SWORD: Symmetry and Wyckoff-sequence of Ordered and Disordered crystals

The paper introduces SWORD, a novel symmetry-aware and Wyckoff-based string representation that standardizes and uniquely identifies both ordered and disordered crystal structures, thereby enabling efficient deduplication, novelty assessment, and curation of large-scale materials databases like the ICSD.

The Gist

Imagine you are trying to organize a massive, chaotic library of crystal structures. This library, called the ICSD, contains hundreds of thousands of books (crystal structures) describing how atoms are arranged in materials. Some books describe neat, perfectly ordered patterns (like a chessboard), while others describe messy, jumbled patterns where atoms are sharing seats or missing entirely (like a crowded subway car where people are shifting around).

The problem? The library is so big and messy that it's impossible to tell if a new book you found is truly a new discovery or just a duplicate of something already there, perhaps written with slightly different words or a slightly different layout.

Enter SWORD (Symmetry and Wyckoff-sequence of Ordered and Disordered crystals). Think of SWORD as a super-smart librarian with a magical new filing system.

The Problem: The "Same Book, Different Cover" Issue

In the world of crystals, the same structure can be described in many ways depending on how you look at it (rotating it, shifting the starting point, or changing the coordinate system).

• Old Methods: Imagine trying to match two books by comparing every single word. If one book says "The cat sat on the mat" and the other says "The feline rested on the rug," a simple computer might think they are different. Or, if the library has 100,000 books, checking every pair takes forever.
• The Disorder Problem: Many materials aren't perfect. Atoms might share a spot (like two people sitting on one chair) or be missing (an empty chair). Old filing systems often get confused by this "messiness" and can't tell if two messy structures are actually the same.

The Solution: SWORD's Magic Filing System

SWORD solves this by creating a unique, standardized ID card for every crystal, regardless of how messy or rotated it is.

1. The "Seat Map" Analogy (Wyckoff Positions)

Imagine a theater with numbered seats (Wyckoff positions).

• Ordered Crystals: Seat 1 has a person named "Iron," Seat 2 has "Oxygen." It's simple.
• Disordered Crystals: Seat 3 is shared by "Lithium" and "Manganese." Maybe 60% of the time it's Lithium, and 40% it's Manganese.
• SWORD's Trick: Instead of just saying "Seat 3 is messy," SWORD writes a specific code: "Seat 3 is a 60/40 mix of Li and Mn." It creates a string of text (a "SWORD label") that acts like a barcode. If two crystals have the same barcode, they are the same structure, even if one was described by a scientist in Japan and another in Germany using different coordinates.

2. The "Mixing Meter" (Degree of Mixing - DOM)

Sometimes, two crystals have the same "Seat Map" (same SWORD label) but the ratio of the mix is different.

• Example: Imagine a smoothie.
  • Smoothie A: 50% Strawberry, 50% Banana.
  • Smoothie B: 90% Strawberry, 10% Banana.
  • They are both "Strawberry-Banana Smoothies" (same SWORD label), but they taste different.
• SWORD's DOM: SWORD adds a Mixing Meter (DOM) to the ID card. It calculates exactly how "evenly" the ingredients are mixed. This allows scientists to group the 50/50 smoothies together and keep them separate from the 90/10 smoothies, ensuring they don't accidentally delete a unique variation.

3. The "Relaxation" Test (The Magic Mirror)

When scientists generate new crystal structures using AI, they often start with a rough, "unrelaxed" sketch (like a messy sketch of a building). It takes a lot of computer power to "relax" it into a perfect, stable shape.

• The Challenge: Can we tell if the messy sketch is the same as the perfect building before we spend the time and money to build the perfect one?
• SWORD's Superpower: SWORD is so good at recognizing the "soul" of the structure that it can look at the messy sketch and say, "Yes, that is definitely the same building as the perfect one, even though it's not finished yet." This saves scientists from wasting time checking duplicates that haven't even been fully built yet.

Why This Matters

1. No More Rediscovering the Wheel: SWORD helps scientists instantly know if a "new" material is actually just an old one in disguise.
2. Cleaning the Library: The authors used SWORD to clean up the ICSD database. They found that 46% of the entries were duplicates! They organized them into neat groups, making the database much smaller and more useful for AI.
3. Handling the Mess: It's the first tool that handles both perfect crystals and messy, disordered ones equally well, which is crucial because real-world materials are often messy.

In a Nutshell

SWORD is a new language for crystals. It translates the complex, messy, and confusing descriptions of atomic structures into a simple, standardized code. It's like giving every crystal a unique fingerprint that works even if the crystal is broken, jumbled, or rotated. This allows scientists to organize the world's crystal library, find true new discoveries faster, and train better AI models to design the materials of the future.

Technical Summary

1. Problem Statement

The discovery of novel materials requires the ability to distinguish genuinely distinct crystal structures from existing ones without redundancy. However, current crystallographic databases (e.g., ICSD) are expanding rapidly in size and complexity, making novelty assessment increasingly difficult.

• The Disorder Challenge: A critical bottleneck is crystallographic disorder (partial occupancies, vacancies, and positional disorder), which affects nearly 50% of entries in the Inorganic Crystal Structure Database (ICSD). Disorder significantly enlarges the structure–composition space.
• Limitations of Existing Methods:
  • Direct Structure Matching: Methods like StructureMatcher or AFLOW-XtalFinder are accurate but computationally expensive ($O(N^2)$ scaling), making them unsuitable for large-scale database screening.
  • Fingerprint Methods: While scalable, many existing fingerprints (e.g., SLICES, PDD, BAWL) fail to natively encode partial occupancies or struggle to distinguish subtle differences in disordered structures.
  • Redundancy: Current tools often fail to group symmetry-equivalent descriptions of the same disordered structure or cannot differentiate between structures with the same symmetry framework but different site stoichiometries.

2. Methodology: SWORD Representation

The authors introduce SWORD (Symmetry and Wyckoff-sequence of Ordered and Disordered crystals), a string-based representation designed to handle both ordered and disordered crystals.

A. Core Representation

SWORD encodes a crystal structure as a sequence of Wyckoff positions paired with their occupied elements.

• Standardization: Unlike anonymous prototypes (e.g., AFLOW), SWORD explicitly maps elements to specific Wyckoff sites.
• Handling Disorder: It explicitly represents co-occupying species on partially occupied sites. For example, a disordered site with Li and Mn is encoded with the specific species and their occupancies.
• Symmetry Normalization: To ensure that symmetry-equivalent structures (described by different origin choices or cell settings) yield the same label, SWORD employs a two-stage standardization process:
  1. Hall Setting Normalization: Maps alternative Hall settings to a default reference.
  2. Affine Transformation: Applies transformations to resolve equivalent Wyckoff letter sequences, ensuring a unique string for any given symmetry-equivalent structure.

B. Degree of Mixing (DOM)

To distinguish between structures that share the same SWORD label (same symmetry and site types) but differ in site stoichiometry (e.g., different doping levels), the authors introduce a scalar metric called Degree of Mixing (DOM).

• Calculation: DOM is calculated as a multiplicity-weighted average of normalized Shannon entropies (Pielou's evenness) across all disordered sites.
• Bias Indicator: A sign indicator ($\delta_{rep}$) is added to capture the direction of mixing bias (e.g., whether Species A or Species B dominates the mixture), ensuring that different stoichiometric ratios are distinguishable even if the total entropy is similar.
• Utility: This allows SWORD to group structures by symmetry framework while refining groups based on chemical composition.

C. Workflow

1. Input: Raw crystallographic data (CIFs).
2. Symmetrization: Uses Pymatgen to identify space groups and standardize coordinates.
3. Disorder Classification: Identifies substitutional, vacancy, or positional disorder based on distance tolerances.
4. Label Generation: Constructs the SWORD string and calculates the DOM.
5. Grouping: Clusters structures by SWORD label, then refines clusters using DOM values.

3. Key Contributions

1. Unified Representation: SWORD is the first fingerprint method that natively supports both ordered and disordered crystals within a single, compact, and indexable string format.
2. Disorder-Aware Metric: The introduction of DOM allows for the quantification and differentiation of site stoichiometry in disordered structures, a capability missing in previous symmetry-based descriptors.
3. Robust Standardization: The Hall-setting and affine transformation protocols ensure that symmetry-equivalent structures (regardless of coordinate origin or cell setting) produce identical labels, solving a major source of false duplicates.
4. Scalability: The method is designed for database-scale deduplication, offering linear time complexity compared to quadratic pairwise matching.

4. Results and Benchmarks

The authors benchmarked SWORD against BAWL, Short-BAWL, Pymatgen StructureMatcher, PDD, and SLICES.

• Invariance & Robustness:
  • SWORD achieved a 100% match rate under rigid site translations, isotropic lattice strain, and symmetry operations, matching the performance of the best existing methods.
  • Under Gaussian noise (structural perturbations), SWORD remained more stable than other fingerprint methods (SLICES, PDD) and showed tunable sensitivity based on symmetry tolerance.
• Relaxation Trajectory Association:
  • A critical test involved matching unrelaxed or intermediate structures to their final relaxed states.
  • SWORD demonstrated superior performance in associating intermediate configurations with their final relaxed basins compared to other methods, particularly in symmetry-constrained DFT relaxations. This suggests SWORD can reliably assess novelty even from partially relaxed or unrelaxed generated structures.
• Scalability:
  • Runtime analysis showed SWORD scales linearly with dataset size, comparable to other fingerprint methods and significantly faster than pairwise structure matchers.
• ICSD Curation Case Study:
  • Applied to 214,637 processed ICSD entries.
  • Identified 46.2% of entries as duplicates at the SWORD-label level.
  • Successfully refined disordered duplicates using DOM, separating entries with different site occupancies (e.g., distinguishing $Mg_{2-x}Fe_xSiO_4$ with varying Fe content).
  • Reduced the dataset to 127,056 unique, curated entries.

5. Significance

• Enabling AI-Driven Materials Discovery: By providing a reliable, disorder-aware method for deduplication and novelty assessment, SWORD creates a "clean" foundation for training machine learning models and generative AI, reducing bias caused by redundant data.
• Handling Real-World Complexity: It addresses the specific challenge of partial occupancies, which are ubiquitous in functional materials (e.g., solid solutions, doped catalysts) but often ignored or mishandled by existing tools.
• Efficiency: The ability to screen and curate massive databases (like ICSD) efficiently allows researchers to focus on genuinely novel candidates rather than rediscovering known materials.
• Future Extensions: The symmetry-based formulation provides a natural framework for analyzing phase transitions and relationships between ordered and disordered parent structures.

In summary, SWORD bridges the gap between high-throughput computational screening and the complex reality of disordered crystal structures, offering a robust, scalable, and chemically meaningful tool for modern materials informatics.