CHILL+ versus cages

Two classifiers live in libyodaLib. They answer different questions and they do not imply each other.

Four-neighbour star

CHILL (seams chill, chill::getCorrel / getIceTypeNoPrint) and CHILL+ (seams chill-plus, chill::getCorrelPlus / getIceTypePlusNoPrint) look at the four-neighbour star. Both stay on the host. A crystal label needs exactly four neighbours inside the cutoff. That shell count is taken before the bond list is cut to four. The old count after the cut was a bug: a shell of five or more neighbours could still receive an ice label. q3 is then the average over the four nearest neighbours of a four-neighbour shell. The staggered and eclipsed pattern of that star labels the atom cubic, hexagonal, interfacial, clathrate, or water.

A molecule is cubic or hexagonal when its four bonds look like ice, even if those neighbours do not close a cage.

CHILL+: Nguyen and Molinero, J. Phys. Chem. B 2015 (10.1021/jp510289t). CHILL: Moore et al., Phys. Chem. Chem. Phys. 2010.

Cage membership

The ice score (seams cages, ring::seededCageAffiliation by default) is cage membership. A molecule in a hexagonal cage (HC) is ice Ih. A molecule in a double-diamond cage (DDC) is ice Ic. A molecule in neither is water. The cages are assembled from six-membered primitive rings on a chosen bond graph.

This is topological unit matching (TUM). The methods paper is the 2020 JCIM article (see Citation).

The GPU path (gpu::analyzeResident) is that TUM ice score only: linkcell::gpu k-nearest, the mutual four-nearest graph, primitive six-rings, and HC/DDC affiliation. CHILL, CHILL+, and \(q_{lm}\) stay on the host. Only cage labels come back.

Why they disagree

A configuration can agree exactly on the closed star of a vertex (same induced graph, same coordinates on those five atoms) and still disagree on that vertex’s cage label, because the cage predicates read a larger neighbourhood of six-rings. The Catch2 suite encodes that fact: cage membership is not a function of the four-neighbour star.

On the in-tree mixed TIP4P dump (input/traj/exampleTraj.lammpstrj), seams chill-plus uses the cutoff list by default and prints 9 interfacial clathrate and 241 water; --graph mutual prints 1 clathrate, 7 interfacial clathrate, and 242 water. Seeded cages report zero HC and zero DDC. A three-eclipsed star is not cage membership. The commands are in the mixed TIP4P tutorial.

Bond graph

seams cages --graph chooses the bond graph those rings are found on:

value

graph

cutoff

pairs inside the distance cutoff

knn

mutual k-nearest (default k=4)

knn-union

union k-nearest (completion graph of the seeded rule)

seeded

mutual seeds, union completions (CLI default)

On a perfect cubic crystal the four graphs coincide. On a mixed nucleation dump they do not have to.

Ring-adjacent completion

ring::seededCageAffiliation takes a fifth argument, ringAdjacentCompletion, default false. When the flag is true, the accepted labels pass through ring::ringAdjacentCompletion: a permissive six-ring with all but one vertex labelled fills that last vertex, iterated to a fixed point. Hexagonal cage (HC) flags and double-diamond cage (DDC) flags run as two walks.

seams cages --complete sets the flag. The rule, the least-fixed-point claim, and the Lean build live on Ring-adjacent completion.