A Comprehensive Survey of Crystal Structural Prototypes: Construction, Classification, and Applications

Abstract

Crystal structural prototypes describe recurring atomic arrangements across chemically distinct materials. Prototype libraries are built through crystallographic curation, standardization, geometric matching, local-environment clustering, and periodic-graph analysis, and support chemical substitution, derivative enumeration, high-throughput screening, structure solution, crystal structure prediction (CSP), and generative design. This survey compares methods by the structural information they retain and the equivalence rules they impose, distinguishing crystallographic structure types, geometric realizations, chemistry-anonymous topologies, and chemical decorations. We review curated and automatically extracted prototype libraries (AFLOW, CSPD, LAE-ICSPD, EMFDTW, TYPIX, ICSD structure types), their source databases, and algorithms for assigning family membership. A GraphID-based analysis of samples from ICSD, the Materials Project, OQMD, and LeMat shows that the 100 most frequent families in each sample cover 30 to 67 percent of analyzed structures, while 72 to 82 percent of observed families are singletons. The GraphIDs of up to 76 percent of structures change when the neighbor tolerance is varied. In our comparison of 12 generative CSP models, template retrieval achieves the highest top-1 success rate. Removing a reduced-stoichiometry class from training introduces new failures in 50 to 78 percent of the test structures of that class, which measures sensitivity to stoichiometric coverage, not generalization to unseen prototypes. We relate these findings to earlier statistics of structure types, space groups, and periodic nets, and to protein-fold coverage studies. We propose independent-database coverage measurements, discovery-rate analyses, and family-held-out tests to assess how prototype reuse affects prediction accuracy, generalization, and novelty.