Average Minimum Distances of periodic point sets

Daniel E. Widdowson, Marco M. Mosca, Angeles Pulido, Vitaliy Kurlin, Andrew I. Cooper

Abstract

Periodic sets of points model all solid crystalline materials (crystals) by representing atoms as labeled points. Crystal structures are determined in a rigid form and are considered up to rigid motions or isometries. Modern tools of Crystal Structure Prediction output thousands of simulated structures, though only few of them can be really synthesized. The first obstacle is the presence of many near duplicate structures that can not be efficiently recognized on the fly by past tools. To continuously quantify a similarity between periodic sets, their isometry invariants should be continuous under perturbations when all discrete invariants such as symmetry groups can break down. This paper studies the isometry classification problem for periodic sets with the new continuity requirement and introduces the Average Minimum Distances, which form an infinite sequence of continuous isometry invariants. Their asymptotic behaviour for a wide class of sets is explicitly described in terms of a point packing coefficient. All results are illustrated by experiments on large datasets of crystals.

Introduction: The Challenge of Crystal Over-Prediction

In the field of materials discovery, Crystal Structure Prediction (CSP) is a double-edged sword. While it allows researchers to simulate potential new materials, it often results in an "embarrassment of over-prediction"—generating thousands of candidate structures for a single molecule.

The challenge lies in identifying which of these thousands are actually the same. Traditional crystallographic tools rely on discrete invariants like symmetry groups or "reduced cells" (e.g., Niggli cells). However, these are mathematically discontinuous: a tiny perturbation in an atom's position (common due to thermal vibrations or simulation noise) can cause the symmetry group to collapse or the reduced cell to jump to a completely different shape. Furthermore, periodic sets can be represented in many different ways depending on the choice of basis or unit cell, making it incredibly difficult to automatically recognize "near-duplicate" structures.

Periodic sets can be defined by different unit cells and bases, complicating comparison.

Methods: Average Minimum Distance (AMD)

The authors propose the Average Minimum Distance ($AMD_k$) as a robust, continuous "fingerprint" for periodic point sets.

  1. Pointwise Distribution of Distances (PDD): For a crystal with $m$ atoms in its unit cell, they calculate the distances to the $k$ nearest neighbors for every atom in the infinite periodic set. This creates an $m \times k$ matrix where each row is sorted by distance.
  2. Averaging for Invariance: To make the descriptor independent of which atom you start with or how you define the unit cell (basis independence), they sort the rows lexicographically and then average the columns. The resulting vector of $k$ values is the AMD.
  3. Algorithmic Efficiency: They provide a fast algorithm using $n$-dimensional trees to find neighbors in the infinite periodic set. The complexity is near-linear ($O(km \log m)$), allowing for rapid processing of large databases.
The AMD framework maps isometry classes of lattices to continuous invariants.

Key Insights and Results

  • Continuity is Essential: The shift from discrete classification (is it symmetry group $X$?) to continuous quantification (how similar are these two structures?) mirrors how computer vision moved from rigid template matching to robust feature descriptors.
  • Asymptotic Behavior: The authors discovered that as $k$ goes to infinity, the AMD curve approaches a specific value related to the crystal's density. This "point packing coefficient" $c(S)$ provides a physical link between the geometric descriptor and the material's density.
  • Experimental Success: The method was evaluated on the T2 Crystal Dataset (5,679 simulated structures). AMD values were computed in just 5 minutes on a standard desktop. The method automatically identified simulated "near-neighbors" for nine experimental crystals and successfully clustered similar structures, effectively filtering out simulation noise.

Limitations and Future Work

While AMDs are highly effective, they might be "incomplete"—meaning it is theoretically possible for two non-isometric crystals to have the exact same AMD profile. To address this, future work explores moving from the averaged AMD vector back toward the full PDD matrix, which contains more granular information while maintaining continuity. Additionally, extending the framework to include chemical labels (e.g., distinguishing Carbon from Oxygen) will further refine similarity metrics for complex materials.

Summary updated

Abstract

Periodic sets of points model all solid crystalline materials (crystals) by representing atoms as labeled points. Crystal structures are determined in a rigid form and are considered up to rigid motions or isometries. Modern tools of Crystal Structure Prediction output thousands of simulated structures, though only few of them can be really synthesized. The first obstacle is the presence of many near duplicate structures that can not be efficiently recognized on the fly by past tools. To continuously quantify a similarity between periodic sets, their isometry invariants should be continuous under perturbations when all discrete invariants such as symmetry groups can break down. This paper studies the isometry classification problem for periodic sets with the new continuity requirement and introduces the Average Minimum Distances, which form an infinite sequence of continuous isometry invariants. Their asymptotic behaviour for a wide class of sets is explicitly described in terms of a point packing coefficient. All results are illustrated by experiments on large datasets of crystals.
BibTeX
@misc{widdowson2020average,
  author = {Daniel E. Widdowson and Marco M. Mosca and Angeles Pulido and Vitaliy Kurlin and Andrew I. Cooper},
  title = {{Average Minimum Distances of periodic point sets}},
  year = {2020},
  eprint = {2009.02488},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2009.02488},
}

Loading related content...

Figure