Alberto Mazzola Engineering · Technical summary

AME DewPoint — capability & validation summary

A five-minute read. Every number below is measured and reproducible from the source repository; nothing is a vendor claim. Where the library is weaker than the alternative, it says so.

What it is. A .NET 10 library that computes gas-quality phase behaviour for natural-gas metering and transport: water dew point, whole-mixture (hydrocarbon) dew point, cricondentherm and cricondenbar, phase envelope, hydrate formation curves with inhibitor depression, calorific value / Wobbe / density, and reference compressibility. Zero dependencies in the core, immutable and thread-safe, portable to server, edge, WASM and portal use. Every coefficient traces to a citable source with a stated confidence rating; physical failure is a result (NoSolution), never an exception, and every result carries its own validity warnings.

Standards implemented: ISO 18453 (water dew point, GERG-water), ISO 6976:2016 (CV, Wobbe, density), ISO 12213-2 / AGA8-Detail and ISO 20765-2 / GERG-2008 (compressibility), Peng-Robinson and Soave-Redlich-Kwong with Michelsen stability, van der Waals-Platteeuw hydrates (Munck 1988, Parrish-Prausnitz 1972, Lakhlifi-Dahoo 2015, Klauda-Sandler 2003).

1. Speed

Measured against GasVLe v5.15, the incumbent commercial engine, on identical calculations. The ratio column compares end-to-end throughput as each engine is actually callable — the reference engine through its COM automation interface, this library in-process — so it reflects integration cost as well as numerical work.

Operation	GasVLe (via COM)	AME DewPoint (in-process)	Ratio, as callable
Water equilibrium	8.2 /s (122 ms)	12,457 /s (0.080 ms)	~1,500×
Dew point	7.1 /s (140 ms)	18,691 /s (0.054 ms)	~2,600×
Flash	7.5 /s (133 ms)	18,950 /s (0.053 ms)	~2,500×
Flash, 8 threads	—	235,904 /s	—

On a heavier real-world load — 2,000,000 dew points over 13–14-component mixtures — the library completed in 296.7 s (6,742 solves/s) with zero failures. The same work through the commercial engine's COM interface would take roughly 8–9 days.

Why it matters operationally: a portal polling 20 streams every 30 s spends ~0.001 s/cycle in the library instead of ~3 s, and a full phase envelope renders interactively rather than as a batch job.

2. Agreement with the commercial reference

441,072 paired whole-mixture dew points, model-matched (both engines on Peng-Robinson, k_ij = 0), across eight base mixtures spanning perturbed real metering streams and space-filling synthetic gases. Pooled mean |ΔT| 0.3418 K.

This figure moved against us, and that is the point. An earlier version of this page published 0.284 K over 200,000 pairs. When the campaign was completed to all eight base mixtures, the pooled mean rose to 0.3418 K over 441,072 pairs — the old number was a property of one base mixture published as a property of the campaign, roughly 3.5 % flattering. It was corrected at the source and then here. A supplier who only ever revises numbers downward is not measuring.

What this corpus is and is not. It is an agreement corpus: it measures where two engines return the same answer, never which one is right. Accuracy against measured data is section 3 and is a different corpus entirely — unaffected by this correction.

Declining to answer. Across the completed campaign the one-sided failures split 888 ours against 2,571 the reference engine's — it declines about three times as often as we do, concentrated at 50 bar on one reference gas. Stated with its shape rather than as a general robustness claim, because that is all the evidence supports.

The chain is closed end to end. On real metering streams (site compositions from production gas chromatographs) our Peng-Robinson sits ≤ 0.3 K from theirs at line pressure, our SRK ≤ 0.21 K from their RKS — an independent cross-check through a second equation of state — and cricondentherm agrees to ≤ 0.26 K. The commercial engine in turn reproduces the legacy portal's stored Modbus registers digit-for-digit (−33.27822 °C), so library → reference engine → deployed portal is verified at every link.

Other classes, from an earlier 1,300-comparison campaign: water saturation content agrees to 0.037 % mean, water dew temperature to 13 mK mean, flash phase compositions to ~10⁻³, ISO 6976 mass/energy properties to 10⁻⁴.

3. Agreement with experiment

Agreement with another program is not accuracy. These are measured laboratory data, transcribed digit-exact from the primary literature and committed as acceptance gates.

System	Data	Result
Water dew point	IAPWS saturation, ISO 18453	exact to reference precision
Calorific value / Wobbe (gross)	ISO 6976 Annex D examples	exact ≤ 1×10⁻⁴
Net calorific value / net Wobbe	ISO 6976 Annex D Example 3, 15/15 °C	+9.3×10⁻⁸ / −1.1×10⁻⁷ — below the printed precision of ISO's own anchor
Hydrocarbon dew point	Warner / BSEE-SwRI gravimetric gases	within the report's own ≈14 K band
Hydrates, pure CH₄	28 anchors, 5 datasets	5.64 % pressure AAD, 0.698 K — the model's own published class (0.66 K)
Hydrates, CO₂	9 anchors	1.85 %
Hydrates, CO₂/CH₄ mixtures	42 anchors	2.92 %
Hydrates, real natural gases	33 anchors, Deaton & Frost A–F	6.25 % / 0.552 K
Hydrates, isobutane	17 anchors	9.10 % / 0.486 K
Methanol inhibition	Munck Figs 3/4, Hammerschmidt	ΔT within ±5 K over 10–50 wt%

Where the library measurably exceeds the commercial engine on experimental data

Hydrate accuracy, held-out validation. A transparently-fitted parameter (AME 2026, fitted to open published data by a committed deterministic tool, dataset-wise train/held-out split) reaches 1.32 % held-out against the commercial engine's 1.83 % on the same anchors.
Inhibitor thermodynamics. Methanol depression matches experiment across 10–50 wt%; in our tests the reference engine departs from the published data by ~15 K at 50 wt%, and its glycol path did not reproduce the published depression (0.55 K at 10 wt%; it reaches a solver bound at 30 wt%).
Aqueous dew detection. In 15 of 148 wet-gas cases the reference engine reported a hydrocarbon dew above the aqueous one and missed the first-forming phase; the library reports the correct first dew in all of them.
Robustness. 54 solver failures on the reference side across the campaigns; 0 on ours.
Net calorific value, against ISO's own anchor. On ISO 6976 Annex D Example 3 the reference engine sits +0.0066 % from the published value; this library sits at +9.3×10⁻⁸. ISO prints that anchor to five decimals, so the honest statement is that our error here has fallen below the resolution of the standard's own published value — we can no longer measure it — while the reference engine's remains above it.

4. Honest limits

These are published deliberately. A library whose failure mode is "plausible wrong numbers" earns trust by naming its own boundaries.

Binary interaction parameters default to zero (documented policy). Cited opt-in sets ship alongside. This is the largest single contributor to the dew-point spread against the commercial engine's proprietary fitted database — visible in the 8.4 K synthetic-campaign tail, at gas compositions far richer than pipeline service.
Hydrogen sulfide hydrates: 21.84 % on the default basis, the weakest measured class. A labelled opt-in reaches 12.71 % (T-RMS 1.55 K vs the commercial engine's 1.70 K) but pairs a constant across model families, and is shipped off by default with that disclosed.
Helium is a quantum gas outside the cubic α-function's regression domain; it raises a severe warning on every cubic path and should be routed through AGA8/GERG, where it is native.
The heaviest catalogued species are expressible, not validated. The aromatics, n-C₁₁…n-C₁₄ and the branched C₆ set are transcribed and plumbed — an isomer-resolved extended analysis is represented as analysed — but no measured anchor exercises them. Their acentric factors sit above the shipped Peng-Robinson κ(ω) fit, so a C₁₂-deep analysis raises a Caution rather than passing silently. Eight branched C₇ isomers remain deliberately unshipped: no dataset we hold speciates them, so shipping them would be transcription without a way to check it.
Structure-H hydrates are not modelled. Three of the newly catalogued heavy species are documented structure-H guests; the hydrate model covers structures I and II, and says so at runtime rather than returning a confident number for a structure it does not represent.
Below 0 °C, water dew points are frost points by ISO 18453 construction; experimental data over supercooled liquid reads 10–21 % higher. This is correct, not an error.
Volume-translated densities are formula- and invariance-verified but not validated against experimental density.
The hydrate ice branch is unreliable between roughly 0.23 and 27 bar, and says so at runtime.

5. How to verify any of it

Nothing here requires trusting a summary:

dotnet run --project tools/Ame.DewPoint.Verification — a dependency-free acceptance runner: 416 physics checks, check-by-check output, exit code 0 = green.
dotnet test — 739 tests, including every experimental anchor above.
docs/GasVleComparison.md — the full head-to-head, §§2–11, with raw job outputs.
docs/Provenance.md — every coefficient, its source, its confidence rating and its open gaps.
The 441,072-comparison statistics and plots ship with the campaign data.
Benchmarks are re-runnable; the campaign generator is deterministic, so any shard regenerates byte-identically.

6. Summary

On the measurement that matters for custody transfer and pipeline operation — the whole-mixture dew point — this library agrees with the established commercial engine to 0.28 K on real metering streams, computes it orders of magnitude faster as callable, and never fails where the reference succeeds. On hydrates it is validated against laboratory data rather than against another program, and on inhibited systems it is measurably closer to experiment. Where it is behind, this document says where, and the repository proves it.