Abstract
Modern Crystal Structure Prediction (CSP) tools generate thousands of simulated structures, creating an "embarrassment of over-prediction." To identify unique and synthesizable materials, researchers require invariants that are independent of crystal representation and stable under atomic vibrations. This paper introduces Average Minimum Distances (AMDs), a hierarchy of continuous descriptors that quantify geometric similarity. AMDs are shown to be stronger than traditional powder diffraction patterns, capable of distinguishing homometric crystals while remaining computationally efficient for large-scale databases.
The Challenge of Crystal Comparison
In CSP, supercomputers simulate millions of atomic arrangements to identify stable local energy minima. However, comparing these outputs is non-trivial. A single crystal can be represented by infinitely many unit cells, and traditional symmetry-based classifications are "brittle"—minor atomic displacements can cause discrete jumps in symmetry groups. To navigate the energy landscape effectively, a similarity measure must be:
- Invariant: Independent of the chosen unit cell or rigid motion.
- Continuous: Stable under small perturbations (atomic vibrations).
- Computable: Efficient enough to process thousands of structures.
Average Minimum Distances (AMDs)
The authors propose a sequence of descriptors derived from inter-atomic distances:
- Point-wise Distribution of Distances (PDD): For each atom in a unit cell, a list of distances to its $k$ nearest neighbors is generated.
- Weighted Point-wise Distribution (WPD): To ensure independence from unit cell size, weights are assigned to these distance lists, making the descriptor invariant to supercell representations.
- Average Minimum Distance (AMD): The core signature is created by averaging the $k$-th nearest neighbor distances across all atoms in the motif. This results in a vector that serves as a robust geometric "fingerprint."
Distinguishing Homometric Crystals
A significant advantage of AMDs is their ability to distinguish homometric sets—different arrangements of points that yield identical distance distributions and, consequently, identical X-ray diffraction patterns.
While traditional powder diffraction might fail to separate these structures, the AMD hierarchy provides enough resolution to identify unique geometries. The authors provide a fast $O(m^2)$ algorithm to detect such sets and prove a Completeness Theorem, suggesting that for most point sets, these descriptors uniquely reconstruct the crystal structure.
Validation on the T2 Dataset
The method was validated using the T2 database, comprising 5,799 simulated organic crystals. Key findings include:
- Enhanced Matching: AMDs identified a simulated match for the experimental crystal T2-$\delta$ (Crystal 15) that was geometrically closer than the one previously selected by experts (Crystal 14) using standard RMS deviation methods.
- Geometric Clustering: Hierarchical clustering based on AMDs revealed previously hidden "families" of structures within the top 100 most stable crystals.
- Efficiency: The near-linear computational complexity makes AMDs suitable for screening massive materials databases.
Conclusion
AMDs provide a mathematically rigorous and computationally practical framework for quantifying crystal similarity. By prioritizing stability and continuity over discrete symmetry, this approach allows researchers to map the complex energy landscapes of simulated materials more accurately than ever before.
Reference: Mosca, M., & Kurlin, V. (2020). Average Minimum Distances continuously quantify similarities between geometries of any crystal structures. arXiv:2009.02488v1.
Summary updated