WorksPhysics and Astronomy
When Do Tensor Product Paths Matter? A Group-Theoretic Analysis of Equivariant Interatomic Potentials
Abstract
Tensor-product (TP) layers in equivariant interatomic potentials encode angular interactions through symmetry-allowed paths. Yet local site symmetry can make an allowed path carry no signal after neighbor aggregation. We call this effect Aggregation Cancellation. In a scalar-input first layer, equal-species neighbors on a complete site-symmetry orbit share scalar and radial factors, reducing their angular contribution to a spherical-harmonic orbit sum. If the restricted degree- representation has no trivial component, this sum and the corresponding path vanish for any learned radial weight. For an ideal octahedral orbit, this removes the classes. Beyond the first layer, the factorisation no longer holds, so cancellation is no longer guaranteed; we instead ask whether affected paths remain less important on average. We test the two levels separately in NequIP: matched-radius TiO probes examine first-layer cancellation, whereas a within-model CaTiO comparison and six OMat24 models examine the importance bias. We then rank paths on training frames, choose the pruning fraction on validation frames, and evaluate frozen masks on held-out tests. At the highest evaluated fraction (60% of 174 paths), force mean absolute error (MAE) changes are 1.42%, 0.04%, and 1.83% for CaTiO, OMat24 -cubic, and OMat24 low-symmetry data, respectively. The masks imply 67.12-71.94% potential TP-operation reduction and 59.77% potential edge multilayer perceptron (MLP) reduction; these are structural proxies, not measured speedups. The results connect local symmetry to path redundancy and pruning robustness in NequIP; transfer to other architectures remains open.
Introduction
Equivariant interatomic potentials encode atomic geometry by coupling learned features with spherical harmonics through Clebsch–Gordan tensor products (TPs) [4, 3, 19, 16, 11]. A TP instruction combines an input feature of angular momentum with an edge harmonic of degree to produce an output feature . Triangle and parity rules determine which instructions are allowed by equivariance, but they do not determine whether neighbor aggregation supplies a nonzero signal to each allowed instruction. This distinction is consequential because practical models may contain hundreds of weighted paths: the five-layer NequIP architecture studied here has 174 at .
The unresolved question is therefore not only which TP paths an architecture permits, but when those paths matter for the environments presented to a trained model. Recent methods reduce TP cost by changing the computation, selecting paths architecturally, or learning gates [22, 28, 15, 17]; pruning criteria instead rank parameters by magnitude or sensitivity [14, 18]. These approaches can expose redundancy, but a low score does not by itself explain why a path is inactive on a particular neighbor distribution. Likewise, a favorable curve evaluated on the data used to choose a mask does not establish that the mask generalises. A mechanistic account therefore needs both a symmetry-based prediction and an evaluation that separates ranking, selection, and test.
We address this gap by analysing neighbor aggregation under local site symmetry. In a scalar-input first layer, same-species neighbors begin with the same embedding. Members of one site-symmetry orbit also share a radius and therefore a radial-network output. Their angular contribution factorises into the spherical-harmonic orbit sum . Because the site group only permutes the summands, lies in the trivial subspace of the restricted degree- representation. If this subspace is absent, the orbit sum and the associated first-layer TP path vanish for any learned radial weight. We call this effect Aggregation Cancellation.
This argument produces an exact first-layer result and a testable hypothesis for the full network. For an ideal octahedral orbit, the first-layer contributions cancel exactly. In deeper layers, symmetry-related neighbors carry transformed equivariant features rather than a shared scalar, so the factorisation no longer holds in general. The exact selection rule therefore does not imply permanent inactivity. Instead, we test whether the first-layer bottleneck appears as a statistical bias toward lower path importance. Conversely, lowering site symmetry permits more invariant components but does not rank their learned magnitudes; bonding, composition, distortion, and temperature remain relevant.
We test the exact and statistical claims separately. Matched-radius TiO probes test first-layer cancellation and parity-dependent activation under controlled distortions. A within-model CaTiO temperature comparison and six OMat24 NequIP models test the broader importance bias and its limits. Finally, a three-system pruning study ranks paths on training frames, chooses the pruning fraction on validation frames, and evaluates the frozen mask on held-out test frames; post hoc controls are kept separate. All trained-model evidence uses NequIP. The first-layer argument may apply to other architectures that satisfy the same scalar-input and radial-aggregation assumptions, but their importance and pruning behaviour are not tested here.
Contributions.
We derive an orbit-factorisation result and a group-independent criterion, expressed using the trivial irrep , for exact first-layer TP-path cancellation under complete equal-species, equal-radius neighbor orbits.
Controlled TiO probes test the exact first-layer result, while a within-model CaTiO comparison and six OMat24 NequIP models test the predicted importance bias without claiming that symmetry alone orders path magnitude.
Under disjoint ranking, selection, and test roles, the validation-selected 60% masks change held-out force mean absolute error (MAE) by 0.04–1.83% across three systems. They imply 67.12–71.94% potential TP-operation reduction and 59.77% potential edge multilayer perceptron (MLP) reduction; these are structural operation counts, not measured speedups. Post hoc OMat24 controls compare the gradient masks with equal-count random and magnitude baselines.
Related Work
Equivariant interatomic potentials. Tensor field networks and related constructions introduced learned features that transform under irreducible representations of rotations [7, 13, 24, 1]. The e3nn library made these representation-theoretic operations reusable across architectures [11], while SO3krates developed an alternative equivariant attention-based atomistic representation [9]. NequIP showed that such features can produce data-efficient force fields, while MACE connected equivariant message passing to higher-body-order atomic cluster expansions [4, 3]. Allegro localised the representation for scalable simulation, and EquiformerV2 combined equivariant operations with attention [19, 16]. SevenNet further demonstrates scalable equivariant potentials for periodic materials [21]. Unified analyses relate these models to the atomic cluster expansion and clarify their expressivity [8, 20, 23]. These works establish what an architecture can represent; our question is when the neighbor distribution annihilates a specific allowed path before its learned coefficient can matter.
Efficient tensor products. Several recent methods reduce TP cost by changing the computation itself. SO(3)-to-SO(2) reductions exploit rotations into an edge-aligned frame [22]; path-design studies remove couplings according to architectural criteria [28]; and fused or low-rank implementations improve kernel efficiency [15]. Adaptive path gates have also been learned for equivariant Hamiltonian prediction [17]. These methods are complementary to our analysis. Aggregation Cancellation predicts that the utility of a path depends on the site-symmetry orbits contained in the deployment data. It can therefore guide structural path removal, but only after its distributional assumptions are checked.
Symmetry adaptation and pruning. Point-group-adapted representations explicitly restrict features to subspaces compatible with a crystal symmetry [12]. We instead show that an ordinary -equivariant message-passing layer can encounter the same trivial-subspace projection implicitly through orbit aggregation. Classical pruning scores use weight magnitude, second-order sensitivity, or gradient criteria [14, 18, 10]. Our gradient score is a practical ranking device rather than the theoretical contribution. The group argument explains why the data may produce low sensitivity, while the disjoint ranking/selection/test protocol tests whether that sensitivity generalises.
Aggregation Cancellation
First-Layer Orbit Factorisation
Consider a center atom and a weighted TP instruction . Suppressing channel indices, its message is
Here is produced by the edge MLP, is a Clebsch–Gordan coefficient [27, 26], and is a spherical harmonic. Let be the site-symmetry group of the center and partition a complete neighbor shell into orbits under . Within one orbit, symmetry preserves radius and species.
Proposition 1 (Orbit factorisation). For a first-layer path with , the contribution of orbit factorises as
where contains the common species feature, radial weight, channel mixing, and CG factor. Hence makes the whole orbit contribution zero for any learned radial weight.
The assumptions are explicit. First, the input is the scalar species embedding used by NequIP in layer 1. Second, the orbit is complete inside the neighbor list. Third, orbit members have the same species and radius. No assumption is made about the numerical value of . This is why the cancellation is architectural-data interaction rather than a coincidental small learned weight.
Proposition 2 (Site-symmetry criterion). The orbit sum lies in , the invariant subspace of the spherical representation restricted to . Therefore
where is the trivial irrep of the relevant group.
For every , applying to the sum rotates each spherical harmonic and merely permutes the members of . Thus . Equivalently, the Reynolds projector fixes the sum. If the restricted representation contains no trivial copy, the projector has no invariant image and the sum must vanish. The implication is one-way: even if a trivial component is present, the coefficient of that component can be zero for a particular orbit geometry. These restriction and projection arguments follow standard finite-group representation theory [25]. Appendix A gives the complete derivation.
Figure 1. Aggregation Cancellation at an octahedral Ti site. Six equivalent oxygen neighbors form one orbit and their spherical-harmonic sums vanish for . A lower-symmetry distortion partitions the shell into smaller orbits and can activate previously forbidden even components. Exact odd- cancellation survives whenever an orbit retains inversion-related pairs; thermal disorder, non-centrosymmetric sites, or deeper features can relax it.
Concrete Site Groups
For , the trivial irrep is conventionally named . The restrictions through are
Only contains , so an ideal six-neighbor octahedral orbit transmits only the scalar first-layer component for . For a general group we retain the notation : writing for every group would incorrectly assume both the naming convention and inversion parity.
The tetrahedral group illustrates the distinction between permission and importance. Its trivial irrep is , and the degree-three representation contains an component. A tetrahedral orbit can therefore have a nonzero sum even though its and sums are forbidden. Lower-symmetry groups generally admit more invariant components, but the theorem does not rank those components. Their amplitudes depend on the orbit geometry and on how the radial and channel weights respond to chemistry.
Inversion provides a useful special case. If every direction in an orbit is paired with its opposite, then , so all odd- sums cancel. This remains true for a centrosymmetric distorted orbit even when the full selection rule has been lost. It also clarifies why a crystallographic space-group label is only a proxy: a dataset can contain several Wyckoff sites, thermal snapshots can break instantaneous local symmetry, and a finite cutoff can truncate an otherwise complete orbit. Fixed-radius TiO probes in trained CaTiO and OMat24 models reproduce ideal cancellation and parity-dependent activation (Appendix Figure 6); they are not force-accuracy tests.
Why the Effect Persists Beyond Layer 1
The exact proof does not extend mechanically to later layers. After one interaction, symmetry-related neighbors satisfy rather than sharing one scalar value. The feature and edge harmonic are then coupled inside the orbit sum, and the common factor in Equation equation 2 cannot generally be extracted. A direct enumeration on an ideal octahedron finds only three active first-layer triplet types but 31 of 34 types can be active once higher-order neighbor features are supplied.
Nevertheless, the first layer controls the information initially injected into the network. Directions suppressed there must be reconstructed through distortion, other site types, or deeper feature interactions. We therefore expect an upstream symmetry bottleneck to appear statistically as lower sensitivity of many downstream paths. The prediction is deliberately qualitative: near- data should show a lower high- importance profile than a matched lower-symmetry condition, while no universal ordering of follows. This distinction determines the experimental design below.
Experimental Design
Models, Intervention, and Metrics
All models use NequIP 0.17.1 with , five interaction layers, 64 channels per irrep, and a 5.0 Å cutoff. The architecture has 174 weighted layer–instruction blocks. The edge MLP emits one contiguous coefficient block for each instruction, so a path can be intervened on by zeroing the corresponding final-linear columns. For path , we report
This measures functional sensitivity of a trained model; because the dense instruction list and tensors are unchanged, ablation alone does not reduce runtime.
The CaTiO comparison is the most controlled symmetry experiment. One model is trained jointly on 603 MatterSim-driven MD frames ([29]) sampled at 1000, 1500, and 2000 K. The full split contains 482/60/61 train/validation/test frames. The ablation analysis contrasts the 1000 and 2000 K endpoints, avoiding a comparison between separately trained networks. The 1000 K structures are predominantly distorted Pnma environments, whereas the 2000 K trajectory is closer to cubic Pmm. The temperature label is not treated as an exact site-symmetry annotation.
OMat24 ([2]) supplies broader but observational evidence. We train two 50k-frame perovskite models stratified by temperature and four models stratified by equilibrium space-group class: -cubic, -cubic, mid-symmetry, and low-symmetry. The class label is a dataset-level proxy and covaries with chemistry. We therefore use OMat24 to test whether the predicted trend recurs across data regimes, not as a clean causal isolation of point group.
Two additional datasets provide diagnostic contrasts. MPF denotes the MPF2021 relaxation corpus released with M3GNet ([5]); our F–Fe–O subset contains 372 materials and 1116 density functional theory (DFT) relaxation frames. Aspirin from MD-17 ([6]) is a low-symmetry molecular case. These datasets contextualise path sensitivity but are not included in the three-system confirmatory pruning claim. Their construction and split limitations are described in Appendix § 10.
Pruning Evaluation Protocol
The pruning evaluation assigns a single statistical role to each split. Training data estimate a path ranking, validation data select a pruning fraction, and held-out test data estimate the final accuracy change. For every weighted path with edge-MLP block , we accumulate force-loss gradients over training mini-batches and define
Paths are sorted in ascending . For each cumulative fraction , the unmodified checkpoint is restored before applying the corresponding mask. We select the largest prefix-safe fraction: every evaluated fraction up to it must remain within the predeclared 5% validation force-MAE tolerance. The selected mask is then frozen and used for one baseline/pruned comparison on the test split. There is no mask fine-tuning or test-guided choice in this primary evaluation. Later controls reuse the inspected test sets and are not a fresh confirmatory sample.
For CaTiO, ranking uses 10 balanced endpoint frames drawn from 327 endpoint training frames; selection uses all 36 endpoint validation frames; and testing uses all 39 endpoint test frames (23 at 1000 K and 16 at 2000 K). For each OMat24 model, ranking uses 10 frames from 40,000 training frames, selection uses 200 of 5,000 validation frames, and testing uses all 5,000 held-out frames. File-level audits find no exactly duplicated frame across the OMat24 splits. The available files do not contain a material or trajectory identifier, so these OMat24 splits are accurately described as frame-disjoint, not proven material-disjoint.
Results
Controlled CaTiO Importance Shift
Figure 2 groups the 174 single-path ablations by . Relative to the near-cubic 2000 K endpoint, the lower-symmetry 1000 K endpoint has larger mean normalized ablation for every , with endpoint ratios , , and . The result matches the predicted direction: angular paths suppressed by an ideal octahedral orbit become more useful when distortions and heterogeneous local environments increase. We do not interpret the ordering of the three ratios. It is affected by finite-temperature symmetry breaking, contributions from Ca and O centers, and paths in layers where exact factorisation no longer applies.
Figure 2. CaTiO normalized single-path ablation grouped by edge angular momentum at the 1000 and 2000 K endpoints. One jointly trained model is evaluated under both conditions. The consistent increase for in the lower-symmetry endpoint is the controlled test of the predicted importance bias; the theory does not predict an ordering among .
The gradient and intervention rankings agree strongly, with Spearman correlations of 0.85 at 1000 K and 0.82 at 2000 K. This supports the use of accumulated gradients as a low-cost mask ranking signal. It does not make the ranking data-independent: the score is still calibrated on a specific distribution, which is why the pruning experiment keeps selection and test structures separate.
Symmetry Permission Does Not Fix Magnitude
Figure 3 separates the OMat24 space-group models. The -cubic model has the smallest mean normalized ablation for (0.38%, 0.30%, and 0.34%), consistent with a strong first-layer octahedral bottleneck. The -cubic model instead reaches 3.17%, 2.77%, and 3.36%. In particular,

Figure 3. OMat24 mean normalized ablation for models stratified by equilibrium space-group class. The -cubic subset has the lowest high- importance, while the subset demonstrates that symmetry permission and learned magnitude are different questions. The full-width panel is shown separately from Figure 2 to preserve readable labels and uncertainty bars.
is symmetry-allowed for a tetrahedral orbit, so its large learned importance is compatible with the theorem. Directional covalent bonding offers a chemical explanation for why an allowed channel is strongly used.
The mid- and low-symmetry models lie between these extremes for the reported averages. This is a useful negative lesson: a lower group permits more invariant components but need not make all of them more important than in every higher-symmetry chemistry. Because the four models differ in composition and training distribution, the comparison is observational. The evidence is strongest when read together with the within-model endpoint shift rather than as a standalone causal ranking of space groups.
Held-Out Pruning Robustness
All three validation curves remain below the 5% tolerance through the highest evaluated fraction, 60% (Figure 4). After freezing those masks, held-out force MAE changes by +1.42% for , +0.04% for OMat24 -cubic, and +1.83% for OMat24 low-symmetry (Table 1). The corresponding absolute force MAEs are 23.68 to 24.02, 229.19 to 229.28, and 167.22 to 170.27 , respectively. For , the held-out changes are also small when separated by endpoint: +1.80% at 1000 K and +1.13% at 2000 K.
Figure 4. Validation selection curves for the three confirmatory systems. Lines use validation frames only; stars show the held-out test changes after the 60% masks were frozen. All validation curves remain below the predeclared 5% tolerance. Sixty percent is the highest evaluated pruning level, not a claimed global maximum.
| System | Frames | Base | Pruned | (%) | TP (%) | MLP (%) |
| CaTiO endpoints | 10/36/39 | 23.68 | 24.02 | 1.42 | 71.94 | 59.77 |
| OMat24 | 10/200/5000 | 229.19 | 229.28 | 0.04 | 67.12 | 59.77 |
| OMat24 low-sym. | 10/200/5000 | 167.22 | 170.27 | 1.83 | 68.45 | 59.77 |
Table 1. Held-out evaluation at the validation-selected 60% path mask. Frame counts are ranking/selection/test. Force MAEs are in meV/Å. TP and MLP columns are potential operation reductions, not measured latency.
The result is nearly invariant to the mask, whereas the low-symmetry result has the largest degradation. This ordering is consistent with the proposed mechanism, but it is not by itself a controlled symmetry intervention because the two OMat24 models differ in data distribution. More importantly, both remain below the tolerance after the ranking, fraction choice, and test evaluation are separated. The result therefore supports pruning robustness under the observed splits rather than merely reproducing a mask on the data used to choose it.
At fixed 60% pruning, three OMat24 gradient rankings change test MAE by 0.040–0.065% () and 1.825–2.241% (low symmetry); every layer-matched random mask degrades more (Appendix Figure 8). This post hoc result does not isolate symmetry as the cause.
The selected masks share substantial structure. The and low-symmetry masks overlap on 89 of 104 paths (Jaccard index 0.75), while overlaps with them at 0.81 and 0.87 Jaccard, respectively. None of the three 60% masks removes an instruction from the zero-indexed first interaction layer; the pruned paths are concentrated in later layers. For example, the mask removes 11, 41, and 52 paths from layers 2–4 (zero-based layers 1–3), and the low-symmetry mask removes 7, 40, and 57. This does not contradict the theorem. Exact cancellation creates an upstream information bottleneck, while gradient ranking identifies the downstream blocks whose removal least changes the trained computation. The masks should therefore be understood as mostly shared downstream sparse subnetworks compatible with, but not uniquely determined by, the first-layer selection rule.
Potential Operation Reduction
Pruning a fraction of paths does not translate linearly to TP arithmetic because instructions have different representation dimensions. For path with multiplicity , define the dimension-weighted dense-contraction proxy
The proxy counts products over representation components before implementation-specific Clebsch–Gordan sparsity. Because the masks preferentially select some high-dimensional paths, removing 60% of the instructions corresponds to 67.12–71.94% of this potential TP work.
For the final edge-MLP linear layer, each 64-wide path block has the same output width in these models. Its removable multiply–accumulate fraction is
where is the input width of the layer-specific final linear map. This component-level accounting deliberately omits scatter, radial basis evaluation, nonlinearities, memory traffic, force backpropagation, kernel launch overhead, and any CG sparsity exploited by e3nn. Moreover, setting columns to zero still executes the original dense graph. Realizing the potential reduction requires deleting TP instructions, compacting the edge-MLP outputs, rebuilding the model, and benchmarking energy-and-force inference on the target hardware. We therefore use the phrase pruning robustness and potential operation reduction, not compression speedup.
Discussion
Scope and Limitations. Aggregation Cancellation connects local site-symmetry orbits to the internal use of an equivariant network: architectural equivariance permits a coupling, whereas orbit aggregation can annihilate its actual input. The criterion is exact only for a complete first-layer orbit satisfying the stated equal-species and equal-radius assumptions. It does not rank symmetry-allowed channels, and thermal disorder, heterogeneous sites, cutoff truncation, and deeper features turn the selection rule into an empirical importance bias. All trained-model evidence in this work uses NequIP. Although the first-layer factorisation also applies to architectures that share the assumed scalar input and radial-weighted neighbor aggregation, we do not empirically establish the same importance or pruning behavior for MACE, Allegro, or other equivariant interatomic potentials.
Design and evaluation implications. Persistently symmetric domains may remove provably forbidden first-layer instructions before training; mixed phases are better served by validation-calibrated masks or conditional gates. The observed mask overlap suggests a partially reusable sparse backbone, but transfer requires fresh validation. Three post hoc ranking draws show stability at fixed 60% pruning; OMat24 evaluation remains frame-disjoint rather than material-disjoint. The operation proxies identify removable arithmetic, while acceleration requires structural compilation and hardware timing.
Conclusion
Within the scalar-input first layer studied here, equivariance determines which angular couplings are permitted, but the deployment data determine which of those couplings are activated. We formalised this distinction through Aggregation Cancellation: equal-weight neighbors on a site-symmetry orbit project the first-layer spherical signal onto the trivial subspace, exactly eliminating a TP path when that subspace is absent. Because deeper equivariant features break the scalar factorisation, the exact first-layer rule becomes a network-wide importance bias rather than permanent inactivity.
Matched-radius probes reproduce the predicted first-layer cancellation and distortion response. Across the NequIP models evaluated here, CaTiO ablations and six OMat24 models support the broader importance bias while showing that symmetry permission does not determine learned magnitude. In the primary study, the original 60% masks change held-out force MAE by 0.04–1.83% and imply 67.12–71.94% potential TP-operation reduction. Post hoc OMat24 controls show that gradient masks are stable across ranking seeds and outperform the random and magnitude-based controls. These findings support pruning robustness, not measured compression or speedup. Material-disjoint evaluation, compiled benchmarks, and validation on architectures such as MACE and Allegro remain necessary before making a broader claim about equivariant interatomic potentials.
AI Use Statement
Generative AI tools were used to critique and refine the presentation of theoretical material, provide feedback on experimental design, assist with code implementation and scientific figures, support the interpretation of experimental results, and help organize, draft, and edit the manuscript and references. All AI-assisted content was reviewed and verified by the authors. All reported numerical results were obtained from the authors’ computational experiments. The authors take full responsibility for all mathematical claims, code, citations, experimental results, and final content.
Proofs and Group-Theoretic Details
Orbit Factorisation
In the first layer, a neighbor of species carries the scalar feature . For all in a site-symmetry orbit , the distance and species are constant. Restricting Equation (1) to this orbit and using gives
where the CG selection rule fixes for . Thus the learned radial function enters only through the common scalar and cannot undo a vanishing angular sum.
Invariant-Subspace Criterion
Let act on the directions in . For every ,
because permutes orbit members. Hence belongs to . Equivalently, the Reynolds projector satisfies . If the trivial representation is absent from the restricted representation, this projector is zero and .
For a centrosymmetric orbit, every direction has an opposite partner and ; all odd- sums therefore vanish. This parity argument is a special case of the invariant-subspace criterion. The notation is used only for centrosymmetric point groups such as ; for the trivial irrep is , and is the group-independent notation.
| Contains ? | Ideal orbit sum | ||
| 0 | yes | non-zero allowed | |
| 1 | no | zero | |
| 2 | no | zero | |
| 3 | no | zero |
Table 2. Restrictions of to for $l\leq 3.
Pnma, , and Deeper Layers
Octahedral tilting in Pnma splits the ideal oxygen shell into smaller orbits and can activate even- components that forbids. The ideal Ti site nevertheless remains centrosymmetric, so opposite-pair cancellation still removes odd in its first-layer scalar orbit sum. Odd- importance measured for the full network can instead arise from thermally broken instantaneous geometry, non-centrosymmetric atomic sites, or paths in deeper layers. We therefore do not use the CaTiO ratio as a direct proof of first-layer Ti-site activation.
For a tetrahedral orbit, symmetry forbids but permits a trivial component at . Lower-symmetry groups permit progressively more components, although permission alone does not determine their learned magnitude. In deeper NequIP layers, symmetry-related neighbors obey rather than carrying identical scalars. A direct e3nn calculation on an ideal octahedron finds only three non-zero layer-1 triplets but 31 of 34 path types active once higher-order neighbor features are present.
Figure 5. Norms of spherical-harmonic orbit sums for representative site groups. A zero bar indicates a symmetry-forbidden component; an allowed component may still vanish for a particular orbit geometry.
Experimental and Dataset Details
Architecture and Training
Models use NequIP 0.17.1, five interaction layers, 64 features per irrep, , and Å. Training uses Adam, exponential moving average (EMA) weights with decay 0.999, and a force-dominated energy–force objective. All reported evaluations use EMA checkpoints.
CaTiO. MatterSim-driven trajectories contain 603 total frames: 201 each at 1000, 1500, and 2000 K. The 482/60/61 train/validation/test files contain temperature counts 163/155/164, 15/24/21, and 23/22/16, respectively. One model is trained jointly on all three temperatures. The symmetry and pruning analyses contrast the 1000 and 2000 K endpoints; 1500 K remains part of training but is excluded from the endpoint evaluation. This design makes the importance comparison within-model rather than a comparison between independently trained networks.
OMat24. For the temperature study, two models are trained on 50k ABO frames at 1000 and 3000 K. For the space-group study, structures are partitioned into -cubic (space groups 221–230), -cubic (216–220), mid-symmetry (tetragonal, trigonal, and hexagonal), and low-symmetry (orthorhombic, monoclinic, and triclinic) subsets. Equilibrium space group is a dataset-level proxy and does not imply exact instantaneous site symmetry at finite temperature.
MPF2021 F–Fe–O. MPF2021 is the DFT relaxation corpus released with M3GNet [5]. We filter structures whose element set is exactly {F, Fe, O}, obtaining 372 materials and 1116 frames. Using seed 42, the current dataset files contain 780 training, 112 validation, and 224 test frames. The existing preparation script performs a frame-level split; a material-disjoint split is preferred for a final generalisation claim and should be regenerated before submission if multiple relaxation frames share a material identifier across partitions.
Aspirin. The molecular contrast uses aspirin from MD-17 [6], with 950 training, 50 validation, and 1000 test structures generated using seed 123. It uses the same five-layer, , 64-channel NequIP architecture and is included as a low-symmetry counterexample, not as an accuracy benchmark against inorganic models.
Per-Path Ablation
For each of the 174 layer–instruction blocks, we restore the original checkpoint, zero only the associated edge-MLP output columns, and evaluate force MAE. This intervention measures functional sensitivity but leaves the dense execution graph unchanged.
Algorithm 1: Per-Path Ablation for TP Path Importance
Input: Trained NequIP module $M$; test loader $D$; number of eval batches $B$
Output: Per-path importance scores $\{\widehat{\Delta\mathrm{MAE}}_k\}_{k=1}^{174}$
1 entries $\leftarrow$ []
2 for $n \in \{0,\ldots,4\}$ do
3 conv $\leftarrow M.\mathrm{layer}_n.\mathrm{conv}$
4 tp $\leftarrow$ conv.tp_scatter.tp
5 $W \leftarrow$ conv.edge_mlp.last_linear.weight
6 $c \leftarrow 0$
7 foreach instruction $\iota \in$ tp.instructions do
8 $l_{\mathrm{edge}} \leftarrow$ tp.irreps_in2[$\iota.i_{\mathrm{in2}}$].ir.l
9 entries.append({$W$, col_start=$c$, col_end=$c+64$, $l_{\mathrm{edge}}$})
10 $c \leftarrow c+64$
11 end
12 end
13 $\mathrm{MAE}_{\mathrm{base}} \leftarrow \mathrm{EvalForceMAE}(M,D,B)$
14 $W_{\mathrm{orig}} \leftarrow \{W.\mathrm{data.clone()}\ \forall W\}$
15 for $k \leftarrow 1$ to $|\mathrm{entries}|$ do
16 $e \leftarrow \mathrm{entries}[k]$
17 $e.W.\mathrm{data}[:,e.\mathrm{col\_start}:e.\mathrm{col\_end}] \leftarrow 0$
18 $\mathrm{MAE}_k \leftarrow \mathrm{EvalForceMAE}(M,D,B)$
19 foreach $(W,W_0) \in W_{\mathrm{orig}}$ do
20 $W.\mathrm{data} \leftarrow W_0$
21 end
23 $\widehat{\Delta\mathrm{MAE}}_k \leftarrow (\mathrm{MAE}_k-\mathrm{MAE}_{\mathrm{base}})/\mathrm{MAE}_{\mathrm{base}}$
24 end
25 return $\{\widehat{\Delta\mathrm{MAE}}_k\}$
Pruning Protocol and Operation Counts
The implemented protocol uses training data for ranking, validation data for mask selection, and the held-out test split only after the mask has been frozen. The complete procedure is given in Algorithm 2. For CaTiO, the ranking pool contains 327 endpoint training frames and the ten ranking frames are sampled in a balanced 5/5 split between 1000 and 2000 K. All 36 endpoint validation frames are used for selection and all 39 endpoint test frames are used for the final report. For both OMat24 models, the ranking sample is 10 of 40,000 training frames, the selection sample is 200 of 5,000 validation frames, and the final evaluation uses all 5,000 test frames. The test set is never used to rank paths, choose a threshold, or select among masks.
The CaTiO manifest records file hashes and a temperature-aware split. Exact structure hashes for the OMat24 and low-symmetry exports show zero exact duplicates across every train/validation, train/test, and validation/test pair. Stable material and trajectory identifiers are absent, however, so this verifies frame-level disjointness rather than material-level disjointness.
For a directed edge, define and
Counting one multiply and one add as two operations changes numerator and denominator by the same factor. For the final edge-MLP linear layer with hidden width ,
Algorithm 2: Gradient-Ranked Cumulative TP Path Pruning
Input: Trained model ; training-ranking, validation-selection, and held-out test loaders , , ; pruning fractions ; tolerance ; ranking batches
Output: Frozen mask ; one held-out test record
collect all weighted (layer, TP instruction) entries; foreach do
locate its edge-MLP column block ; compute on : ; compute ;
end
in ascending order; clone original edge-MLP weights;
; foreach do
restore ; ; ; foreach do
end
;
end
; restore ; apply the frozen mask ; evaluate exactly once and compute operation reductions; restore and return and the test record;
These proxies omit CG sparsity, fused-kernel schedules, nonlinearities, scatter, memory traffic, and reverse-mode force evaluation. They quantify removable arithmetic only after structural instruction deletion and edge-MLP compaction.
Additional Results
Matched-Radius First-Layer Distortion Probe
To test the orbit argument within trained networks, we construct a nonperiodic seven-atom TiO cluster with one Ti center and six identical O neighbors at a fixed Ti–O radius of 2.0 Å. The ideal oxygen directions are , , and . At each angle , the inversion-preserving case rotates the oxygen by toward and moves its opposite partner to the negative of the new direction. The inversion-broken case moves only the oxygen. Species, bond lengths, and the number of incoming Ti–O edges are unchanged.
We compute the norm of the degree- spherical-harmonic sum with e3nn component normalization and capture the pre-linear output of the trained first-layer TP scatter at the Ti center. The latter is grouped by output degree and divided in Figure 6 by the model’s ideal-geometry norm. For both the CaTiO and OMat24 checkpoints, the ideal responses are below in raw norm. Under inversion-preserving tilts, the odd- responses remain below while increases with angle. Moving only one oxygen also activates and . The probe isolates geometric symmetry in one trained layer; it does not evaluate forces or simulate a thermodynamic distortion trajectory.
Figure 6. Trained first-layer TP-scatter norms at the Ti center, normalized separately for each model by its ideal norm. Solid lines break inversion; the dashed line preserves opposite oxygen pairs. In the latter case, the responses remain at numerical zero and are omitted from the plotted series. All six Ti–O bonds remain at 2.0 Å.
CaTiO3 Ratios
| 0 | 1 | 2 | 3 | |
| Paths across five layers | 21 | 46 | 56 | 51 |
| 1.23 | 1.38 | 1.07 | 1.69 |
Table 3. Normalized CaTiO3 ablation ratio between 1000 and 2000 K. The selection rule predicts a directional bias for symmetry-suppressed channels, not an ordering of the ratios.
OMat24 Temperature Study
Figure 7. Mean normalized ablation for OMat24 perovskite models. Near-equilibrium 1000 K structures show negative high- ablation, whereas thermal disorder at 3000 K activates the same path groups.
| -cubic | -cubic | mid-sym. | low-sym. | |
| 0 | 20.3 | 77.7 | 48.5 | 21.1 |
| 1 | 0.38 | 3.17 | 1.88 | 0.96 |
| 2 | 0.30 | 2.77 | 1.64 | 0.50 |
| 3 | 0.34 | 3.36 | 1.26 | 0.66 |
Table 4. OMat24 mean normalized ablation (%).
Additional Pruning Results
| System | Validation | Test | Test frames |
| CaTiO endpoints | 1.136 | 1.418 | 39 |
| OMat24 -cubic | 0.034 | 0.040 | 5000 |
| OMat24 low-symmetry | 1.931 | 1.825 | 5000 |
Table 5. Validation and held-out test changes at the selected 60% mask. All values are relative force-MAE changes in percent.
Validation and test changes are close for all three systems, with no evidence that the selected fraction exploits the validation subset. The experiment stops at 60%, so these results establish robustness at the highest evaluated level but do not locate the failure point. Earlier same-loader diagnostics are reported in Appendix § 12.§ 6 for transparency, but are not used to support the confirmatory claim.
The 104-path masks are distributed primarily over later interaction layers. Using zero-based layer indices, CaTiO removes 11/41/52 paths from layers 1/2/3, removes 16/43/45, and low-symmetry removes 7/40/57; none removes a layer-0 instruction. The /low-symmetry masks share 89 paths (Jaccard 0.748), while the CaTiO mask has Jaccard overlaps 0.809 and 0.874 with and low-symmetry. This downstream concentration is compatible with an upstream symmetry bottleneck affecting sensitivity, but does not by itself identify the cause of the final mask.
Fixed-Rate OMat24 Pruning Controls
We examine the stability of the original OMat24 result at a fixed 60% pruning fraction, equivalent to 104 of 174 weighted paths. For each model, three gradient rankings use separately sampled 10-frame training subsets with seeds 20260910–20260912. Three unrestricted random masks sample the same number of paths. Three layer-matched random masks preserve the first gradient mask’s per-layer path counts, and one magnitude mask removes the paths with the smallest RMS weight in their final edge-MLP linear blocks. The same 5,000 test frames are used for every mask within a model. No mask is fine-tuned. These controls were designed after inspection of the original test result and therefore do not constitute a second independent confirmatory test.
Across gradient seeds, the force-MAE increase is 0.040–0.065% (mean 0.055 0.013%), and the low-symmetry increase is 1.825–2.241% (mean 2.084 0.226%). The values are sample standard deviations across three masks, not test-set confidence intervals. Layer-matched random masks give means of 32.00 4.40% and 124.99 10.35%, respectively. The unrestricted random means are 149.67% and 158.28%; the single magnitude masks give 220.54% and 307.48%. These equal-path-count comparisons support the practical value of the gradient ranking, including beyond its preference for particular layers. They do not match the masks by TP-operation proxy or angular composition, and they cannot establish that site symmetry alone causes the separation.
Figure 8. Post hoc OMat24 controls at fixed 60% path pruning, evaluated on the same 5,000 test frames per model. The vertical axis is logarithmic and shows the percentage increase in force MAE relative to the unpruned checkpoint. Each dot is one mask; black horizontal ticks show method means. Gradient, layer-matched random, and unrestricted random have three seeds each; magnitude has one mask. This is a robustness diagnostic, not a new test split.
Exploratory Same-Loader Pruning Diagnostics
Before fixing the confirmatory evaluation protocol, we ran diagnostic pruning sweeps on the OMat24 perovskite models at 1000 and 3000 K, the MPF2021 F–Fe–O model, and the MD-17 Aspirin model. These experiments use the same loader both to estimate gradient importance and to evaluate the pruned masks. They are retained to document how the pruning strategy behaves across different chemical domains, but they do not provide an independent estimate of mask generalisation.
The descriptive curves are shown in Figure 9.
Figure 9. Exploratory same-loader pruning curves, normalized by each model’s unpruned force MAE and shown on a logarithmic vertical scale. At 60% pruning, the relative changes are 0.02% for OMat24 perovskite at 1000 K, 4.05% at 3000 K, and 0.60% for MPF2021 F–Fe–O. Aspirin is qualitatively different: degradation begins between 20% and 25% pruning and reaches 13.81× the baseline MAE at 60%. Because ranking and evaluation reuse the same loader, these curves are descriptive diagnostics only.
For each trained model, all 174 weighted layer–instruction paths are enumerated. If denotes the edge-MLP final-linear block associated with path , its diagnostic score is the Frobenius norm of the accumulated force-MAE gradient over the first ten mini-batches,
The paths are sorted in ascending and cumulative prefixes corresponding to are zeroed without retraining. The original checkpoint weights are restored before every value of . The intervention zeros the relevant edge-MLP columns while leaving the dense tensor-product instruction list unchanged, so the reported operation reductions remain structural proxies rather than measured speedups.
| System | Ranking/evaluation source | Batch size | Maximum |
| OMat24 perovskite, 1000 K | same exported test loader | 10 | 60% |
| OMat24 perovskite, 3000 K | same exported test loader | 10 | 60% |
| MPF2021 F–Fe–O | same 224-frame test loader | 10 | 60% |
| Aspirin (MD-17) | same molecular test loader | 20 | 60% |
Table 6. Data roles in the exploratory pruning sweeps. “Same loader” means that the data used to form the gradient ranking are also used to compute the pruning curve. All rankings use ten mini-batches.
References
- [1]Brandon Anderson, Truong Son Hy, and Risi Kondor. Cormorant: Covariant molecular neural networks. In Advances in Neural Information Processing Systems, volume 32, 2019.arxiv.org/abs/1906.04015
- [2]Luis Barroso-Luque, Muhammed Shuaibi Muhammed, Xiang Fu, Brandon M. Wood, Misko Dzamba, Jingyu Guo, Matthieu Schmid, C. Lawrence Zitnick, and Zachary W. Ulissi. Open Materials 2024 (OMat24) inorganic materials dataset and models. arXiv preprint arXiv:2410.12771, 2024.
- [3]Ilyes Batatia, Dávid Péter Kovács, Gregor N. C. Simm, Christoph Ortner, and Gábor Csányi. MACE: Higher order equivariant message passing neural networks for fast and accurate force fields. In Advances in Neural Information Processing Systems, volume 35, 2022.
- [4]Simon Batzner, Albert Musaelian, Lixin Sun, Mario Geiger, Jonathan P. Mailoa, Mordechai Kornbluth, Nicola Molinari, Tess E. Smidt, and Boris Kozinsky. E(3)-equivariant graph neural networks for data-efficient and accurate interatomic potentials. Nature Communications, 13:2453, 2022. doi: 10.1038/s41467-022-29939-5.DOI
- [5]Chi Chen and Shyue Ping Ong. A universal graph deep learning interatomic potential for the periodic table. Nature Computational Science, 2:718–728, 2022.DOI
- [6]Stefan Chmiela, Alexandre Tkatchenko, Huziel E. Sauceda, Igor Poltavsky, Kristof T. Schütt, and Klaus-Robert Müller. Machine learning of accurate energy-conserving molecular force fields. Science Advances, 3(5):e1603015, 2017. doi: 10.1126/sciadv.1603015.DOI
- [7]Taco Cohen and Max Welling. Group equivariant convolutional networks. In International Conference on Machine Learning, pp. 2990–2999, 2016.arxiv.org/abs/1602.07576
- [8]Ralf Drautz. Atomic cluster expansion for accurate and transferable interatomic potentials. Physical Review B, 99(1):014104, 2019. doi: 10.1103/PhysRevB.99.014104.DOI
- [9]J. Thorben Franck, Oliver T. Unke, and Klaus-Robert Müller. So3krates: Equivariant attention for interactions on arbitrary length-scales in molecular systems. In Advances in Neural Information Processing Systems, volume 35, 2022.arxiv.org/abs/2205.14276
- [10]Jonathan Frankle and Michael Carlin. The lottery ticket hypothesis: Finding sparse, trainable neural networks. In International Conference on Learning Representations, 2019.
- [11]Mario Geiger and Tess Smidt. e3nn: Euclidean neural networks. arXiv preprint arXiv:2207.09453, 2022.arxiv.org/abs/2207.09453
- [12]Alexander J. Heilman and Qimin Yan. Point group equivariant graph neural networks for materials. arXiv preprint arXiv:2607.16871, 2026.arxiv.org/abs/2607.16871
- [13]Risi Kondor and Shubhendu Trivedi. On the generalization of equivariance and convolution in neural networks to the action of compact groups. In International Conference on Machine Learning, pp. 2747–2755, 2018.arxiv.org/abs/1802.03690
- [14]Yann LeCun, John Denker, and Sara Solla. Optimal brain damage. In Advances in Neural Information Processing Systems, volume 2, 1989.
- [15]Seung Yul Lee, Hojoon Kim, Yutack Park, Dawoon Jeong, Seungwu Han, Yeonhong Park, and Jae W. Lee. FlashTP: Fused, sparsity-aware tensor product for machine learning interatomic potentials. In Proceedings of the 42nd International Conference on Machine Learning, volume 267 of Proceedings of Machine Learning Research, pp. 33143–33156, 2025.
- [16]Yi-Lun Liao, Brandon Wood, Abhishek Das, and Tess Smidt. EquiformerV2: Improved equivariant transformer for scaling to higher-degree representations. In International Conference on Learning Representations, 2024.
- [17]Erpai Luo, Xinran Wei, Lin Huang, Yunyang Li, Han Yang, Zaishuo Xia, Zun Wang, Chang Liu, Bin Shao, and Jia Zhang. Efficient and scalable density functional theory hamiltonian prediction through adaptive sparsity. In Proceedings of the 42nd International Conference on Machine Learning, volume 267 of Proceedings of Machine Learning Research, pp. 41368–41390, 2025.
- [18]Pavlo Molchanov, Arun Mallya, Stephen Tyree, Iuri Frosio, and Jan Kautz. Importance estimation for neural network pruning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp. 11264–11272, 2019.arxiv.org/abs/1906.10771
- [19]Albert Musaelian, Simon Batzner, Anders Johansson, Lixin Sun, Cameron J. Owen, Mordechai Kornbluth, and Boris Kozinsky. Learning local equivariant representations for large-scale atomistic dynamics. Nature Communications, 14:579, 2023. doi: 10.1038/s41467-023-36329-y.
- [20]Jigyasa Nigam, Sergey Pozdnyakov, Gábor Csányi, and Michele Ceriotti. Unified theory of atom-centered representations and message-passing machine-learning schemes. The Journal of Chemical Physics, 156(20):204115, 2022. doi: 10.1063/5.0087042.
- [21]Yutack Park, Jaesun Kim, Seungwoo Hwang, and Seungwoo Han. SevenNet: A interatomic potential benchmark for bulk phase materials. arXiv preprint arXiv:2402.03789, 2024.
- [22]Saro Passaro and C. Lawrence Zitnick. Reducing SO(3) convolutions to SO(2) for efficient equivariant GNNs. In International Conference on Machine Learning, pp. 27420–27438, 2023.
- [23]Sergey N. Pozdnyakov, Michael J. Willatt, Albert P. Bartók, Christoph Ortner, Gábor Csányi, and Michele Ceriotti. Incompleteness of atomic structure representations. Physical Review Letters, 125(16):166001, 2020. doi: 10.1103/PhysRevLett.125.166001.DOI
- [24]Nathaniel Thomas, Tess Smidt, Steven Kearnes, Lusann Yang, Li Li, Kai Kohlhoff, and Patrick Riley. Tensor field networks: Rotation- and translation-equivariant neural networks for 3d point clouds. arXiv preprint arXiv:1802.08219, 2018.
- [25]Wu-Ki Tung. Group Theory in Physics. World Scientific, 1985.
- [26]D. A. Varshalovich, A. N. Moskalev, and V. K. Khersonskii. Quantum Theory of Angular Momentum. World Scientific, 1988.DOI
- [27]Eugene P. Wigner. Group Theory and its Application to the Quantum Mechanics of Atomic Spectra. Academic Press, 1959.
- [28]Yuqing Xie, Ameya Daigavane, Mit Kotak, and Tess E. Smidt. The price of freedom: Exploring expressivity and runtime tradeoffs in equivariant tensor products. In Proceedings of the 42nd International Conference on Machine Learning, volume 267 of Proceedings of Machine Learning Research, pp. 68599–68625, 2025.
- [29]Han Yang, Chenxi Hu, Yichi Zhou, Xixian Liu, Yu Shi, Jielan Li, Guanzhi Li, Zekun Chen, Shuizhou Chen, Claudio Zeni, et al. Mattersim: A deep learning atomistic model across elements, temperatures and pressures. arXiv preprint arXiv:2405.04967, 2024.







