Inorganic Crystals Graphlet Bank: Atomic Cluster-Based Representation of Inorganic Crystals

Machine learning offers powerful opportunities for discovering new inorganic crystalline materials with desired properties, yet its effectiveness depends critically on how materials information is featurized for ML. Simple chemical formulas without structural information clearly carry limited information, the difference between graphite and diamond being a case in point. The Inorganic Crystals Graphlet Bank (ICGB) provides an experimentally grounded framework that unifies atomic, chemical, and structural information using systematic graphlet and symmetry features. When these features are combined with training data labels of target features, ICGB can enable ML algorithms to discover underlying trends and predict properties for new candidate materials. Using nearly 200k materials from the ICSD, each crystal is represented as a hierarchy of graphlets—atomic clusters encoding local geometry, chemistry, and symmetry. For each graphlet order, physically interpretable elemental properties such as ionization energy, electron affinity, and valence-electron count combined with structural information relevant at that order, form a uniform framework for capturing chemically relevant information. We illustrate this approach through a case study on predicting superconductivity using ICGB.