TriForces: Triadic Structures in Graphs, Magnetism & MLIPs

• TriForces is a term describing irreducible tri-part organization across disciplines, including forcing triples in graphs, triforce magnetic order, and three-stream augmentations in MLIPs.
• In graph theory, TriForces denote minimal three-element forcing families that ensure quasirandomness through combinatorial constructions without relying on any forcing subpairs.
• In condensed-matter physics and atomistic MLIPs, TriForces separate complex interactions by isolating dipole–quadrupole lock-ins and disentangling latent composition, structure, and interaction information.

TriForces is a nonstandard label applied to several distinct three-part constructions in recent research. In graph theory it naturally designates forcing triples—3-element forcing families in the Chung–Graham–Wilson quasirandom-graph framework, including explicit forcing triples with no forcing subpairs (Spasić, 2023). In condensed-matter physics it names a specific triple-$\mathcal Q$ partial magnetic order in ${\rm UNi_4B}$ stabilized by quadrupolar interactions (Ishitobi et al., 2022). In machine learning for materials it denotes a model-agnostic three-stream augmentation for atomistic GNNs that separates composition, structure, and interaction information (Ramlaoui et al., 20 May 2026). The available literature also uses the same triadic motif, sometimes explicitly and sometimes interpretively, in triangle-based quasirandomness, three-nucleon forces, tripartite entanglement, triangle-center theory, flavour gauge extensions, triality-based generation models, and triangle-free triple systems.

1. Terminological scope

The literature associates the label with several unrelated technical objects rather than a single unified definition.

Context	Meaning	Representative paper
Quasirandom graphs	forcing triples with no forcing pairs	(Spasić, 2023)
UNi$_4$B magnetism	triple-$\mathcal Q$ partial magnetic order called “triforce” order	(Ishitobi et al., 2022)
Atomistic MLIPs	three-stream augmentation for transferable representations	(Ramlaoui et al., 20 May 2026)

A plausible implication is that “TriForces” functions primarily as a descriptor of triadic organization. In the graph-theoretic case, the triad is a minimal forcing family whose every 2-element subfamily fails to force quasirandomness. In the magnetic case, the triad is a particular triple-$\mathcal Q$ real-space order. In the machine-learning case, it is an explicit factorization of latent information into three streams. Other occurrences preserve the same three-part logic but refer to different mathematical or physical objects.

2. Forcing triples in quasirandom graph theory

Within the Chung–Graham–Wilson framework, a family $\mathcal H$ of graphs is forcing if the conditions

$$\operatorname{hom}(H,G_n)=\bigl(p^{|E(H)|}+o(1)\bigr)n^{|V(H)|}\qquad \forall H\in\mathcal H$$

already imply quasirandomness of the graph sequence $\{G_n\}$, equivalently $t(H)=p^{|E(H)|}$ for all $H\in\mathcal H$ forces property ${\rm UNi_4B}$0 (Spasić, 2023). The classical example is ${\rm UNi_4B}$1, and the paper “A note on forcing triples with no forcing pairs” establishes the existence of genuinely 3-element forcing families such that none of their 2-element subfamilies is forcing.

The main result is Theorem 1.1: ${\rm UNi_4B}$2 More explicitly, for any fixed connected non-bipartite graph ${\rm UNi_4B}$3 with distinguished vertex ${\rm UNi_4B}$4 and ${\rm UNi_4B}$5, the paper defines

${\rm UNi_4B}$6

${\rm UNi_4B}$7

${\rm UNi_4B}$8

and proves that each ${\rm UNi_4B}$9 is forcing while no two graphs in $_4$0 form a forcing pair (Spasić, 2023).

The constructions use labeled-vertex operations $_4$1 and doubling $_4$2. The forcing direction is obtained through Jensen- and Cauchy–Schwarz-type inequalities that convert homomorphism counts for doubled graphs into constraints on intermediate conditional counts, ultimately forcing the correct edge density $_4$3. Once the edge density is pinned down, classical forcing pairs such as $_4$4 or $_4$5 imply quasirandomness. The non-forcing direction uses the Lovász–Szegedy finite graphon model and two nonconstant 2-vertex weight patterns, together with the criterion $_4$6, to construct non-quasirandom sequences on which any chosen 2-element subfamily has matching density parameter $_4$7 (Spasić, 2023).

Among the three families, $_4$8 has two structural features emphasized in the paper: all three graphs have the same chromatic number as $_4$9, and the number of vertices in each graph is linear in $\mathcal Q$0. This contrasts with earlier forcing-pair constructions in which the auxiliary graph $\mathcal Q$1 can be exponential in $\mathcal Q$2. In this precise graph-theoretic sense, TriForces are minimal 3-element forcing families whose forcing power is irreducibly ternary.

3. “Triforce order” in $\mathcal Q$3

In $\mathcal Q$4, “triforce order” denotes a specific triple-$\mathcal Q$5 partial magnetic order on the U triangular lattice, obtained in a localized pseudo-triplet crystalline-electric-field model with quadrupole degrees of freedom (Ishitobi et al., 2022). The ordered state is proposed as an alternative to an earlier toroidal triple-$\mathcal Q$6 order.

Within a $\mathcal Q$7 magnetic unit cell containing 9 U sites, 6 of 9 sites carry finite in-plane magnetic dipole moments, while the remaining 3 of 9 sites are magnetically disordered but host ordered quadrupole moments. The six magnetic sites form two interpenetrating triangles, one small inverted triangle and one larger triangle, and the in-plane spins on each triangle form a $\mathcal Q$8 non-collinear configuration. The authors name this pattern “triforce” because the arrangement resembles three mutually touching triangles (Ishitobi et al., 2022).

The magnetic order is described by

$\mathcal Q$9

with $\mathcal Q$0. For the triforce domain discussed in the paper,

$\mathcal Q$1

The quadrupole field

$\mathcal Q$2

uses

$\mathcal Q$3

together with $\mathcal Q$4-point components that generate quadrupole order both on the six magnetic sites and on the three magnetically disordered sites (Ishitobi et al., 2022).

The paper stresses that triforce and toroidal triple-$\mathcal Q$5 states have identical spin structure factors at the primary wave vectors $\mathcal Q$6. The amplitudes $\mathcal Q$7 are the same, while the phase factors differ. Accordingly, neutron diffraction at the primary Bragg vectors does not distinguish the two states. What changes is the real-space phase structure and the quadrupolar pattern. In the toroidal state, the three core sites are disordered both magnetically and quadrupolarly at mean field; in the triforce state they are magnetically disordered but quadrupolarly ordered (Ishitobi et al., 2022).

The stabilization mechanism is analyzed by a Landau expansion in dipole and quadrupole fields, with a local cubic coupling

$\mathcal Q$8

which acts as a dipole–quadrupole lock-in term. The relative stability of single-$\mathcal Q$9, toroidal triple-$\mathcal H$0, and triforce triple-$\mathcal H$1 states is governed by negative quartic corrections obtained by integrating out quadrupolar modes at $\mathcal H$2, $\mathcal H$3, and $\mathcal H$4. For $\mathcal H$5, the phase-shifted triple-$\mathcal H$6(2) family, which includes the triforce pattern, is favored (Ishitobi et al., 2022).

The triforce state also has a distinct cluster-multipole decomposition. The magnetic sector contains $\mathcal H$7 and $\mathcal H$8 components, while the electric sector contains $\mathcal H$9 and $$\operatorname{hom}(H,G_n)=\bigl(p^{|E(H)|}+o(1)\bigr)n^{|V(H)|}\qquad \forall H\in\mathcal H$$0. When realistic orthorhombic distortion and site-dependent crystalline-electric-field effects are included, the resulting canted triforce order is consistent with the observed current-induced magnetization in $$\operatorname{hom}(H,G_n)=\bigl(p^{|E(H)|}+o(1)\bigr)n^{|V(H)|}\qquad \forall H\in\mathcal H$$1, including the anisotropy problem that motivated re-examination of the pure toroidal interpretation (Ishitobi et al., 2022).

4. TriForces as a three-stream framework for atomistic GNNs

In atomistic machine learning, TriForces is a model-agnostic augmentation for machine-learning interatomic potentials that separates composition and structure information and combines this separation with self-supervised learning to preserve transferable representations (Ramlaoui et al., 20 May 2026). The central node representation is

$$\operatorname{hom}(H,G_n)=\bigl(p^{|E(H)|}+o(1)\bigr)n^{|V(H)|}\qquad \forall H\in\mathcal H$$2

where the three streams correspond to composition, structure, and interaction.

The composition stream is coordinate-blind. A structure is compressed into unique element tokens $$\operatorname{hom}(H,G_n)=\bigl(p^{|E(H)|}+o(1)\bigr)n^{|V(H)|}\qquad \forall H\in\mathcal H$$3, with learned element embeddings and a count-weighted Transformer. For head $$\operatorname{hom}(H,G_n)=\bigl(p^{|E(H)|}+o(1)\bigr)n^{|V(H)|}\qquad \forall H\in\mathcal H$$4,

$$\operatorname{hom}(H,G_n)=\bigl(p^{|E(H)|}+o(1)\bigr)n^{|V(H)|}\qquad \forall H\in\mathcal H$$5

The $$\operatorname{hom}(H,G_n)=\bigl(p^{|E(H)|}+o(1)\bigr)n^{|V(H)|}\qquad \forall H\in\mathcal H$$6 bias makes attention over unique tokens exactly equivalent to attention over a multiset in which token $$\operatorname{hom}(H,G_n)=\bigl(p^{|E(H)|}+o(1)\bigr)n^{|V(H)|}\qquad \forall H\in\mathcal H$$7 is repeated $$\operatorname{hom}(H,G_n)=\bigl(p^{|E(H)|}+o(1)\bigr)n^{|V(H)|}\qquad \forall H\in\mathcal H$$8 times (Ramlaoui et al., 20 May 2026).

The structural stream is type-agnostic and rotation-invariant. It builds a SOAP-like local density from displacements $$\operatorname{hom}(H,G_n)=\bigl(p^{|E(H)|}+o(1)\bigr)n^{|V(H)|}\qquad \forall H\in\mathcal H$$9, radial basis functions, real spherical harmonics, and multi-scale cutoffs, forming coefficients

$\{G_n\}$0

and the power spectrum

$\{G_n\}$1

A small invariant message-passing stack then produces $\{G_n\}$2. The interaction stream is the original MLIP backbone, such as Orb-v3, eSEN, or MACE, operating on both species and positions (Ramlaoui et al., 20 May 2026).

Self-supervised pretraining combines three losses: $\{G_n\}$3 The denoising term predicts Gaussian position noise; the masking term reconstructs masked atom types; and the LeJEPA-style latent objective aligns graph- and node-level embeddings across augmented views while using SIGReg regularization to avoid collapse. The paper argues that the architectural separation is the main source of gains when abundant supervised data is available, whereas self-supervision becomes especially important in low-data transfer and for retrieval quality (Ramlaoui et al., 20 May 2026).

The reported empirical gains are substantial. On OMat24 in the limited-data regime, TriForces reduces energy MAE by 57% at 20K samples only, from 81.3 to 34.6 meV/atom in the cited eSEN direct setting, and improves force MAE from 151.4 to 121.8 meV/$\{G_n\}$4 (Ramlaoui et al., 20 May 2026). On full OMat24, Orb-v3 conservative improves from 107 meV/atom and 150 meV/$\{G_n\}$5 to 19.4 meV/atom and 95.5 meV/$\{G_n\}$6, while eSEN conservative improves from 80.3 meV/atom and 84.2 meV/$\{G_n\}$7 to 18.8 meV/atom and 78.0 meV/$\{G_n\}$8. On MatBench, TriForces variants achieve best or near-best results on 6 out of 8 tasks, and on QM9 the molecularly pretrained variants improve standard targets such as dipole moment, polarizability, and HOMO energy (Ramlaoui et al., 20 May 2026).

The latent-space analysis is equally central. Frozen TriForces embeddings support crystal-system accuracy of 96–100%, majority-element prediction near 100%, and mean nearest-neighbor distance MAE of 0.05–0.08 $\{G_n\}$9, whereas the cited baseline MLIPs obtain roughly 55–73%, 61–62%, and 0.75–1.10 $t(H)=p^{|E(H)|}$0, respectively. Separate retrieval in composition and structure space also behaves as intended: the composition stream yields high element-set recall, while the structure stream yields the best space-group recall (Ramlaoui et al., 20 May 2026). In this usage, TriForces denotes an explicitly factorized latent representation rather than a physical force.

5. Other research uses of the triadic motif

In quasirandom-graph theory, triangle-based forcing provides a related but distinct use of the motif. Reiher and Schacht proved that $t(H)=p^{|E(H)|}$1 is a forcing pair and that if $t(H)=p^{|E(H)|}$2 is forcing, then $t(H)=p^{|E(H)|}$3 is also forcing, where $t(H)=p^{|E(H)|}$4 is obtained by replacing every edge of $t(H)=p^{|E(H)|}$5 by a triangle with a fresh vertex. This establishes triangles as effective forcing objects and systematically transfers edge-based forcing pairs into triangle-based ones (Reiher et al., 2017).

In nuclear theory, “TriForces” has been used interpretively for three-nucleon forces. The microscopic Hamiltonian

$t(H)=p^{|E(H)|}$6

includes a genuine three-body term $t(H)=p^{|E(H)|}$7. In chiral EFT, the N$t(H)=p^{|E(H)|}$8LO three-nucleon force consists of long-range two-pion exchange, one-pion–exchange–contact, and pure contact topologies, with only the low-energy couplings $t(H)=p^{|E(H)|}$9 and $H\in\mathcal H$0 beyond the two-pion-exchange coefficients $H\in\mathcal H$1. These interactions shift the oxygen dripline from $H\in\mathcal H$2O in NN-only calculations to the observed $H\in\mathcal H$3O, restore shell structure in calcium, and constrain neutron matter and neutron-star radii (Schwenk, 2013).

In quantum information, a triangle-based construction governs genuine tripartite entanglement for three-qubit pure states. The squared one-vs-rest concurrences

$H\in\mathcal H$4

obey triangle inequalities and therefore define a concurrence triangle. Its normalized area, the concurrence fill $H\in\mathcal H$5, is proposed as a genuine tripartite entanglement measure. The paper proves a Triangle No-Area Theorem: the area is zero if and only if at least one side vanishes, so $H\in\mathcal H$6 exactly on product or biseparable states. The normalization gives $H\in\mathcal H$7 and $H\in\mathcal H$8 (Xie et al., 2021).

In triangle geometry, a proposed trio of centers comprises the equiareal disk center, the illuminating center, and the thermodynamic center. These are defined, respectively, by minimizing Fraenkel asymmetry to an equal-area disk, maximizing a renormalized integrated brightness functional, and maximizing the first Dirichlet eigenfunction of the Laplacian on the triangle (Finch, 2014).

In flavour model building, tri-hypercharge replaces Standard Model hypercharge by three family-specific factors,

$H\in\mathcal H$9

with ${\rm UNi_4B}$00 as the low-energy hypercharge. If the Higgs doublets carry only third-family hypercharge, only third-family renormalisable Yukawa couplings are allowed. Hyperons break the three ${\rm UNi_4B}$01 factors to the diagonal subgroup and generate non-renormalisable Yukawas, producing fermion mass hierarchies, small CKM mixing, and a low-scale seesaw; one ${\rm UNi_4B}$02 boson can be as light as a few TeV (Navarro et al., 2023).

In algebraic generation models, a trio of trialities organizes both internal symmetry and family structure. The Standard Model gauge algebra is embedded inside

${\rm UNi_4B}$03

acting on the triality triple ${\rm UNi_4B}$04 with ${\rm UNi_4B}$05. The paper identifies two generations directly in ${\rm UNi_4B}$06 and ${\rm UNi_4B}$07, while a third generation is encoded in ${\rm UNi_4B}$08 through a Cartan factorization in which vector degrees of freedom are products of spinor degrees of freedom (Furey et al., 2024).

In extremal hypergraph theory, triangle-free triple systems classify the four non-isomorphic 3-edge triangle configurations in a 3-uniform hypergraph and study all 15 extremal problems obtained by forbidding subsets of those four patterns. The paper solves the new cases exactly or asymptotically and in many instances characterizes the extremal constructions (Frankl et al., 2024).

6. Comparative interpretation

A likely source of confusion is the word “force.” In quasirandom graph theory, the relevant object is a forcing family in the precise Chung–Graham–Wilson sense. In atomistic machine learning, the term names a three-stream representation architecture that improves energy and force prediction. In ${\rm UNi_4B}$09, it denotes a specific magnetic order. In nuclear theory, the expression aligns most literally with genuine three-body forces, but there it is an interpretive gloss rather than the paper’s native title.

This suggests three recurring functions of the label. First, it marks irreducible ternary structure: no forcing subpairs in graph theory, three symmetry-related ${\rm UNi_4B}$10 modes in triple-${\rm UNi_4B}$11 order, and three disentangled latent streams in MLIPs. Second, it often separates components that standard formulations entangle: edge density versus higher homomorphism data, dipole versus quadrupole order, and composition versus geometry versus interaction. Third, it frequently introduces a hidden or inaccessible degree of freedom that becomes explicit only after reformulation: edge density recovered from doubled-graph counts, quadrupole order revealed on magnetically disordered sites, and transferable chemical or structural information recovered from latent-space probing.

A second misconception would be to read the shared name as evidence of a shared formalism. The cited works do not support that conclusion. The graph-theoretic TriForces are families of small graphs in dense graph limits; the magnetic triforce is a symmetry-broken state of a correlated uranium compound; the machine-learning TriForces framework is an architectural wrapper for atomistic GNNs; the remaining usages concern EFT many-body terms, concurrence geometry, triangle centers, gauge flavour structure, triality representations, or forbidden hypergraph configurations. The commonality is therefore triadic design, not disciplinary continuity.