured, not projected.

RTO cycle-solve scaling

Timed benchmark of the real plant-wide multistart Ipopt NLP,flowsim.solver.rto.run_rto_schedule, at increasing multistart-restart counts — measured, not projected.