betaFeedback

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-ll representation has no trivial component, this sum and the corresponding path vanish for any learned radial weight. For an ideal OhO_h octahedral orbit, this removes the ledge=1,2,3l_\mathrm{edge}=1,2,3 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 TiO6_6 probes examine first-layer cancellation, whereas a within-model CaTiO3_3 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 CaTiO3_3, OMat24 OhO_h-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 linl_{\mathrm{in}} with an edge harmonic of degree ledgel_{\mathrm{edge}} to produce an output feature loutl_{\mathrm{out}}. 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 lmax⁡=3l_{\max} = 3.

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 Sl=∑jYl(r^ij)S_l = \sum_j Y_l(\hat{r}_{ij}). Because the site group only permutes the summands, SlS_l lies in the trivial subspace of the restricted degree-ll 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 OhO_h octahedral orbit, the l=1,2,3l = 1, 2, 3 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 TiO6_6 probes test first-layer cancellation and parity-dependent activation under controlled distortions. A within-model CaTiO3_3 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 1Gi1_{G_i}, for exact first-layer TP-path cancellation under complete equal-species, equal-radius neighbor orbits.

  • Controlled TiO6_6 probes test the exact first-layer result, while a within-model CaTiO3_3 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 E(3)E(3)-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 ii and a weighted TP instruction p=(lin,ledge,lout)p=(l_{\mathrm{in}},l_{\mathrm{edge}},l_{\mathrm{out}}). Suppressing channel indices, its message is

Mi,mlout=∑j∈N(i)wp(rij)∑min,medgeCmin,medgemhj,minlinYledgemedge(r^ij).(1)M^{l_{\mathrm{out}}}_{i,m} = \sum_{j\in\mathcal{N}(i)} w_p(r_{ij}) \sum_{m_{\mathrm{in}},m_{\mathrm{edge}}} C^{m}_{m_{\mathrm{in}},m_{\mathrm{edge}}} h^{l_{\mathrm{in}}}_{j,m_{\mathrm{in}}} Y^{m_{\mathrm{edge}}}_{l_{\mathrm{edge}}}(\hat{r}_{ij}). \tag*{(1)}

Here wpw_p is produced by the edge MLP, CC is a Clebsch–Gordan coefficient [27, 26], and YY is a spherical harmonic. Let GiG_i be the site-symmetry group of the center and partition a complete neighbor shell into orbits Oa\mathcal{O}_a under GiG_i. Within one orbit, symmetry preserves radius and species.

Proposition 1 (Orbit factorisation). For a first-layer path with lin=0l_{\mathrm{in}}=0, the contribution of orbit Oa\mathcal{O}_a factorises as

Mi,mlout(Oa)=αa,pSledge,m(a),Sl,m(a)=∑j∈OaYlm(r^ij),(2)M^{l_{\mathrm{out}}}_{i,m}(\mathcal{O}_a)=\alpha_{a,p}S^{(a)}_{l_{\mathrm{edge},m}}, \qquad S^{(a)}_{l,m}=\sum_{j\in\mathcal{O}_a}Y^m_l(\hat{r}_{ij}), \tag*{(2)}

where αa,p\alpha_{a,p} contains the common species feature, radial weight, channel mixing, and CG factor. Hence Sl(a)=0S^{(a)}_l=0 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 wp(ra)w_p(r_a). This is why the cancellation is architectural-data interaction rather than a coincidental small learned weight.

Proposition 2 (Site-symmetry criterion). The orbit sum Sl(a)S^{(a)}_l lies in Inv⁡Gi(Dl↓Gi)\operatorname{Inv}_{G_i}(D^l\downarrow G_i), the invariant subspace of the spherical representation restricted to GiG_i. Therefore

1Gi⊄Dl↓Gi⟹Sl(a)=0,(3)\mathbf{1}_{G_i}\not\subset D^l\downarrow G_i \quad\Longrightarrow\quad S^{(a)}_l=0, \tag*{(3)}

where 1Gi\mathbf{1}_{G_i} is the trivial irrep of the relevant group.

For every g∈Gig\in G_i, applying Dl(g)D^l(g) to the sum rotates each spherical harmonic and merely permutes the members of Oa\mathcal{O}_a. Thus Dl(g)Sl(a)=Sl(a)D^l(g)S^{(a)}_l=S^{(a)}_l. Equivalently, the Reynolds projector ∣Gi∣−1∑g∈GiDl(g)|G_i|^{-1}\sum_{g\in G_i}D^l(g) 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 OhO_h orbit and their spherical-harmonic sums vanish for l=1,2,3l = 1, 2, 3. A lower-symmetry distortion partitions the shell into smaller orbits and can activate previously forbidden even components. Exact odd-ll cancellation survives whenever an orbit retains inversion-related pairs; thermal disorder, non-centrosymmetric sites, or deeper features can relax it.

Concrete Site Groups

For OhO_h, the trivial irrep is conventionally named A1gA_{1g}. The restrictions through l=3l = 3 are

D0↓Oh=A1g,D1↓Oh=T1u,D2↓Oh=Eg⊕T2g,D3↓Oh=A2u⊕T1u⊕T2u.(4)\begin{aligned} D^0 \downarrow O_h &= A_{1g}, & D^1 \downarrow O_h &= T_{1u},\\ D^2 \downarrow O_h &= E_g \oplus T_{2g}, & D^3 \downarrow O_h &= A_{2u} \oplus T_{1u} \oplus T_{2u}. \tag*{(4)} \end{aligned}

Only l=0l = 0 contains A1gA_{1g}, so an ideal six-neighbor octahedral orbit transmits only the scalar first-layer component for l≤3l \le3. For a general group we retain the notation 1Gi\mathbf{1}_{G_i}: writing A1gA_{1g} for every group would incorrectly assume both the OhO_h naming convention and inversion parity.

The tetrahedral group TdT_d illustrates the distinction between permission and importance. Its trivial irrep is A1A_1, and the degree-three representation contains an A1A_1 component. A tetrahedral orbit can therefore have a nonzero l=3l = 3 sum even though its l=1l = 1 and l=2l = 2 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 Ylm(−r^)=(−1)lYlm(r^)Y_l^m(-\hat{r}) = (-1)^lY_l^m(\hat{r}), so all odd-ll sums cancel. This remains true for a centrosymmetric distorted orbit even when the full OhO_h 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 TiO6_6 probes in trained CaTiO3_3 and OMat24 OhO_h 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 hgjl′=Dl′(g)hjlh_{gj}^{l'} = D^{l'}(g)h_j^l rather than sharing one scalar value. The feature and edge harmonic are then coupled inside the orbit sum, and the common factor αa,p\alpha_{a,p} in Equation equation 2 cannot generally be extracted. A direct e3nne3nn enumeration on an ideal OhO_h 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-OhO_h data should show a lower high-ll importance profile than a matched lower-symmetry condition, while no universal ordering of l=1,2,3l = 1, 2, 3 follows. This distinction determines the experimental design below.

Experimental Design

Models, Intervention, and Metrics

All models use NequIP 0.17.1 with lmax⁡=3l_{\max} = 3, 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 pp, we report

ΔMAE^p=MAEp-ablated−MAEbaseMAEbase.(5)\widehat{\Delta\mathrm{MAE}}_p = \frac{\mathrm{MAE}_{p\text{-ablated}} - \mathrm{MAE}_{\mathrm{base}}}{\mathrm{MAE}_{\mathrm{base}}}. \tag*{(5)}

This measures functional sensitivity of a trained model; because the dense instruction list and tensors are unchanged, ablation alone does not reduce runtime.

The CaTiO3_3 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 Pm3ˉ\bar{3}m. 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: OhO_h-cubic, TdT_d-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 pp with edge-MLP block WpW_p, we accumulate force-loss gradients over B=10B = 10 training mini-batches and define

Ip=∥1B∑b=1B∂LF(b)∂Wp∣F∥.(6)I_p = \left\| \left.\frac{1}{B}\sum_{b=1}^{B}\frac{\partial\mathcal{L}_{F}^{(b)}}{\partial W_p}\right|_{F} \right\|. \tag*{(6)}

Paths are sorted in ascending IpI_p. For each cumulative fraction q∈{0,5,10,15,20,25,30,40,50,60}%q \in\{0, 5, 10, 15, 20, 25, 30, 40, 50, 60\}\%, 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 CaTiO3_3, 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 CaTiO3_3 Importance Shift

Figure 2 groups the 174 single-path ablations by ledgel_{\mathrm{edge}}. Relative to the near-cubic 2000 K endpoint, the lower-symmetry 1000 K endpoint has larger mean normalized ablation for every ledge≥1l_{\mathrm{edge}} \ge1, with endpoint ratios r1=1.38r_1 = 1.38, r2=1.07r_2 = 1.07, and r3=1.69r_3 = 1.69. 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. CaTiO3_3 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 ledge≥1l_{\mathrm{edge}} \ge1 in the lower-symmetry endpoint is the controlled test of the predicted importance bias; the theory does not predict an ordering among l=1,2,3l = 1, 2, 3.

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 OhO_h-cubic model has the smallest mean normalized ablation for l=1,2,3l = 1, 2, 3 (0.38%, 0.30%, and 0.34%), consistent with a strong first-layer octahedral bottleneck. The TdT_d-cubic model instead reaches 3.17%, 2.77%, and 3.36%. In particular,

Exp B: TP path importance by symmetry class, showing mean normalized ablation MAE by symmetry class and angular momentum

Figure 3. OMat24 mean normalized ablation for models stratified by equilibrium space-group class. The OhO_h-cubic subset has the lowest high-ll importance, while the TdT_d 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.

l=3l = 3 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 CaTiO3\mathrm{CaTiO}_3 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 CaTiO3\mathrm{CaTiO}_3, +0.04% for OMat24 OhO_h-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 meV/A˚\mathrm{meV}/\text{\AA}, respectively. For CaTiO3\mathrm{CaTiO}_3, 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.

SystemFramesBasePrunedΔ\Delta (%)TP (%)MLP (%)
CaTiO3_3 endpoints10/36/3923.6824.021.4271.9459.77
OMat24 OhO_h10/200/5000229.19229.280.0467.1259.77
OMat24 low-sym.10/200/5000167.22170.271.8368.4559.77

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 OhO_h 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% (OhO_h) 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 OhO_h and low-symmetry masks overlap on 89 of 104 paths (Jaccard index 0.75), while CaTiO3\mathrm{CaTiO}_3 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 CaTiO3\mathrm{CaTiO}_3 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 p=(lin,ledge,lout)p = (l_{\mathrm{in}}, l_{\mathrm{edge}}, l_{\mathrm{out}}) with multiplicity CpC_p, define the dimension-weighted dense-contraction proxy

ϕp=Cp(2lin+1)(2ledge+1)(2lout+1),RTP=∑p∈Pϕp∑pϕp.(7)\phi_p = C_p(2l_{\mathrm{in}} + 1)(2l_{\mathrm{edge}} + 1)(2l_{\mathrm{out}} + 1), \qquad R_{\mathrm{TP}} = \frac{\sum_{p \in\mathcal{P}} \phi_p}{\sum_p \phi_p}. \tag*{(7)}

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

RMLP=∑p∈PHn(p)Cp∑pHn(p)Cp=59.77%,(8)R_{\mathrm{MLP}} = \frac{\sum_{p \in\mathcal{P}} H_n(p) C_p}{\sum_p H_n(p) C_p} = 59.77\%, \tag*{(8)}

where Hn(p)H_n(p) 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, CaTiO3_3 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 jj of species sas_a carries the scalar feature hj0,0=hsah_j^{0,0}=h_{s_a}. For all jj in a site-symmetry orbit Oa\mathcal{O}_a, the distance rij=rar_{ij}=r_a and species are constant. Restricting Equation (1) to this orbit and using min=0m_{\mathrm{in}}=0 gives

Mi,mlout(Oa)=w(ra)hsa∑j∈Oa∑medgeC0,medgemYledgemedge(r^ij)(9)M_{i,m}^{l_{\mathrm{out}}}(\mathcal{O}_a)=w(r_a)h_{s_a}\sum_{j\in\mathcal{O}_a}\sum_{m_{\mathrm{edge}}}C_{0,m_{\mathrm{edge}}}^{m}Y_{l_{\mathrm{edge}}}^{m_{\mathrm{edge}}}(\hat{r}_{ij}) \tag*{(9)}
=αaSledge,m(a),(10)=\alpha_a S_{l_{\mathrm{edge}},m}^{(a)}, \tag*{(10)}

where the CG selection rule fixes m=medgem=m_{\mathrm{edge}} for lin=0l_{\mathrm{in}}=0. Thus the learned radial function enters only through the common scalar αa\alpha_a and cannot undo a vanishing angular sum.

Invariant-Subspace Criterion

Let GiG_i act on the directions in Oa\mathcal{O}_a. For every g∈Gig\in G_i,

Dl(g)Sl(a)=∑j∈OaDl(g)Yl(r^ij)=∑j∈OaYl(gr^ij)=Sl(a),(11)D^l(g)S_l^{(a)}=\sum_{j\in\mathcal{O}_a}D^l(g)Y_l(\hat{r}_{ij})=\sum_{j\in\mathcal{O}_a}Y_l(g\hat{r}_{ij})=S_l^{(a)}, \tag*{(11)}

because gg permutes orbit members. Hence Sl(a)S_l^{(a)} belongs to Inv⁡Gi(Dl∣Gi)\operatorname{Inv}_{G_i}(D^l|_{G_i}). Equivalently, the Reynolds projector PGi(l)=∣Gi∣−1∑g∈GiDl(g)P_{G_i}^{(l)}=|G_i|^{-1}\sum_{g\in G_i}D^l(g) satisfies PGi(l)Sl(a)=Sl(a)P_{G_i}^{(l)}S_l^{(a)}=S_l^{(a)}. If the trivial representation 1Gi\mathbf{1}_{G_i} is absent from the restricted representation, this projector is zero and Sl(a)=0S_l^{(a)}=0.

For a centrosymmetric orbit, every direction has an opposite partner and Ylm(−r^)=(−1)lYlm(r^)Y_l^m(-\hat{r})=(-1)^lY_l^m(\hat{r}); all odd-ll sums therefore vanish. This parity argument is a special case of the invariant-subspace criterion. The notation A1gA_{1g} is used only for centrosymmetric point groups such as OhO_h; for TdT_d the trivial irrep is A1A_1, and 1Gi\mathbf{1}_{G_i} is the group-independent notation.

llDl↓OhD^l\downarrow O_hContains A1gA_{1g}?Ideal orbit sum
0A1gA_{1g}yesnon-zero allowed
1T1uT_{1u}nozero
2Eg⊕T2gE_g\oplus T_{2g}nozero
3A2u⊕T1u⊕T2uA_{2u}\oplus T_{1u}\oplus T_{2u}nozero

Table 2. Restrictions of DlD^l to OhO_h for $l\leq 3.

Pnma, TdT_d, and Deeper Layers

Octahedral tilting in Pnma splits the ideal OhO_h oxygen shell into smaller orbits and can activate even-ll components that OhO_h forbids. The ideal Ti site nevertheless remains centrosymmetric, so opposite-pair cancellation still removes odd ll in its first-layer scalar orbit sum. Odd-ll 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 CaTiO3_3 l=3l=3 ratio as a direct proof of first-layer Ti-site activation.

For a tetrahedral orbit, TdT_d symmetry forbids l=1,2l=1,2 but permits a trivial component at l=3l=3. Lower-symmetry groups permit progressively more components, although permission alone does not determine their learned magnitude. In deeper NequIP layers, symmetry-related neighbors obey hj′l=Dl(g)hjlh_{j'}^l=D^l(g)h_j^l rather than carrying identical scalars. A direct e3nn calculation on an ideal OhO_h 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, lmax⁡=3l_{\max}=3, and rmax⁡=5.0r_{\max}=5.0 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.

CaTiO3_3. 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 ABO3_3 frames at 1000 and 3000 K. For the space-group study, structures are partitioned into OhO_h-cubic (space groups 221–230), TdT_d-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, lmax⁡=3l_{\max}=3, 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 CaTiO3_3, 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 CaTiO3_3 manifest records file hashes and a temperature-aware split. Exact structure hashes for the OMat24 OhO_h 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 dl=2l+1d_l=2l+1 and

ϕp=Cpdlindledgedlout,RTP=∑p∈Pϕp∑pϕp.(12)\phi_p=C_p d_l^{\mathrm{in}}d_{l_{\mathrm{edge}}}d_l^{\mathrm{out}},\qquad R_{\mathrm{TP}}=\frac{\sum_{p\in\mathcal{P}}\phi_p}{\sum_p\phi_p}. \tag*{(12)}

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 HnH_n,

RMLP=∑p∈PHn(p)Cp∑pHn(p)Cp.(13)R_{\mathrm{MLP}}=\frac{\sum_{p\in\mathcal{P}}H_n(p)C_p}{\sum_p H_n(p)C_p}. \tag*{(13)}

Algorithm 2: Gradient-Ranked Cumulative TP Path Pruning

Input: Trained model M\mathcal{M}; training-ranking, validation-selection, and held-out test loaders Drank\mathcal{D}_{\mathrm{rank}}, Dselect\mathcal{D}_{\mathrm{select}}, Dtest\mathcal{D}_{\mathrm{test}}; pruning fractions Q\mathcal{Q}; tolerance ϵ\epsilon; ranking batches BB

Output: Frozen mask Pq∗\mathcal{P}_{q^*}; one held-out test record

  1. E←\mathcal{E} \leftarrow collect all weighted (layer, TP instruction) entries; foreach p∈Ep \in\mathcal{E} do

  2. locate its edge-MLP column block WpW_p; compute on Drank\mathcal{D}_{\mathrm{rank}}: Ip←∥B−1∑b=1B∂LF(b)/∂Wp∥FI_p \leftarrow\left\lVert B^{-1}\sum_{b=1}^{B}\partial\mathcal{L}_{F}^{(b)}/\partial W_p\right\rVert_F; compute ϕp←Cp(2lin+1)(2ledge+1)(2lout+1)\phi_p \leftarrow C_p(2l_{\mathrm{in}}+1)(2l_{\mathrm{edge}}+1)(2l_{\mathrm{out}}+1);

  3. end

  4. π←argsort⁡p∈EIp\pi\leftarrow\operatorname{argsort}_{p\in\mathcal{E}} I_p in ascending order; Θ0←\Theta_0 \leftarrow clone original edge-MLP weights;

  5. MAE⁡0select←EvalForceMAE⁡(M,Dselect)\operatorname{MAE}_{0}^{\mathrm{select}} \leftarrow\operatorname{EvalForceMAE}(\mathcal{M},\mathcal{D}_{\mathrm{select}}); foreach q∈Qq \in\mathcal{Q} do

  6. restore Θ0\Theta_0; Kq←round⁡(q∣E∣/100)K_q \leftarrow\operatorname{round}(q|\mathcal{E}|/100); Pq←{π1,…,πKq}\mathcal{P}_q \leftarrow\{\pi_1,\ldots,\pi_{K_q}\}; foreach p∈Pqp \in\mathcal{P}_q do

  7. Wp←0W_p \leftarrow0

  8. end

  9. MAE⁡select(q)←EvalForceMAE⁡(M,Dselect)\operatorname{MAE}^{\mathrm{select}}(q) \leftarrow\operatorname{EvalForceMAE}(\mathcal{M},\mathcal{D}_{\mathrm{select}});

  10. end

  11. q∗←max⁡{q:∀q′∈Q, q′≤q⇒MAE⁡select(q′)≤(1+ϵ)MAE⁡0select}q^* \leftarrow\max\{q : \forall q' \in\mathcal{Q},\ q' \le q \Rightarrow\operatorname{MAE}^{\mathrm{select}}(q') \le(1+\epsilon)\operatorname{MAE}_{0}^{\mathrm{select}}\}; restore Θ0\Theta_0; apply the frozen mask Pq∗\mathcal{P}_{q^*}; evaluate Dtest\mathcal{D}_{\mathrm{test}} exactly once and compute operation reductions; restore Θ0\Theta_0 and return Pq∗\mathcal{P}_{q^*} 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 TiO6_6 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 ±x^\pm\hat{x}, ±y^\pm\hat{y}, and ±z^\pm\hat{z}. At each angle θ∈{2∘,5∘,10∘,15∘}\theta\in\{2^\circ, 5^\circ, 10^\circ, 15^\circ\}, the inversion-preserving case rotates the +z+z oxygen by θ\theta toward +x+x and moves its opposite partner to the negative of the new direction. The inversion-broken case moves only the +z+z oxygen. Species, bond lengths, and the number of incoming Ti–O edges are unchanged.

We compute the norm of the degree-ll 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 ll and divided in Figure 6 by the model’s ideal-geometry l=0l=0 norm. For both the CaTiO3_3 and OMat24 OhO_h checkpoints, the ideal l=1,2,3l=1,2,3 responses are below 3.1×10−83.1 \times10^{-8} in raw norm. Under inversion-preserving tilts, the odd-ll responses remain below 2.1×10−82.1 \times10^{-8} while l=2l=2 increases with angle. Moving only one oxygen also activates l=1l=1 and l=3l=3. 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 l=0l = 0 norm. Solid lines break inversion; the dashed line preserves opposite oxygen pairs. In the latter case, the l=1,3l = 1, 3 responses remain at numerical zero and are omitted from the plotted series. All six Ti–O bonds remain at 2.0 Å.

CaTiO3 Ratios

ledgel_{\mathrm{edge}}0123
Paths across five layers21465651
rlr_l1.231.381.071.69

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-ll ablation, whereas thermal disorder at 3000 K activates the same path groups.

ledgel_{\mathrm{edge}}OhO_h-cubicTdT_d-cubicmid-sym.low-sym.
020.377.748.521.1
10.383.171.880.96
20.302.771.640.50
30.343.361.260.66

Table 4. OMat24 mean normalized ablation ΔMAE‾\overline{\Delta\mathrm{MAE}} (%).

Additional Pruning Results

SystemValidationTestTest frames
CaTiO3_3 endpoints1.1361.41839
OMat24 OhO_h-cubic0.0340.0405000
OMat24 low-symmetry1.9311.8255000

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, CaTiO3_3 removes 11/41/52 paths from layers 1/2/3, OhO_h removes 16/43/45, and low-symmetry removes 7/40/57; none removes a layer-0 instruction. The OhO_h/low-symmetry masks share 89 paths (Jaccard 0.748), while the CaTiO3_3 mask has Jaccard overlaps 0.809 and 0.874 with OhO_h 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 OhO_h force-MAE increase is 0.040–0.065% (mean 0.055 ±\pm 0.013%), and the low-symmetry increase is 1.825–2.241% (mean 2.084 ±\pm 0.226%). The ±\pm values are sample standard deviations across three masks, not test-set confidence intervals. Layer-matched random masks give means of 32.00 ±\pm 4.40% and 124.99 ±\pm 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 lmax⁡=3l_{\max}=3 model, all 174 weighted layer–instruction paths are enumerated. If WpW_p denotes the edge-MLP final-linear block associated with path pp, its diagnostic score is the Frobenius norm of the accumulated force-MAE gradient over the first ten mini-batches,

Ipdiag=∥∑b=110∂LC(b)∂Wp∥F.(14)I_p^{\mathrm{diag}}=\left\|\sum_{b=1}^{10}\frac{\partial\mathcal{L}_{\mathrm{C}}^{(b)}}{\partial W_p}\right\|_F . \tag*{(14)}

The paths are sorted in ascending IpdiagI_p^{\mathrm{diag}} and cumulative prefixes corresponding to q∈{0,5,10,15,20,25,30,40,50,60}%q \in\{0, 5, 10, 15, 20, 25, 30, 40, 50, 60\}\% are zeroed without retraining. The original checkpoint weights are restored before every value of qq. 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.

SystemRanking/evaluation sourceBatch sizeMaximum qq
OMat24 perovskite, 1000 Ksame exported test loader1060%
OMat24 perovskite, 3000 Ksame exported test loader1060%
MPF2021 F–Fe–Osame 224-frame test loader1060%
Aspirin (MD-17)same molecular test loader2060%

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. [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. [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. [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. [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. [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. [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. [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. [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. [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. [10]Jonathan Frankle and Michael Carlin. The lottery ticket hypothesis: Finding sparse, trainable neural networks. In International Conference on Learning Representations, 2019.
  11. [11]Mario Geiger and Tess Smidt. e3nn: Euclidean neural networks. arXiv preprint arXiv:2207.09453, 2022.arxiv.org/abs/2207.09453
  12. [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. [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. [14]Yann LeCun, John Denker, and Sara Solla. Optimal brain damage. In Advances in Neural Information Processing Systems, volume 2, 1989.
  15. [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. [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. [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. [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. [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. [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. [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. [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. [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. [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. [25]Wu-Ki Tung. Group Theory in Physics. World Scientific, 1985.
  26. [26]D. A. Varshalovich, A. N. Moskalev, and V. K. Khersonskii. Quantum Theory of Angular Momentum. World Scientific, 1988.DOI
  27. [27]Eugene P. Wigner. Group Theory and its Application to the Quantum Mechanics of Atomic Spectra. Academic Press, 1959.
  28. [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. [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.

Paper details

vlnc:W_7rt3ge7a
Version
v1 · latest
Published
Fresh — added to the Hub within the last 30 days
Added
License
CC BY 4.0
Usage

views · 0

downloads · 0

A view counts once per browser per day, a download each time the PDF is requested.

  • Physics and Astronomy

Generative AI tools

View statement
Looking for related papers…
Contents