Method Queue

The order follows one principle: a method that can be checked against an open implementation (T2) comes before one that can only be built from its paper. Requests that come with a fixture and its expected output move up. See Adding a Method for the tiers and rules.

Next: methods with an open reference (T2)

#	Method	Paper	Reference to check against	Licence of reference
1	CHILL+ to T2 (already shipped at T1)	Nguyen & Molinero 2015, doi 10.1021/jp510289t	OVITO’s Chill+ modifier, including its gas-hydrate class	`ovito` pip module, MIT
2	Polyhedral template matching (PTM)	Larsen et al. 2016, doi 10.1088/0965-0393/24/5/055007	`pmla` reference code; OVITO	MIT
3	Ring statistics (King / Franzblau shortest-path rings)	King 1967, doi 10.1038/2131112a0; Franzblau 1991, doi 10.1103/PhysRevB.44.4925	matscipy ring statistics	LGPL-2.1
4	Topological cage building blocks (`genice2-cage`)	Matsumoto et al. 2007, doi 10.1063/1.2772627	genice-cage plugin of GenIce2	MIT
5	Unsupervised hydrate classification (DBSCAN on F3/F4 features, with a cavity finder)	Cai, Striolo & Salvalaglio, J. Phys. Chem. C 130 (2026) 3074	the authors’ repository `mme-ucl/clathrate_hydrate_classification`	GPL-3.0

Then: paper-only methods (T1 with crystal fixtures)

#	Method	Paper	Why it waits
6	Local structure index (LSI)	Shiratani & Sasai 1996, doi 10.1063/1.471475	No open reference; the definition is short, so the spec gaps are few.
7	Jacobson cage order parameters	Jacobson et al., doi 10.1021/ja1051445; doi 10.1021/jp107269q; doi 10.1063/1.3613667	Written for the coarse-grained mW model; carrying it over to atomistic water needs its spec gaps listed first.
8	MCG-1 and other MCG variants	Barnes et al. 2014, doi 10.1063/1.4871898	MCG-3 is shipped. The variants differ only in the neighbour count, so they mostly need fixtures.

Deferred

Method	Paper	Reason
ICA / FSICA	Guo et al. 2008, doi 10.1063/1.2919558; Guo et al. 2011, doi 10.1039/c1cp20070d	No public code, and the cage-search details the papers leave open exceed the two-gap limit.
ICO	Hao et al. 2021, doi 10.1021/acs.jpcb.0c08964	No public code, and its spec gaps are not yet listed.

Machine-learning methods: after the benchmark

No released model classifies hydrate cages in a form that can run in the browser.

GCIceNet (doi 10.1039/D0CP03456H) ships MIT code but no weights.
DeepIce has weights but no licence, so it cannot be redistributed.

The plan is to train a small model on labels produced by the verified methods above, using the benchmark crystals and deposited trajectories, and ship its weights in the repository with their training data named. It will carry its own level, “ML · trained here”, and will not be presented as a published method.