Completeness of Atomic Structure Representations

Jigyasa Nigam, Sergey N. Pozdnyakov, Kevin K. Huguenin-Dumittan, Michele Ceriotti

Abstract

In this paper, we address the challenge of obtaining a comprehensive and symmetric representation of point particle groups, such as atoms in a molecule, which is crucial in physics and theoretical chemistry. The problem has become even more important with the widespread adoption of machine-learning techniques in science, as it underpins the capacity of models to accurately reproduce physical relationships while being consistent with fundamental symmetries and conservation laws. However, the descriptors that are commonly used to represent point clouds -- most notably those adopted to describe matter at the atomic scale -- are unable to distinguish between special arrangements of particles. This makes it impossible to machine learn their properties. Frameworks that are provably complete exist but are only so in the limit in which they simultaneously describe the mutual relationship between all atoms, which is impractical. We present a novel approach to construct descriptors of finite correlations based on the relative arrangement of particle triplets, which can be employed to create symmetry-adapted models with universal approximation capabilities. Our strategy is demonstrated on a class of atomic arrangements that are specifically built to defy a broad class of conventional symmetric descriptors, showcasing its potential for addressing their limitations.

Introduction

To predict the properties of molecules or materials using machine learning, 3D atomic structures must be converted into mathematical vectors called descriptors. A robust descriptor must be invariant to "nuisance" symmetries, such as rotation, translation, and the permutation of identical atoms. However, a significant "blind spot" exists in popular descriptors like SOAP or the Bispectrum: they are often incomplete. This incompleteness leads to "degenerate" configurations—distinct 3D shapes that produce identical descriptor values. When two different structures appear identical to a model, it becomes physically impossible to predict different energies for them, causing numerical instabilities and fundamental failures in learning physical relationships.

Methods: Non-Linear Triplet Correlations

The core of this work is a novel construction based on 3-center-1-neighbor features. Unlike standard "body-order" approaches that look at neighbors relative to a central atom, this method "tags" two specific neighbors to define a local, rigid coordinate system.

Schematic of the triplet-based construction using local frames and non-linear layers to achieve completeness.

By computing correlations within these triplets and applying a non-linear transformation (such as a neural network layer) before summing over all possible neighbor pairs, the model restores permutation invariance while effectively reconstructing high-order geometric information. This approach follows the "Deep Sets" principle, providing a provably universal way to approximate functions on sets. The authors provide a formal proof that this construction is complete for all non-collinear atomic arrangements.

Results: Resolving Structural Degeneracies

The method was validated using a challenging benchmark of 4,000 pairs of $B_8$ boron clusters. These clusters were specifically engineered to be "bispectrum-degenerate," making them indistinguishable to standard 3-body and 4-body descriptors. Ground truth energies were calculated using high-fidelity Density Functional Theory (PBE-DFT).

Performance on degenerate boron clusters, showing how the new representation (blue) resolves structural differences that standard 3-body descriptors (red) cannot.

As shown in Figure 3, standard models (red) exhibited a 100% relative error when attempting to predict energy differences between these degenerate pairs, as they were unable to perceive any structural difference. In contrast, the new non-linear triplet model (blue) resolved these differences with high precision. Furthermore, the authors demonstrated that an encoder-decoder model could reconstruct "full-body" features (correlations of all 7 neighbors) using only this triplet-based representation, confirming that essential geometric information is preserved.

Discussion and Conclusion

This research highlights that the failure of standard descriptors stems from their reliance on linear contractions of density correlations. While linear bases require impractical, infinitely high body-orders to achieve completeness, non-linear functions of triplets are sufficient to uniquely determine atomic positions.

While the $O(N^2)$ scaling of this method is more computationally intensive than simpler descriptors, it remains manageable for most molecular systems. Future work will focus on integrating this complete layer into state-of-the-art message-passing neural networks to ensure robustness against structural degeneracies in complex materials and high-pressure phases.

Summary updated

Abstract

In this paper, we address the challenge of obtaining a comprehensive and symmetric representation of point particle groups, such as atoms in a molecule, which is crucial in physics and theoretical chemistry. The problem has become even more important with the widespread adoption of machine-learning techniques in science, as it underpins the capacity of models to accurately reproduce physical relationships while being consistent with fundamental symmetries and conservation laws. However, the descriptors that are commonly used to represent point clouds -- most notably those adopted to describe matter at the atomic scale -- are unable to distinguish between special arrangements of particles. This makes it impossible to machine learn their properties. Frameworks that are provably complete exist but are only so in the limit in which they simultaneously describe the mutual relationship between all atoms, which is impractical. We present a novel approach to construct descriptors of finite correlations based on the relative arrangement of particle triplets, which can be employed to create symmetry-adapted models with universal approximation capabilities. Our strategy is demonstrated on a class of atomic arrangements that are specifically built to defy a broad class of conventional symmetric descriptors, showcasing its potential for addressing their limitations.
BibTeX
@inproceedings{nigam2024completeness,
  author = {Jigyasa Nigam and Sergey N. Pozdnyakov and Kevin K. Huguenin-Dumittan and Michele Ceriotti},
  title = {{Completeness of Atomic Structure Representations}},
  booktitle = {APL},
  year = {2024},
  doi = {10.1063/5.0160740},
}
Figure