Geometric Algebra Meets Cartesian Tensors: Higher-Order Equivariance for Interatomic Potentials

Technical Summary: Geometric Algebra Meets Cartesian Tensors: Higher-Order Equivariance for Interatomic Potentials

Problem Statement
Machine learning interatomic potentials (MLIPs) must encode physical symmetries, specifically ensuring that predicted forces are equivariant under rotations and translations. While force magnitude is often predicted accurately, force direction frequently fails, leading to trajectories that drift from the correct potential-energy basin. The authors identify a structural limitation in existing Clifford (geometric) algebra-based architectures: the geometric product of two vectors in $\mathbb{R}^3$ (within the algebra $Cl(3,0)$) only realizes the $L=0$ (scalar) and $L=1$ (vector/bivector) irreducible representations (irreps) of $SO(3)$. The symmetric-traceless rank-2 component ($L=2$), which is essential for capturing quadrupolar charge distributions, d-orbital interactions, and anisotropic stress, is absent from the per-edge bilinear operations that drive message-passing layers. Consequently, standard Clifford models rely on stacking layers to indirectly reconstruct $L \ge 2$ content, resulting in poor directional fidelity (aggregate force-cosine similarity $< 0.25$ on rMD17) compared to spherical-harmonic-based models.

Methodology: CliffordSTF
The paper proposes CliffordSTF, a hybrid architecture that couples the Clifford multivector substrate with closed-form symmetric-traceless tensor (STF) tracks. The core innovation is a "dual-track" design that avoids Clebsch–Gordan (CG) tables, Wigner-D matrices, or external libraries like $e3nn$.

1. Theoretical Foundation (Theorem 1): The authors establish a constructive equivalence between the $Cl(3,0)$ geometric product and STF operations. They demonstrate that the tensor product $1 \otimes 1$ (two vectors) can be decomposed into $L=0, 1, 2$ components using only:

   • The scalar part of the geometric product ($L=0$).
   • The Hodge dual of the bivector part ($L=1$).
   • A closed-form symmetric-traceless projector ($L=2$).
     This construction spans the full $SO(3)$ irrep content up to $L=3$ at the per-edge bilinear level without learned CG coefficients.

2. Architecture:

   • Substrates: The model maintains two parallel feature tracks per atom: a Clifford multivector (dimension 8, covering $2 \cdot L=0 \oplus 2 \cdot L=1$) and optional STF channels at ranks 2 and 3.
   • Message Passing: Messages are generated via the geometric product on the Clifford track and closed-form STF contractions on the tensor track.
   • Cross-Track Coupling: A critical component is the bidirectional coupling between tracks. STF features are generated from Clifford vectors, and conversely, STF features contract with Clifford vectors (via $S \cdot v$) to inject $L=2$ information back into the vector channel. This coupling is learned via a single bilinear operation.
   • Readout: Forces are predicted using a combination of vector channels (grade-1) and Hodge duals of bivectors (grade-2), augmented by explicit $L=2 \otimes L=1 \to L=1$ contractions ($S \cdot \hat{r}$) that capture angular force components inaccessible to $L \le 1$ architectures.

Key Contributions

• Diagnosis: The paper isolates the missing $L \ge 2$ substrate, rather than model capacity or depth, as the primary cause of directional failure in $L \le 1$ Clifford baselines.
• Architecture: A closed-form construction that spans spherical-harmonic content through $L=3$ using a single learned bilinear and classical STF operations, verified against Wigner-3j symbols.
• Empirical Validation: An eleven-variant ablation study confirms that the hybrid scaffold is necessary; neither the Clifford track alone nor the STF track alone (without coupling) approaches the performance of the combined system.

Results
Evaluated on the rMD17, OC20, and OC22 benchmarks with a fixed parameter budget ($\sim 10^6$ parameters):

• rMD17 (Force Directionality): CliffordSTF achieves an aggregate force-cosine similarity of 0.551, a tenfold improvement over the base Clifford model (0.055) and outperforming all other CG-free or body-ordered baselines (all $\le 0.17$). It approaches the performance of full spherical-harmonic models (EquiformerV2 $\approx 0.96$) without using spherical harmonics.
• Magnitude Accuracy: The directional gains do not come at the expense of magnitude accuracy. CliffordSTF reduces force MAE by 15.8% and energy MAE by 10.9% compared to the base Clifford model.
• Catalysis (OC22):
  • IS2RE: Attains the best in-distribution energy MAE among $L \ge 2$ methods (4.09 eV) and competitive out-of-distribution performance.
  • S2EF: Achieves the best out-of-distribution energy MAE in the study and matches the best in-distribution energy MAE among $L \ge 2$ methods.
• Efficiency: At inference, CliffordSTF (0.352 s/step) is competitive with slower spherical-harmonic baselines and faster than Cartesian tensor baselines like ICTP. Training is approximately 2–3x slower than the fastest spherical-harmonic baselines (MACE, EquiformerV2) due to the autograd graph overhead of the dual-track scaffold, but remains competitive with other tensor-field methods.

Significance and Claims
The paper claims that representational choice in equivariant architectures should be driven by the specific angular content required ($L$-values) rather than loyalty to a single algebraic formalism. By demonstrating that the geometric product and closed-form STF operations together span the same $SO(3)$-irrep content as spherical harmonics through $L=3$, the authors show that a hybrid approach can recover the missing structural expressivity of pure geometric algebra.

The authors modestly note that while CliffordSTF closes the gap significantly, it does not yet match the absolute directional fidelity of fully spherical-harmonic architectures (EquiformerV2). They identify the cross-track coupling as a "stability lever" essential for training, rather than just an accuracy driver, and suggest future work could extend this to higher $L_{max}$, optimize training costs via fused kernels, and achieve full $O(3)$ equivariance through explicit parity tracking.