Rey.BEng 4th September 2026
Title:
Geometry Is the Memory
Coarse-Graining Confirms the Open Invitation
The Future Begins
Lin, Dahiya & Cersonsky (arXiv:2609.01911) put geometry back into coarse-grained machine-learned potentials. Isotropic beads lose direction. Energies, forces and torques then fail. Two complementary fixes restore what was thrown away: anisotropic density descriptors (AniSOAP) when the body is roughly ellipsoidal, and symmetry-adapted equivariant networks (MACE-CG) when the point group is arbitrary. Water is the sharp case. Ellipsoids are not enough. Orientational degeneracy must be resolved. Information loss is governed not only by mapping resolution but by the symmetry and geometry retained in the representation.
That is the laboratory statement of the invitation already filed, and seen in Neuroscience and Holographic Cosmology.:
Memory is not a second store bolted onto a scalar particle. Memory is the geometry that was kept.
What the paper records
- Coarse-graining that discards orientation predicts worse energies, forces and torques.
- Keep anisotropy: prediction improves.
- Water is the worst isotropic case; rigid-body orientation repairs it.
- Transferability tracks the geometric information that survives the map.
Geometric reading
A coarse grain that forgets direction is a Rest-Mass scalar: one number, no residual. The residual origin is
0^i2 (k.g.s^2) = r^2 m
State A — open residual, directional disc
E = 2c / h
State B — locked residual, compressive, −1/2 phase
E = hbar / c
Torque is the moment of the contact patch. Force without torque is a completed 2 with the twist deleted. Water already carries two local structures; those structures are orientations of the same residual, not two extra species. Drop the orientation and the model cannot tell HDL from LDL, disc from lock.
The 27-sphere ledger and the photonic synapse say the same thing at other scales: the 2D memory surface stores quality by contact-patch geometry; colour writes or erases modes on that surface. C. elegans stores aversion as parallel integrators — layers of geometric distortion in an elastic plenum, not addresses in a hard drive. Lin et al. show the engineering cost of refusing that ledger. If you coarse-grain geometry away, the potential forgets.
Catalogue entry
Discipline: Computational chemistry / coarse-grained ML potentials
Field: Geometry and directionality in CG representations
Observation:Lin, Dahiya & Cersonsky, arXiv:2609.01911
Canon reading: Information loss is residual loss. Keep direction and the memory remains. Future data memory is geometric because present matter already is.
Linked invitation pages:
geometric memory in an elastic plenum; T3 / 27-sphere ledger; photonic synapse on the π-tensor surface.
The observer shares r^2. The machine is the orientation that isotropic beads discarded.
Unity is already present.
The Superior Perspective is already proved.
The mechanical ontology of geometric memory is already derived.
The recursion holds.
The geometry continues to reveal itself.
Pirate Canon Sealed.
The Future Begins.
ace-consultancy.uk 