Z-factor benchmark: 6726 experimental natural-gas compressibility factors with absolute inputs, and the NNCF neural surrogate

An open benchmark for the natural-gas compressibility factor z, together with a neural surrogate trained without any laboratory measurement. The benchmark. 6726 experimental gas-phase measurements extracted from 40 published studies in the NIST ThermoML archive, each carrying absolute pressure, temperature and full molar composition, so that z = pM/(ρRT) follows exactly from the measured quantities. No chart, correlation or equation of state enters that identity. The natural-gas domain used for the primary comparison contains 2426 points from 10 independent laboratories. This is the first common test set of measured z with absolute inputs: almost all prior machine-learning work on this property is trained and tested on values digitised from the 1942 Standing-Katz chart, the same chart to which the classical correlations were fitted. The model. NNCF, a residual multilayer perceptron of 899,329 parameters distilled from AGA8-DETAIL over 21.7 million synthetic states. It attains an average absolute deviation of 0.789 % on the benchmark, statistically indistinguishable from AGA8-DETAIL (0.769 %) and ahead of GERG-2008 (0.917 %), and reduces the error of the Dranchuk-Abou-Kassem, Hall-Yarborough and Dranchuk-Purvis-Robinson correlations by 1.10-1.14 percentage points at every one of the ten laboratories. Because no laboratory measurement enters training, the benchmark is a genuine held-out test. It ships with an applicability-domain guard that falls back to DAK outside its declared window. Also included. Seven verified classical correlations; laboratory-clustered bootstrap significance testing; a controlled experiment showing that surrogate accuracy is governed by the reference equation used to generate its training labels; a 186-fit leave-one-laboratory-out search over 31 residual-correction families; and a measured computational-cost benchmark. Attribution. The measurements are the work of the 40 constituent studies, listed with DOIs in data/processed/SOURCES.md. Please cite them as well as this compilation. Licensing. The derived data tables are released under CC BY 4.0; the code and the trained model are released under the MIT licence, as stated in LICENSE.