Graphlet Histogram Representation Database of Inorganic Crystals

This paper introduces Graphlet-MP, a comprehensive database and open-source toolkit that provides interpretable, data-efficient graphlet histogram representations for over 149,000 inorganic crystals to enable materials property prediction even with scarce experimental data.

The Gist

Imagine you are trying to describe a complex building to a friend who has never seen it. You could just list the ingredients: "It has 500 bricks, 20 windows, and a red door." This is like looking only at a material's composition (what atoms are inside). But this description fails to tell you if the windows are on the second floor or the roof, or if the bricks are stacked in a wall or a spiral. In materials science, this missing detail is crucial because the arrangement of atoms determines how the material behaves (like whether it conducts electricity or bends).

This paper introduces a new, smarter way to describe crystals called Graphlet-MP. Here is how it works, broken down into simple concepts:

1. The Problem: "Black Box" vs. "Blueprint"

Most modern computer models try to learn how to describe materials by reading millions of expensive computer simulations (called Density Functional Theory). It's like trying to learn how to bake a cake by tasting thousands of cakes without ever seeing the recipe. This works if you have endless data, but fails when you only have a few real-world examples (which is common for new, rare materials).

Other methods try to use "domain knowledge" (human rules) but often ignore the shape of the building, treating it like a bag of ingredients rather than a structured house.

2. The Solution: The "Graphlet" Blueprint

The authors created a system that breaks a crystal down into a hierarchical blueprint using three levels of detail, much like describing a city:

• Level 1: The People (Atomic Sites)
  Instead of just saying "there are 100 people," they count who is there and what they are like. They track 10 different traits for every atom (like their "personality," such as how strongly they attract electrons or their size). They create a histogram (a bar chart) showing the distribution of these traits across the whole crystal.
• Level 2: The Handshakes (Bonded Pairs)
  Now, they look at who is standing next to whom. They map out every pair of connected atoms. They don't just say "A is next to B"; they measure the distance between them and how their "personalities" differ. This captures the connectivity of the structure.
• Level 3: The Angles (Bond-Angle Triplets)
  Finally, they look at three atoms at a time to see the angles between them. This is like checking if a corner is a sharp 90-degree turn or a wide, open curve. This captures the 3D geometry that previous methods often missed.

By combining these three levels, they generate 79 different "histograms" (distributions) for every single material. Think of this as a unique 79-page ID card for every crystal, describing its local neighborhood in extreme detail.

3. The "Voronoi" Rule: Who is a Neighbor?

To know who is standing next to whom, the authors didn't use a simple "everyone within 5 feet" rule (which can be inaccurate in crowded or sparse areas). Instead, they used a method called Screened Voronoi Tessellation.

Imagine dropping a drop of water on a surface; it spreads out until it hits other drops. The boundary where two drops meet is their shared border. The authors use this geometric logic to decide which atoms are true neighbors. They then apply a "screen" (a filter) to ignore tiny, meaningless connections, ensuring they only count physically meaningful bonds. This creates a robust map of the crystal's structure.

4. The "Moving Earth" Metric: Comparing Materials

Once you have these 79 histograms for two different materials, how do you say how similar they are?

• Bad Way: Counting how many bars are different in the charts. If a bar shifts slightly to the right, a simple count might say they are totally different, even though they are very similar.
• The Paper's Way (Earth Mover's Distance): Imagine the histogram bars are piles of dirt. To turn Material A's pile into Material B's pile, you have to move the dirt. The "distance" is the amount of work required to move that dirt. If the piles are slightly shifted, it takes very little work (they are similar). If the piles are in completely different places, it takes a lot of work (they are different).

This method is robust against small errors and respects the physical reality that atoms close to each other are more similar than atoms far apart.

5. The Result: A Massive Library

The authors didn't just invent the method; they built a massive library called Graphlet-MP.

• They processed 149,082 inorganic crystals from the Materials Project database.
• They pre-calculated all 79 histograms for every single one.
• They made the code open-source, so anyone can take a new crystal structure (even one from a real lab experiment) and instantly generate its 79-page ID card to compare it with the library.

Why This Matters

This approach is like giving scientists a universal translator for materials. Instead of needing millions of examples to teach a computer what a material is, researchers can use these pre-made, human-understandable blueprints. This allows them to predict properties (like superconductivity or piezoelectricity) even when they only have a small amount of experimental data, bridging the gap between computer simulations and real-world discovery.

Technical Summary

Technical Summary: Graphlet Histogram Representation Database of Inorganic Crystals

Problem Statement
Current machine learning approaches for materials property prediction increasingly rely on end-to-end learning from large density-functional-theory (DFT) databases. While effective in data-rich regimes, these methods struggle when labeled experimental data is scarce, as is often the case for high-impact discoveries like high-temperature superconductors or novel piezoelectrics. Furthermore, many existing representations fail to capture critical structural nuances: composition-only methods (e.g., MAGPIE) cannot distinguish polymorphs, while graph neural networks (GNNs) often omit bond angles or encode them opaquely, requiring massive datasets to learn meaningful representations. Additionally, DFT-based training data often propagates systematic errors due to deviations from experimental reality. There is a need for domain-knowledge-driven representations that are data-efficient, interpretable, and capable of capturing local structural geometry without the burden of learning the representation from scratch.

Methodology
The authors introduce Graphlet-MP, a database and representation framework based on graphlet histograms. This approach explicitly embeds physical invariances and chemical intuition rather than learning them from data. The methodology proceeds through the following steps:

1. Neighbor Graph Construction: Instead of using rigid radial cutoffs or arbitrary $k$-nearest neighbor heuristics, the method employs a screened Voronoi tessellation. Two atoms are considered neighbors if their Wigner–Seitz cells share a geometric face. To ensure physical rigor, two screening conditions are applied:

   • The Voronoi face weight (solid angle normalized by the maximum for that site) must exceed 1%.
   • The interatomic distance must be within 1.5 times the sum of the effective atomic radii.
     This yields a robust neighbor map that adapts to varying packing fractions.

2. Graphlet Extraction: The method extracts deterministic, interpretable subgraphs (graphlets) in three hierarchical orders:

   • 1st Order (Atomic Sites): Captures the identity and electronic state of individual sites. Ten elemental attributes (e.g., Pauling electronegativity, electron affinity, ionization potential, covalent radius, atomic weight, periodic table column, and valence electron counts) are tracked. For non-stoichiometric materials, attributes are weighted by site occupancy.
   • 2nd Order (Bonded Pairs): Captures connectivity and bond-level properties. For every valid connected pair, permutation-invariant features are computed using the mean and absolute difference of the 10 site attributes, plus the interatomic distance.
   • 3rd Order (Bond-Angle Triplets): Captures angular constraints. Defined as a center site with two valid Voronoi neighbors, these triplets include the mean, standard deviation, skewness, and kurtosis of the 10 site attributes, with bond angles serving as an additional geometric degree of freedom.

3. Histogram Representation: To standardize the variable number of graphlets across different crystals, the extracted features are binned into histograms. This results in 79 distributions per material: 10 for 1st order, 21 for 2nd order, and 48 for 3rd order.

4. Distance Metric: To compare materials within this space, the authors utilize the Earth Mover's Distance (EMD), or Wasserstein-1 distance. Unlike pointwise metrics (e.g., Euclidean distance) that treat bins independently, EMD measures the minimum "work" to transform one distribution into another, penalizing shifts proportionally to their physical magnitude. The total inter-material distance is the sum of normalized EMDs across all 79 features. This metric is robust to small perturbations and resolves CIF multiplicity (e.g., primitive vs. supercell representations yield identical histograms).

Key Contributions

• Graphlet-MP Database: A precomputed database of graphlet histogram representations for 149,082 inorganic crystals from the Materials Project (MP). The dataset comprises 60 GB of data, with each material described by 79 distributions.
• Technical Specification: A complete, self-contained definition of the graphlet histogram representation, including the screened Voronoi neighbor construction and the specific feature extraction logic.
• Distance Metric Formalization: A full specification of the EMD-based metric tailored for graphlet histograms, demonstrating its suitability for kernel-based property prediction.
• Open-Source Codebase: A publicly available tool allowing users to generate graphlet histograms from arbitrary Crystallographic Information Files (CIFs), including experimentally determined structures, and to compute the EMD metric.

Results
The paper does not present new property prediction results but references prior work (Ref. [5]) where this representation was successfully applied. In that context, a Gaussian-process model (GP-Tc) trained on only ~4,350 experimentally obtained superconductors achieved an $R^2$ of 0.93 for predicting superconducting transition temperatures ($T_c$). The interpretability of the graphlet representation allowed for the compression of the feature space to four descriptors, identifying electron-affinity differences between neighboring atoms as the most informative predictor. The model successfully predicted and experimentally confirmed superconductivity in PtPb$_3$Bi ($T_c \approx 3$ K).

Significance
The authors position Graphlet-MP as a shared foundation for the materials informatics community, particularly for scenarios where experimental data is scarce. By decoupling the representation from the property predictor, the framework enables researchers to train interpretable models for diverse properties (superconductivity, dielectricity, piezoelectricity) without the data-hunger of black-box end-to-end models. The representation is designed to be extensible, allowing laboratories to integrate proprietary or experimental CIFs into the same representational space. The authors argue that as the community recognizes the limitations of data-hungry models for guiding real experimental programs, transparent, precomputable representations backed by ready-to-use databases and extensible codebases will play a central role in bridging computational prediction and laboratory discovery.