Rotation-Invariant Machine Learning for Highly Symmetric Molecular Configurations

Logan Bolton, Tomas Karella, Nicholas Lubbers, Nathaniel Philip Gorski, Emily Shinkle, Alice Elisabeth Anastasia Allen, Pieter J Swart, Roxana BujackarXiv 2026 · arXiv

TL;DR

Researchers have developed a more expressive rotation-invariant basis for atomistic machine learning that resolves long-standing geometric "blind spots" in highly symmetric molecular environments. By extending the HIP-HOP architecture to higher angular orders, the new model achieves accuracy comparable to state-of-the-art equivariant models like EquiformerV3 while running up to 3.2× faster during inference.

What Problem Does This Paper Address?

In the world of atomistic machine learning, we want to predict the energy and forces of molecules without running expensive quantum mechanical simulations. To do this, we must represent 3D atomic environments in a way that the computer understands.

There are two main schools of thought:

  1. Invariant models: These use simple scalar quantities (like distances and angles). They are computationally efficient but can suffer from "degeneracies"—meaning they might see two geometrically different atomic arrangements as identical, especially in highly symmetric molecules.
  2. Equivariant models: These use complex tensor-valued features that track directional information. They are highly expressive and accurate but come with a heavy computational price tag.

The authors ask a fundamental question: Is the heavy machinery of equivariance strictly necessary? Or can we build "complete" invariant representations that are both fast and smart enough to distinguish complex symmetries?

Methods

The authors build upon HIP-HOP, an invariant architecture that uses moment invariants. They interpret local atomic configurations as a density on a sphere and project it onto a polynomial basis.

Key technical contributions include:

  • Increasing Angular Order: They extended the basis from $\ell_{max}=3$ to $\ell_{max}=4$. Think of this like increasing the resolution of a camera; higher orders capture finer angular details of how neighbors are distributed around a central atom.
  • Tensor Contractions: They used Langbein’s algorithm to identify a compact, independent subset of tensor contractions, ensuring the representation is rich but not redundant.
  • Parallelization: To handle the increased math required for $\ell_{max}=4$, they parallelized the construction and evaluation of these invariant polynomials, mitigating the potential speed penalty.

Datasets

The study uses three distinct evaluation tools:

  • N-body Benchmark: A standard test to see how well models resolve interactions between multiple atoms.
  • Rings Benchmark (New): The authors introduced this lightweight synthetic dataset. It consists of a central node surrounded by two rings of atoms. By varying the number of atoms and rotating the rings, they created specific $n$-fold symmetries to expose where models fail.
  • Methane Dataset: A massive real-world dataset containing over 7 million randomly displaced methane configurations with high-fidelity quantum mechanical (DFT) labels.

Key Insights and Ideas

  • Symmetry Blindness is Real: The "Rings Benchmark" revealed that standard invariant models (and even some equivariant ones) are "blind" to certain symmetric rotations. For instance, $\ell_{max}=3$ models consistently failed to distinguish orientations in 4-fold symmetric configurations.
  • Resolution vs. Equivariance: Most of these failures can be resolved simply by increasing the angular resolution ($\ell_{max}=4$) of the invariant basis, rather than switching to a full equivariant architecture.
  • The (3/4) Singularity: Interestingly, both the new HIP-HOP and EquiformerV3 failed on a specific (3/4)-ring configuration. This highlights a "singular degeneracy" where even high-order models struggle. While the authors found a specific math contraction to fix this, it didn't improve real-world methane performance, suggesting these perfect planar symmetries are rare in nature.
The Rings Benchmark: Models must distinguish between "Close" and "Far" orientations of rotated rings.

Key Results

The extended HIP-HOP model strikes an impressive balance between speed and precision:

  • Accuracy: At a model size of ~36K parameters, HIP-HOP ($\ell_{max}=4$) matches the accuracy of EquiformerV3 on the methane dataset.
  • Inference Speed: HIP-HOP is significantly faster, achieving a 3.2× speedup over EquiformerV3 at smaller scales and remaining 2.4× faster at larger scales (1.1M parameters).
  • Training Efficiency: The model trains roughly 1.66× faster than its equivariant counterpart.
  • Expressivity: As shown in the benchmark results below, the $\ell_{max}=4$ basis (green squares) successfully resolves many symmetric cases where $\ell_{max}=3$ (red squares) failed.
Benchmark performance comparison. Green indicates success (1.0 accuracy), red indicates failure (0.5 accuracy).

Limitations and Future Work

While the results are promising, the authors note a few areas for growth:

  • Scaling: At very large model sizes (over 1 million parameters), EquiformerV3 still maintains a slight edge in absolute accuracy, even if it is slower.
  • Rare Degeneracies: The specific "singular" failures identified (like the 3/4 ring) require computationally expensive contractions to solve. The authors suggest that for most practical datasets, the current $\ell_{max}=4$ basis is "good enough," but more work is needed to see if these rare cases matter in more complex chemical systems.
  • Generalization: The current study focuses heavily on methane; testing this basis on more diverse molecular datasets (like MD17 or AcS) would further validate the approach.

Summary updated

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 $n$-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 $3.2\times$ faster at inference. These results show that carefully constructed invariant representations can mitigate key expressivity limitations while balancing accuracy and computational efficiency.
BibTeX
@misc{bolton2026rotation,
  author = {Logan Bolton and Tomas Karella and Nicholas Lubbers and Nathaniel Philip Gorski and Emily Shinkle and Alice Elisabeth Anastasia Allen and Pieter J Swart and Roxana Bujack},
  title = {{Rotation-Invariant Machine Learning for Highly Symmetric Molecular Configurations}},
  year = {2026},
  eprint = {2611.21308},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2611.21308},
}
Figure