betaFeedback

WorksPhysics and Astronomy

Rotation-Invariant Machine Learning for Highly Symmetric Molecular Configurations

Abstract

Atomistic machine learning can approximate expensive quantum mechanical calculations at a fraction of their computational cost, but accuracy depends strongly on how the three-dimensional atomic environments are represented. Invariant representations enable fast computation but can fail to distinguish geometrically distinct configurations, especially in highly symmetric environments. We introduce a lightweight benchmark to evaluate expressivity on nn-fold symmetric configurations, and use it to expose degeneracies in both invariant and equivariant architectures. Guided by these results, we construct a new invariant basis for HIP-HOP which resolves most of the identified failures. On a methane simulation dataset, the extended basis achieves the lowest error among the HIP-HOP variants and has similar accuracy to EquiformerV3 while running 3.2×3.2\times faster at inference. These results show that carefully constructed invariant representations can mitigate key expressivity limitations while balancing accuracy and computational efficiency.

Editors: List of editors’ names

© 2026 L. Bolton, T. Karella, N. Lubbers, N.P. Gorski, E. Shinkle, A.E.A. Allen, P.J. Swart & R. Bujack.

Introduction

Accurate atomistic simulation is central to understanding material properties, chemical reactions, and biological processes. While first principles quantum mechanical (QM) calculations [11, 16, 23] provide highly accurate results, their computational cost makes routine large-scale and long-timescale simulations challenging. Machine-learned interatomic potentials (MLIPs) have therefore emerged as efficient surrogates, reproducing QM-level accuracy at a fraction of the cost [4, 21]. Recent large-scale models such as UMA demonstrate how far this paradigm has advanced, scaling to hundreds of millions of training structures while enabling fast inference across diverse chemical systems [28].

A key factor behind the success of MLIPs is the incorporation of physical symmetries as inductive biases. For example, energy predictions are invariant to translations, rotations, and permutations of identical atoms, while forces transform along with the geometric transformations. Encoding these properties directly into neural architectures has become a cornerstone of geometric deep learning Bronstein et al. [5]; Cohen and Welling [7]; Weiler et al. [26], leading to substantial gains in accuracy and data efficiency.

These approaches can be broadly divided into two strategies for incorporating symmetry, each with different tradeoffs. Invariant models [20, 10, 24, 1] use scalar geometric quantities such as distances, angles, and inner products, enabling highly efficient computation in a reduced representation space, but potentially limiting expressivity if the invariants are not sufficiently rich to describe geometry. In contrast, equivariant models [17, 14, 22, 25, 3 a] retain vector and tensor-valued features that transform consistently under symmetry operations, preserving directional information and improving expressivity. This comes at increased computational cost due to operations on multi-component tensor representations.

This raises a central question: is explicit equivariant machinery strictly necessary for all tasks? Invariant representations are not inherently lossy; rather, their adequacy depends on whether they are complete, i.e., whether they distinguish all molecular configurations up to the prescribed symmetries. This suggests an alternative design point, namely constructing complete invariant representations that retain all symmetry-relevant geometric information while enabling fast computation.

In this work, we introduce a new basis for HIP-HOP, an invariant architecture combining completeness with computational efficiency [1]. Using an angular expansion up to ℓmax⁡=4\ell_{\max}=4, the representation captures finer angular structure while retaining invariance. The extended basis resolves the identified degeneracies and improves predictive accuracy. This new model has similar accuracy to EquiformerV3 while offering higher inference throughput, though EquiformerV3 remains more accurate at larger model sizes.

Method

We interpret the local atomistic configuration as a density on a sphere and describe its shape by projecting it onto a polynomial basis truncated at a maximal angular order ℓmax⁡\ell_{\max}, following the theory of moment invariants [12] in the flexible formulation of Bujack et al. [6]. Intuitively, ℓmax⁡\ell_{\max} plays the same role as the truncation order of a Fourier or

Taylor series. Low orders capture only coarse angular features of the neighbor distribution, while higher orders resolve increasingly fine angular detail at increased computational cost. This mirrors the angular truncation used in atom-centered descriptors such as ACE [9], and in equivariant networks built on spherical harmonics [22]. The resulting projections, called geometric moments, form Cartesian tensors. Because products and contractions of tensors are again tensors, every rank-zero contraction of a product of moment tensors is a scalar and therefore a rotation invariant. Infinitely many such contractions exist and most are redundant, so we identify an independent subset using Langbein’s algorithm [15, 6] to obtain a compact basis.

Benchmarks We first evaluate HIP-HOP on the N-body benchmark [19, 13] to compare the ability of invariant and equivariant models to resolve geometric interactions. To complement the N-body benchmark, we introduce a lightweight synthetic dataset designed to evaluate highly symmetric local environments, including configurations with nn-fold symmetry. This benchmark provides a controlled setting that is less contrived than the N-body benchmark while remaining simple enough to isolate differences in model expressivity, and is motivated by known degeneracies of finite-order atom-centered invariant representations [6]. We construct graphs with a central node surrounded by two rings and vary their populations and relative rotation to generate configurations with different nn-fold symmetries. For each configuration, we consider orientations that minimize or maximize the distance between nodes on different rings, as illustrated in Figure 1. Additional details can be found in Appendix § 7.

Two rotated versions of a graph with close and far ring configurations

Figure 1. For the rings benchmark, the model is given two rotated versions of the same graph, and must determine which orientation has the nodes closer together.

Methane Dataset To test how basis expressivity affects real-world performance, we evaluate our model on the randomly displaced methane dataset of Pozdnyakov et al. [18], containing over seven million configurations with DFT energies and forces. Methane provides a challenging four-neighbor environment, making it a natural test of the additional geometric resolution introduced by extending the basis to ℓmax⁡=4\ell_{\max}=4.

Model We extend the original construction of HIP-HOP using the flexible moment-invariant framework of Bujack et al. [6] to obtain a new basis through ℓmax⁡=4\ell_{\max}=4. The resulting basis and its tensor-contraction structure are shown in Appendix § 9. Increasing the basis size improves model expressivity but also introduces higher-order polynomial terms and increased computational cost. To mitigate this overhead, we parallelize both the construction and evaluation of the invariant polynomials.

Results

N-body Benchmark Consistent with the theoretical predictions of Allen et al. [1], HIP-HOP with basis functions up to ℓmax⁡=3\ell_{\max}=3 successfully resolves interactions up to three-body and the expanded ℓmax⁡=4\ell_{\max}=4 basis does not improve the expressiveness (Figure 2). The static performance across all models and invariant bases motivates us to additionally evaluate with the rings benchmark to test changes in expressiveness.

N-body and rings benchmark accuracy matrices

Figure 2. The nn-body benchmark (A) accuracy is saturated for all bases, but the rings benchmark (B) shows the difference in expressivity between bases.

Rings Benchmark On the rings benchmark, the original HIP-HOP basis fails, as expected, to distinguish patterns involving four-fold symmetric configurations. Extending to ℓmax⁡=4\ell_{\max}=4 resolves most cases; however, both HIP-HOP and EquiformerV3 fail on the (3/4) configuration (Figure 2). This behavior appears consistent with the singular degeneracy recently described in Domina and Ceriotti [8], which we discuss further in Appendix § 10. Adding an additional contraction to the basis resolves these special configurations, though it did not provide measurable benefits in subsequent experiments (Appendix § 5).

Methane Dataset The new ℓmax⁡=4\ell_{\max}=4 basis consistently achieves the lowest errors among the small HIP-HOP configurations across all four metrics (Table 1 and Appendix § 5). At roughly 36K parameters, it nearly matches EquiformerV3 in accuracy while running 3.2×\times faster at inference. At 1.1M parameters, EquiformerV3 achieves lower error, but HIP-HOP remains 2.4×\times faster at inference. Parallelization further improves both the one-time basis construction and inference speed, making the new basis 1.57×\times faster (Appendix § 8).

Model / BasisPrediction PerformanceInference Speed
Modelℓmax⁡\ell_{\max}nmax⁡n_{\max}# Par.E MAEE RMSEF MAEF RMSEμ\mus/atomSpeedup
HIP-HOP3425 K0.21±0.020.21\pm0.020.50±0.350.50\pm0.350.61±0.030.61\pm0.032.63±2.562.63\pm2.562.281.07×1.07\times
HIP-HOP4324 K0.21±0.000.21\pm0.000.38±0.020.38\pm0.020.62±0.010.62\pm0.011.54±0.331.54\pm0.333.381.05×1.05\times
HIP-HOP4436 K0.18±0.00\mathbf{0.18}\pm0.000.32±0.010.32\pm0.010.55±0.010.55\pm0.011.24±0.03\mathbf{1.24}\pm0.034.421.57×1.57\times
EquiformerV34N/AN/A36 K0.18±N/A0.18\pm N/A0.29±N/A\mathbf{0.29}\pm N/A0.51±N/A\mathbf{0.51}\pm N/A1.25±N/A1.25\pm N/A14.3N/AN/A
HIP-HOP441.13 M0.04±N/A0.04\pm N/A0.11±N/A0.11\pm N/A0.17±N/A0.17\pm N/A0.61±N/A0.61\pm N/A12.71.72×1.72\times
EquiformerV34N/AN/A1.12 M0.03±N/A\mathbf{0.03}\pm N/A0.05±N/A\mathbf{0.05}\pm N/A0.10±N/A\mathbf{0.10}\pm N/A0.35±N/A\mathbf{0.35}\pm N/A30.2N/AN/A

Table 1. The largest (ℓmax⁡=4\ell_{\max}=4, nmax⁡=4n_{\max}=4) basis has the lowest prediction error of the HIP-HOP variants while also being significantly faster than EquiformerV3.

Discussion and Conclusion

We introduced a benchmark for systematically identifying failures of invariant and equivariant representations in highly symmetric atomic environments. Using this benchmark, we showed that low-order representations fail to distinguish several geometrically distinct configurations, whereas increasing ℓmax⁡\ell_{\max} and introducing targeted higher-order contractions can recover much of the missing geometric information.

The rings benchmark suggests that increasing angular resolution can recover geometric distinctions inaccessible to lower-order invariant bases. The failure on the (3/4)-rings configuration shows that finite invariant representations can retain isolated degeneracies even at higher ℓmax⁡\ell_{\max}. Interestingly, the additional contractions designed to resolve these special cases did not improve performance on the methane dataset, indicating that formal expressivity and practical predictive accuracy are not always directly aligned.

We then investigated whether the additional expressivity translates to improved molecular modeling on the methane dataset. The larger basis consistently achieved the lowest energy and force errors among the evaluated HIP-HOP configurations. At 36K parameters, HIP-HOP nearly matched the accuracy of EquiformerV3 while running 3.2×\times faster at inference. At 1.1M parameters, EquiformerV3 was more accurate, though HIP-HOP remained more than twice as fast. Together with the reduced basis-construction runtimes provided by our parallel implementation, these results demonstrate a favorable trade-off between predictive accuracy and computational efficiency.

Overall, our results provide a practical framework for diagnosing representation failures, constructing more expressive invariant bases, and evaluating the relationships among formal expressivity, predictive accuracy, and computational efficiency in machine-learned interatomic potentials.

Acknowledgment We thank Kei Davis for feedback on the manuscript. This work is published under LANL LA-UR-26-27280. We gratefully acknowledge the support of the U.S. Department of Energy through the LANL Laboratory Directed Research Development Program under project number 20250145ER.

Methane Training Details

We train each model using a fixed pool of 10610^{6} methane configurations and evaluate performance on a separate held-out test set of 80,000 configurations. Energies and forces are converted to kcal/mol and kcal/mol/A˚\AA, respectively, and models are optimized using the sum of energy and force MAE and RMSE with L2L2 regularization. Training uses early stopping based on validation energy MAE with a patience of 300 epochs, while the batch size is increased after 150 epochs without improvement.

Hyperparameter36.2 K model1.12 M model
Message-passing layers11
Hidden dimension1632
ℓmax⁡\ell_{\max}44
mmax⁡m_{\max}ℓmax⁡\ell_{\max}ℓmax⁡\ell_{\max}
Attention heads28
Attention hidden dimension432
FFN hidden dimension16128
Edge hidden dimension864
Radial basis functions2020
Maximum neighbors44
Cutoff radius10.3 A˚\AA10.3 A˚\AA
OptimizerAdamAdam
Batch size4096256
Learning rate2.5×10−32.5 \times10^{-3}1×10−41 \times10^{-4}
Gradient-norm clipping100100
L2L2 regularization1×10−61 \times10^{-6}1×10−61 \times10^{-6}
LR reduction factor0.50.5
LR scheduler patience150 epochs150 epochs
Early stopping patience300 epochsnone
Max training epochs10,00010,000

Table 2. EquiformerV3 hyperparameters used for methane training.

Hyperparameter∼25–36 K models1.13 M model
Message-passing layers11
Hidden dimension32192
Cutoff radius10.3 Å10.3 Å
Minimum batch size2564096
Maximum batch size40964096
Batch size doubling patience150 epochs–
OptimizerAdamAdam
Learning rate2.5×10−32.5 \times10^{-3}2.5×10−32.5 \times10^{-3}
LR reduction factor–0.5
LR scheduler patience–150 epochs
L2 regularization1×10−61 \times10^{-6}1×10−61 \times10^{-6}
Early stopping patience300 epochs300 epochs
Max training epochs10,00010,000

Table 3. HIP-HOP-NN hyperparameters used for methane training.

N-body Details

Each counterexample is a pair of star-shaped local environments centered on a single node with binary labels {0,1}\{0, 1\}. We restrict message passing to only be allowed from the center node to the leaf nodes (leaf–leaf messages are blocked), making the distance cutoff irrelevant. Random accuracy is 0.5. We report performance over 2 seeds.

Ring Dataset Details

For each ring configuration (ninner,noutern_{\mathrm{inner}}, n_{\mathrm{outer}}), we construct a local environment consisting of a central node and two rings containing ninnern_{\mathrm{inner}} and noutern_{\mathrm{outer}} nodes. All nodes have the same species. Edges connect the central node to every ring node, with no edges between nodes on either ring. For each configuration, we generate two graphs that differ only in the distance between the rings. The graphs are assigned binary labels according to the minimum Euclidean distance between an inner-ring and outer-ring node, with the smaller-distance configuration labeled “close” and the larger-distance configuration labeled “far.” Because each task contains one graph from each class, random accuracy is 0.5. We experimented with assigning the same radius for the inner and outer rings, but found that the accuracy of the models was identical to when the radii of the rings was different (Figure 3).

Figure 3. Ring benchmark configurations with (a) equal radii for the inner and outer rings and (b) different radii for the inner and outer rings.

HyperparameterValue
Message-passing layers1
Hidden dimension32
Sensitivity functions32
Atom layers per interaction2
Lossbinary cross-entropy
OptimizerAdam
Learning rate1×10−31 \times10^{-3}

Table 5. HIP-HOP-NN hyperparameters used for the n-body benchmark.

HyperparameterValue
Graphs per configuration2
Message-passing layers1
Hidden dimension32
Sensitivity functions32
Atom layers per interaction2
Lossbinary cross-entropy
OptimizerAdam
Learning rate1×10−31 \times10^{-3}
Max training epochs500

Table 6. HIP-HOP-NN hyperparameters used for the ring benchmark.

Speedups

Invariant-construction time was measured on a single NVIDIA Tesla V100 PCIe 32 GB GPU. Inference speed was measured with a batch size of 128 on an NVIDIA GeForce RTX 3090 GPU. For each model, the times were averaged over 4 runs after 1 warm up run.

Model / BasisInference Speed (μs\mu\mathrm{s}/atom)
Modelℓmax⁡\ell_{\max}nmax⁡n_{\max}# Par.BaselineOptimizedSpeedup
HIP-HOP3424.7 K2.452.281.07×
HIP-HOP341.02 M10.199.661.05×
HIP-HOP3525.8 K3.132.951.06×
HIP-HOP351.05 M16.0415.761.02×
EquiformerV33N/AN/A23.1 K11.3
HIP-HOP4324.4 K3.553.381.05×
HIP-HOP430.99 M13.9013.311.04×
HIP-HOP4436.3 K6.934.421.57×
HIP-HOP441.13 M21.7612.661.72×
HIP-HOP4737.4 K49.8826.971.85×
HIP-HOP471.39 M318.70202.541.57×
EquiformerV34N/AN/A36.2 K14.29
EquiformerV34N/AN/A1.12 M30.20

Table 7. Inference speed comparison between HIP-HOP and EquiformerV3 models at different model sizes.

ModelParamsTime / Epoch (s)
EquiformerV3 (ℓmax⁡=4\ell_{\max}=4)1.12M388.8
HIP-HOP (ℓmax⁡=4\ell_{\max}=4, nmax⁡=4n_{\max}=4)1.13M234.8

Table 8. Average training time per epoch on 1M methane configurations at a matched parameter count, averaged over 5 training-only epochs on one RTX 3090. HIP-HOP trains 1.66× faster than EquiformerV3.

BasisInvariant Construction (seconds)
ℓmax⁡\ell_{\max}nmax⁡n_{\max}BaselineOptimizedSpeedup
340.8990.07911.40×
351.7910.08421.35×
430.5000.0756.69×
444.0280.13829.29×
47267.8110.862310.70×

Table 9. Parallelizing the construction of polynomial invariants substantially reduces construction time, especially for the larger ℓmax⁡=4\ell_{\max}=4 bases.

New basis for ℓmax⁡=4\ell_{\max}=4

To extend HIP-HOP to higher angular resolution, we construct an invariant basis including channels up to ℓmax⁡=4\ell_{\max}=4, following the flexible moment-invariant framework of Bujack et al. [6].

The resulting tensor-contraction basis used in our model is summarized in Figure 5. This figure provides a visual representation of how the different invariant scalars are calculated. Each individual graph specifies a particular tensor contraction that evaluates to a scalar. Each node labeled ℓ\ell represents a rank-ℓ\ell tensor T(ℓ)T^{(\ell)}, while each edge represents a pair of tensor indices that are contracted with one another. Multiple edges between two nodes indicate that several indices are contracted between the same pair of tensors. The degree of each node therefore equals the rank of its corresponding tensor.

Invariant tensor-contraction basis for HIP-HOP with angular channels up to $\ell_{\max}=4$

Figure 5. Invariant tensor-contraction basis for HIP-HOP with angular channels up to (ℓmax⁡=4)(\ell_{\max}=4)

The Einstein-summation expression above each graph specifies the same contraction algebraically (Weisstein and Stover). The individual comma-separated terms correspond to the nodes, while each repeated index corresponds to an edge connecting the two tensors in which that index appears. For example, the graph labeled i,i-> contains two rank-1 nodes joined by one edge and represents Ti(1)Ti(1)T_i^{(1)}T_i^{(1)}. Similarly, the graph labeled ij,ik,jk-> contains three rank-2 tensors. The edge corresponding to ii joins the first and second tensors, the edge corresponding to jj joins the first and third tensors, and the edge corresponding to kk joins the second and third tensors.

Singularity on single plane

Representing a spatial structure in a reduced or truncated basis can result in a loss of geometric information, such that distinct configurations become indistinguishable at finite ℓmax⁡\ell_{\max}, as recently discussed in Domina and Ceriotti [8]. Motivated by this behavior, we arrived at a related observation independently, without the use of AI, by studying how (n)(n)-fold symmetric structures behave under projection to the invariant basis. In the following, we examine this effect in the simpler setting of a single plane and characterize the problematic configurations. For simplicity, we describe the problem in the original HIP-HOP ℓmax⁡=3\ell_{\max}=3 basis. As shown in Figure 4, the original HIP-HOP basis unexpectedly fails for the (2/3)(2/3)-ring case (described in Figure 3). We therefore analyze this case in more detail to understand the origin of the failure. An analogous behavior is observed for the (3/4)-ring configuration in the HIP-HOP ℓmax⁡=4\ell_{\max}=4 basis.

Ring benchmark heatmaps showing graph configurations $(1,3)$, $(1,4)$, $(2,2)$, $(2,3)$, $(2,4)$, $(3,3)$, $(3,4)$, and $(4,4)$ for $\ell_{\max}=3$ and $\ell_{\max}=4$ models, with green 1.0 and red 0.5 performance cells.

Figure 4. Ring benchmark performance across graph configurations for different model and invariant-basis configurations.

To analyze the (2/3)-ring configuration, we continuously rotate the two-fold symmetric component relative to the three-fold symmetric component and evaluate the corresponding basis functions, plotting only the nonzero responses shown in Figure 6. The original HIP-HOP basis is shown on the left, while the improved basis is shown on the right, with the two separated by a dashed line. As seen, all contractions in the original basis remain constant under the rotation, even though the configurations are distinct. Thus, the original invariants fail to distinguish the relative orientation of the two rings. The additional contraction is functionally independent of the original contractions on this specific manifold and therefore varies with the relative rotation between the two rings. We constructed this contraction using an algorithm similar to that described in Bujack et al. [6].

Figure 6. Nonzero contractions for the 2/3-ring configuration as a function of the relative rotation between the two- and three-fold components. The original contractions (left) remain constant, while the additional contraction Bujack et al. [6] (right) varies with the relative rotation.

From a tensor perspective, only the rank-2 and rank-3 tensors, T(2)T^{(2)} and T(3)T^{(3)}, are nonzero for this configuration. Choosing the plane of the rings to be the (x,y)(x,y) plane, they can be written as

T(2)(θ)=(a+bcos⁡(2θ)bsin⁡(2θ)0bsin⁡(2θ)a−bcos⁡(2θ)000−2a),T111(3)=c,T122(3)=T212(3)=T221(3)=−c,(1)T^{(2)}(\theta)=\begin{pmatrix} a+b\cos(2\theta) & b\sin(2\theta) & 0 \\ b\sin(2\theta) & a-b\cos(2\theta) & 0 \\ 0 & 0 & -2a \end{pmatrix},\qquad T^{(3)}_{111}=c,\qquad T^{(3)}_{122}=T^{(3)}_{212}=T^{(3)}_{221}=-c, \tag*{(1)}

with all remaining components of T(3)T^{(3)} equal to zero. Here, θ\theta denotes the relative rotation of the two-fold symmetric component, while aa, bb, and cc can be arbitrary real numbers.

Using the Einstein summation convention, the contractions are

f1=Tij(2)Tij(2)=einsum⁡(ij,ij->),(2)f_1=T^{(2)}_{ij}T^{(2)}_{ij}=\operatorname{einsum}(\text{ij,ij->}), \tag*{(2)}
f2=Tij(2)Tik(2)Tjk(2)=einsum⁡(ij,ik,jk->),(3)f_2=T^{(2)}_{ij}T^{(2)}_{ik}T^{(2)}_{jk}=\operatorname{einsum}(\text{ij,ik,jk->}), \tag*{(3)}
f3=Tij(2)Tikl(3)Tklj(3)=einsum⁡(ij,ikl,klj->),(4)f_3=T^{(2)}_{ij}T^{(3)}_{ikl}T^{(3)}_{klj}=\operatorname{einsum}(\text{ij,ikl,klj->}), \tag*{(4)}
f4=Tij(2)Tik(2)Tkml(3)Tlmj(3)=einsum⁡(ij,ik,kml,lmj->),(5)f_4=T^{(2)}_{ij}T^{(2)}_{ik}T^{(3)}_{kml}T^{(3)}_{lmj}=\operatorname{einsum}(\text{ij,ik,kml,lmj->}), \tag*{(5)}
f5=Tij(2)Tkl(2)Tijm(3)Tklm(3)=einsum⁡(ij,kl,ijm,klm->),(6)f_5=T^{(2)}_{ij}T^{(2)}_{kl}T^{(3)}_{ijm}T^{(3)}_{klm}=\operatorname{einsum}(\text{ij,kl,ijm,klm->}), \tag*{(6)}
fc=Tijk(3)Tlmn(3)Til(2)Tjm(2)Tkn(2)=einsum⁡(ijk,lmn,il,jm,kn->).(7)f_c=T^{(3)}_{ijk}T^{(3)}_{lmn}T^{(2)}_{il}T^{(2)}_{jm}T^{(2)}_{kn}=\operatorname{einsum}(\text{ijk,lmn,il,jm,kn->}). \tag*{(7)}

To make the origin of the degeneracy explicit, we evaluate the basis contractions after restricting T(2)T^{(2)} and T(3)T^{(3)} to the forms given above. On this symmetry-constrained manifold, the tensor components depend only on aa, bb, cc and relative rotation angle θ\theta. Substituting these expressions into the contractions and simplifying them allows us to determine directly which invariants retain information about the relative orientation of the two symmetric components.

(6a2+2b26a(−a2+b2)4ac24c2(a2+b2)4b2c24c2(a3−3b3sin⁡2(2θ)cos⁡(2θ)+b3cos⁡3(2θ)))(8)\begin{pmatrix} 6a^2+2b^2\\ 6a(-a^2+b^2)\\ 4ac^2\\ 4c^2(a^2+b^2)\\ 4b^2c^2\\ 4c^2\left(a^3-3b^3\sin^2(2\theta)\cos(2\theta)+b^3\cos^3(2\theta)\right) \tag*{(8)} \end{pmatrix}

None of the contractions in the original basis contains any dependence on the relative rotation angle θ\theta. Consequently, these contractions are insensitive to changes in the relative orientation of the two- and three-fold symmetric components on this manifold. Unlike the contractions in the original basis, fcf_c retains explicit dependence on θ\theta and therefore contains information about the relative orientation of the two symmetric components.

An analogous situation occurs for the 3/4-ring configuration in the ℓmax⁡=4\ell_{\max}=4 basis, where the missing angular dependence can be recovered by the additional contraction

einsum⁡(abc,def,ghi,jkl,ihck,gfda,lebj->).(9)\operatorname{einsum}(\text{abc,def,ghi,jkl,ihck,gfda,lebj->}). \tag*{(9)}

Although these additional contractions are important from a theoretical perspective, they are computationally expensive and provide only limited practical improvement in the experiments, as shown in Table 4. We interpret this as being consistent with the expectation that exactly planar atomic configurations associated with these singularities are very rare in realistic datasets.

ℓmax⁡\ell_{\max}nmax⁡n_{\max}# Par.Energy MAEEnergy RMSEForces MAEForces RMSE
HIP-HOP-NN
3424.7 K0.213±0.0160.213 \pm0.0160.500±0.3470.500 \pm0.3470.607±0.0320.607 \pm0.0322.631±2.5572.631 \pm2.557
3525.8 K0.225±0.0310.225 \pm0.0310.389±0.0460.389 \pm0.0460.642±0.0460.642 \pm0.0463.842±4.4983.842 \pm4.498
4324.4 K0.213±0.0040.213 \pm0.0040.378±0.0240.378 \pm0.0240.623±0.0130.623 \pm0.0131.537±0.3331.537 \pm0.333
4436.3 K0.184±0.0030.184 \pm0.0030.322±0.0060.322 \pm0.0060.551±0.0110.551 \pm0.0111.244±0.0261.244 \pm0.026
441.13 M0.0410.1070.1740.607
EquiformerV3
4N/A36.2 K0.1760.2910.5111.248
4N/A1.12 M0.0260.0550.1020.351

Table 4. Energy and force prediction performance for HIP-HOP-NN and EquiformerV3 on the methane dataset.

References

  1. [1]Alice E. A. Allen, Emily Shinkle, Roxana Bujack, and Nicholas Lubbers. Optimal invariant sets for atomistic machine learning. npj Computational Materials, 12:75, 2026. doi: 10.1038/s41524-025-01948-0.DOI
  2. [3]Ilyes Batatia, David Peter Kovacs, Gregor N. C. Simm, Christoph Ortner, and Gabor Csanyi. MACE: Higher order equivariant message passing neural networks for fast and accurate force fields. In Alice H. Oh, Alekh Agarwal, Danielle Belgrave, and Kyunghyun Cho, editors, Advances in Neural Information Processing Systems, 2022b. URL https://openreview.net/forum?id=YPpSngE-ZU.
  3. [4]Jörg Behler and Michele Parrinello. Generalized neural-network representation of high-dimensional potential-energy surfaces. Phys. Rev. Lett., 98:146401, Apr 2007. doi: 10.1103/PhysRevLett.98.146401. URL https://link.aps.org/doi/10.1103/PhysRevLett.98.146401.DOI
  4. [5]Michael M. Bronstein, Joan Bruna, Taco Cohen, and Petar Velickovic. Geometric deep learning: Grids, groups, graphs, geodesics, and gauges. CoRR, abs/2104.13478, 2021. URL https://arxiv.org/abs/2104.13478.
  5. [6]Roxana Bujack, Emily Shinkle, Alice Allen, Tomas Suk, and Nicholas Lubbers. Flexible moment-invariant bases from irreducible tensors, 2025. URL https://arxiv.org/abs/2503.21939.
  6. [7]Taco Cohen and Max Welling. Group equivariant convolutional networks. In International conference on machine learning, pages 2990–2999. PMLR, 2016.arxiv.org/abs/1602.07576
  7. [8]Michelangelo Domina and Michele Ceriotti. Using large language models to probe the limits of atom-centered structural descriptors. arXiv preprint arXiv:2607.26984, 2026.arxiv.org/abs/2607.26984
  8. [9]Ralf Drautz. Atomic cluster expansion for accurate and transferable interatomic potentials. Phys. Rev. B, 99:014104, Jan 2019. doi: 10.1103/PhysRevB.99.014104. URL https://link.aps.org/doi/10.1103/PhysRevB.99.014104.DOI
  9. [10]Johannes Gasteiger, Janek Groß, and Stephan G􏰀nnemann. Directional message passing for molecular graphs. In International Conference on Learning Representations, 2020. URL https://openreview.net/forum?id=B1eWbxStPH.arxiv.org/abs/2003.03123
  10. [11]Pierre Hohenberg and Walter Kohn. Inhomogeneous electron gas. Physical review, 136(3B):B864, 1964.
  11. [12]Ming-Kuei Hu. Visual pattern recognition by moment invariants. IRE Transactions on Information Theory, 8(2):179–187, 1962. doi: 10.1109/TIT.1962.1057692.DOI
  12. [13]Chaitanya K. Joshi, Cristian Bodnar, Simon V. Mathis, Taco Cohen, and Pietro Lio. On the expressive power of geometric graph neural networks. In Proceedings of the 40th International Conference on Machine Learning, volume 202 of Proceedings of Machine Learning Research, pages 15330–15355. PMLR, 2023.arxiv.org/abs/2301.09308
  13. [14]Risi Kondor, Zhen Lin, and Shubhendu Trivedi. Clebsch–gordan nets: a fully fourier space spherical convolutional neural network. Advances in neural information processing systems, 31, 2018.arxiv.org/abs/1806.09231
  14. [15]Max Langbein. Higher Order Moment Invariants and their Applications. PhD Dissertation, Technical University Kaiserslautern, 2014.
  15. [16]Mel Levy. Universal variational functionals of electron densities, first-order density matrices, and natural spin-orbitals and solution of the iινi/ιi-representability problem. Proceedings of the National Academy of Sciences, 76(12):6062–6065, 1979. doi: 10.1073/pnas.76.12.6062. URL https://www.pnas.org/doi/abs/10.1073/pnas.76.12.6062.DOI
  16. [17]Yi-Lun Liao, Alexander J. Hoffman, Sabrina C. Shen, Alexandre Duval, Sam Walton Norwood, and Tess Smidt. EquiformerV3: Scaling efficient, expressive, and general se(3)-equivariant graph attention transformers. arXiv preprint arXiv:2604.09130, 2026.arxiv.org/abs/2604.09130
  17. [18]Sergey Pozdnyakov, Michael Willatt, and Michele Ceriotti. Randomly-displaced methane configurations, September 2020a. URL https://doi.org/10.24435/materialscloud:qy-dp.
  18. [19]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, 2020b.DOI
  19. [20]Kristof T Schütt, Huziel E Sauceda, P-J Kindermans, Alexandre Tkatchenko, and K-R Müller. Schnet–a deep learning architecture for molecules and materials. The Journal of chemical physics, 148(24), 2018.DOI
  20. [21]J. S. Smith, O. Isayev, and A. E. Roitberg. Ani-1: an extensible neural network potential with dft accuracy at force field computational cost. Chemical Science, 8(4):3192–3203, 04 2017. ISSN 2041-6520. doi: 10.1039/c6sc05720a. URL https://doi.org/10.1039/c6sc05720a.
  21. [22]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.arxiv.org/abs/1802.08219
  22. [23]G. Vignale and Mark Rasolt. Density-functional theory in strong magnetic fields. Phys. Rev. Lett., 59:2360–2363, Nov 1987. doi: 10.1103/PhysRevLett.59.2360. URL https://link.aps.org/doi/10.1103/PhysRevLett.59.2360.DOI
  23. [24]Limei Wang, Yi Liu, Yuchao Lin, Haoran Liu, and Shuiwang Ji. Comenet: Towards complete and efficient message passing for 3d molecular graphs. Advances in Neural Information Processing Systems, 35:650–664, 2022.arxiv.org/abs/2206.08515
  24. [25]Maurice Weiler, Mario Geiger, Max Welling, Wouter Boomsma, and Taco Cohen. 3d steerable cnns: Learning rotationally equivariant features in volumetric data. Advances in Neural information processing systems, 31, 2018.arxiv.org/abs/1807.02547
  25. [26]Maurice Weiler, Patrick Forré, Erik Verlinde, and Max Welling. Equivariant and Coordinate Independent Convolutional Networks, 2023. URL https://maurice-weiler.gitlab.io/cnn_book/EquivariantAndCoordinateIndependentCNNs.pdf.
  26. [28]Brandon Wood, Misko Dzamba, Xiang Fu, Meng Gao, Muhammed Shuaibi, Luis Barroso-Luque, Kareem Abdelmaqsoud, Vahe Gharakhanyan, John Kitchin, Daniel Levine, et al. Uma: A family of universal models for atoms. Advances in Neural Information Processing Systems, 38:129391–129427, 2026.arxiv.org/abs/2506.23971

Paper details

vlnc:W_gjs2rtmc
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

Not recorded for this paper

View statement
Looking for related papers…
Contents