Licensing and scientific-data provenance¶

This page records the release boundary for code, runtime parameter databases, validation data, and external-software results. It is a conservative engineering audit, not legal advice. Copyright, database rights, contract terms, and fair-use or quotation exceptions vary by jurisdiction; obtain qualified advice before publishing an artifact whose permission remains unclear.

Release boundary¶

Artifact class	Git repository	PyPI wheel and source distribution	Policy
`src/torch_flash/*/.py`	Yes	Yes	Original package implementation under LGPL-2.1-only
Runtime YAML under `src/torch_flash/data/`	Yes	Yes	Minimal model/component parameters, with primary scientific citations and applicable third-party notices
Cleared literature-derived CSV under `tests/data/`	Yes, subject to the rights audit below	No	Repository validation fixture with a recorded reusable basis
Unresolved research CSV under `tests/data/not-cleared/`	No	No	Local-only input; the entire subdirectory is ignored by Git
Independently generated software baselines under `tests/data/`	Yes	No	Project-generated numerical output; generator, software, version, and model inputs must be recorded
Notebooks, tests, scripts, and rendered outputs	Yes	No	Research/development material, not package runtime content
Books, papers, publisher supplements, and source workbooks	No	No	Never vendored; users obtain them from the cited publisher or archive

The build is checked in CI so that .ipynb, CSV, workbook, image, and PDF research artifacts cannot enter a Python distribution accidentally.

Current data-rights audit¶

The detailed per-file scientific provenance is in the repository's tests/data/README.md. The machine-readable tests/data/rights.yaml ledger covers all 33 declared CSV artifacts: 5 have a verified open basis, 12 are project-generated, and 16 are marked not-cleared. CI rejects an unclassified CSV and enforces that locally present not-cleared inputs are confined to the ignored tests/data/not-cleared/ subdirectory. Tests that consume those inputs skip in a public checkout where the directory is absent. The rights result does not change whether a file is experimental or model-generated.

Explicitly reusable sources¶

NIST ThermoML data are obtained from NIST's public ThermoML archive. The archive still requires the original articles to be cited and the values to be independently checked.
The H2ThermoBank archive, DOI 10.6084/m9.figshare.12063297, declares CC BY 4.0.
Moortgat's H2/H2O article and supporting table, DOI 10.1371/journal.pone.0332157, are CC BY 4.0.
The EOS-CG-2021 article and its supplementary coefficient tables, DOI 10.1007/s10765-023-03263-6, are CC BY 4.0 unless a third-party credit line says otherwise.
The MDEA article, experimental tables, and parameter supplement, DOI 10.1007/s10765-021-02933-7, are CC BY 4.0 unless a third-party credit line says otherwise.
The bundled original-UNIFAC table is derived from the MIT-licensed thermo data distribution. The scientific publications and DDBST page are cited as model provenance; the applicable software/data notice is retained in THIRD_PARTY_NOTICES.md.

Attribution, license links, and modification notices must remain with CC BY material. Public availability alone is not interpreted as an open license.

Project-generated numerical baselines¶

Files whose names identify ThermoPack, teqp, or NeqSim are independently generated outputs used for numerical comparison. They do not contain copied source code. The frozen-data manifest records the exact software version and model inputs or points to the generator. The third-party programs themselves are optional and are not bundled in torch-flash.

Redistribution permission not yet established¶

Several validation CSV files are transcriptions, normalized subsets, or visual digitizations of publisher-controlled articles, books, theses, or supporting workbooks. A DOI and full citation establish provenance, but no blanket redistribution license has been verified for these sources. They remain excluded from Git and every Python distribution under tests/data/not-cleared/. For a public repository archive or data release, each such file must be replaced by an openly licensed source, generated after a user supplies or downloads the source lawfully, covered by written permission, or kept outside the release.

The three normalized Jaubert-databank fixtures require particular caution: jaubert_2020_co2_binary_vle.csv, jaubert_2020_hv_bac5_vle.csv, and jaubert_ppr78_hydrocarbon_vle.csv. The ACS article is identified as CC BY-NC-ND 4.0, while no separate permissive license was found for the supporting workbook. The normalized subsets are adaptations and the package's license does not impose a noncommercial restriction. Consequently, this audit does not clear those three files for a public data release; obtain explicit permission or replace them with a compatibly licensed source first.

This unresolved group currently includes the Segovia density table, Jaubert binary/PPR78/Huron-Vidal subsets, Yücelen N2/CO2, Ahmadi and Wang CO2/H2O, Portier brine, Verschoyle H2/CO, Folas and Yan CPA tables/digitizations, Pedersen and Whitson book tables, and the Gernert-Span Table 8 transcription. Their scientific use in a private checkout is distinct from permission to redistribute the normalized tables publicly.

Executed notebook outputs that reproduce not-cleared values share the same unresolved publication risk. Excluding notebooks from PyPI prevents package redistribution but does not itself clear publication in a public Git repository.

The current documentation boundary publishes selected rendered plots, including observation markers, while withholding the paired local-only notebooks, source CSVs, machine-readable observation values, and notebook tables. A marker still graphically communicates an approximate observation coordinate; converting a table into a plot does not change the source dataset's license. Each published figure therefore declares its executed notebook checksum, source cell, inputs, evidence class, and publication form in docs/assets/validation/manifest.yaml. CI enforces the declared figure-only boundary and rejects tracked not-cleared research CSVs. This is a repository publication decision, not a determination that the source data are openly licensed or that an exception applies in every jurisdiction.

Runtime parameter databases¶

Published equations and numerical constants are cited at the file level. When a bundled coefficient inventory was transcribed or cross-checked with a licensed software database, the software/data license is also retained:

chemicals 1.5.2 (MIT) for selected shared component and ideal-gas data;
thermo 0.6.0 (MIT) for the original-UNIFAC parameter table;
Clapeyron.jl (MIT) for cross-checking the EOS-CG coefficient inventory; and
NIST teqp 0.23.2 (NIST notice) for cross-checking the GERG inventory.

Cross-checking identifies the independent numerical source; it does not mean that external implementation code was translated into this package. A 2026-07-24 audit compared whitespace-normalized sequences of six consecutive nonblank source lines against the locally inspected ThermoPack, NeqSim, teqp, and Clapeyron source trees. It found no exact implementation-block matches. That scan is evidence against obvious copying, not a proof of independent authorship or a substitute for contributor review.

Minimal parameter subsets transcribed from books and journal tables remain scientifically cited. Because compilation or database rights can apply even where individual numerical facts are not protected, larger future imports require an explicit license or permission before bundling.

Contribution checklist¶

For every new dataset or parameter file:

identify the primary scientific source and persistent identifier;
distinguish measurement, published-model output, fitted project output, and external-software output;
record units, transformations, selection rules, uncertainty, and generator;
record an explicit license or written permission for redistribution;
retain required copyright and license notices; and
keep the artifact outside package distributions unless it is required at runtime and cleared for redistribution.