Completeness of atomic structure representations

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

TL;DR

Researchers have developed a new mathematical framework for representing atomic structures that is "complete"—meaning it can distinguish between any two distinct 3D arrangements of atoms. By using non-linear functions of neighbor triplets to define local coordinate frames, this approach overcomes a fundamental limitation where previous machine learning models were "blind" to certain structural differences.

What Problem Does This Paper Address?

To predict the properties of molecules or materials using machine learning, we first need to turn a 3D cloud of atoms into a mathematical vector called a descriptor. A good descriptor must be invariant to "nuisance" symmetries: if you rotate a molecule, translate it, or swap the labels of two identical atoms, the descriptor shouldn't change.

However, a major "blind spot" was recently discovered in the most popular descriptors (like SOAP or the Bispectrum). These descriptors are often incomplete: there exist "degenerate" configurations—distinct 3D shapes that produce the exact same descriptor values. If the input is identical for two different structures, an ML model is physically incapable of predicting different energies for them, leading to "outright impossibility" in learning certain physical relationships and causing numerical instabilities in simulations.

Methods

The authors propose a novel construction based on 3-center-1-neighbor features.

  1. Local Frames: Instead of just looking at neighbors relative to a central atom (the standard "body-order" approach), they "tag" two specific neighbors to define a local, rigid coordinate system.
  2. Triplet Correlations: They compute correlations between the central atom and these two tagged neighbors (forming a triplet).
  3. The Non-Linear "Trick": The core innovation is applying a non-linear layer (like a neural network) to these triplet features before summing over all possible pairs of neighbors to restore permutation invariance.
  4. Deep Sets Architecture: This follows the "Deep Sets" principle, which is a provably universal way to approximate functions on sets. By processing triplets non-linearly, the model can effectively "reconstruct" high-order geometric information (like 4-body, 5-body, etc.) that linear descriptors miss.

The paper includes a formal proof showing that this construction is complete as long as the atoms aren't all collinear.

Datasets

The authors validated their method on a particularly difficult benchmark:

  • $B_8$ Clusters: A dataset of 4,000 pairs of boron clusters. These clusters were specifically engineered to be "bispectrum-degenerate"—meaning they are indistinguishable to standard 3-body and 4-body descriptors.
  • DFT Ground Truth: Energies for these clusters were calculated using high-fidelity Quantum Chemistry (PBE-DFT).
  • Benchmark Comparison: They compared their new triplet features against standard atom-centered symmetry functions, the SOAP power spectrum, and the bispectrum.

Key Insights and Ideas

  • Linearity is the Enemy of Completeness: The "aha moment" is that standard descriptors fail because they rely on linear contractions of density correlations. While you can make a linear basis complete by going to infinitely high body-orders (e.g., looking at 10 atoms at once), it's computationally impractical.
  • Triplets are Enough: You don't need to look at 10 atoms simultaneously if you use non-linear functions of 3 atoms. A triplet is the minimum required to define a 3D coordinate frame; once that frame is set, every other atom's position is uniquely determined.
  • Smoothness over Rigidity: Previous attempts to use local frames often suffered from discontinuities (the "flipping" of the frame as atoms moved). This method remains smooth because it sums over all possible neighbor pairs, weighted by their positions.

Key Results

  • Resolving the Unresolvable: Standard models (like the Bispectrum) showed a 100% relative error when trying to predict the energy difference between degenerate $B_8$ pairs—they literally could not see a difference.
  • High Accuracy: The new non-linear triplet model ($B_{NL}$) resolved these differences with high precision, achieving very low validation errors where previous models failed entirely.
  • Feature Reconstruction: The authors showed that an encoder-decoder model could reconstruct "full-body" features (correlations of all 7 neighbors at once) using only their triplet-based representation, proving that the information is indeed preserved.

Limitations and Future Work

  • Computational Scaling: Summing over all pairs of neighbors $(i_1, i_2)$ within an environment scales as $O(N^2)$, where $N$ is the number of neighbors. This is more expensive than simple 2-body or 3-body descriptors, though the authors argue it is still manageable for most molecular systems.
  • Discretization Sensitivity: While the framework is theoretically complete, the practical performance still depends on the resolution of the basis functions (radial and angular) used to discretize the neighbor density.
  • Integration: The next step is integrating this "complete" layer into state-of-the-art message-passing neural networks (like MACE or Allegro) to ensure they are robust against structural degeneracies in complex materials like high-pressure ice or complex alloys.

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, some of the descriptors that are commonly used to represent point clouds— notably those based on discretized correlations of the neighbor density that power most of the existing ML models of matter at the atomic scale—are unable to distinguish between special arrangements of particles in three dimensions. This makes it impossible to machine learn their properties. Atom-density correlations are provably complete 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, and have the resolution of the neighbor discretization as the sole convergence parameter. Our strategy is demonstrated on a class of atomic arrangements that are specifically built to defy a broad class of conventional symmetric descriptors, showing 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},
}

Loading related content...

Figure