Title: From isomorphism to polymorphism: connecting interzeolite transformations to structural and graph similarity

Abstract: Zeolites are nanoporous crystalline materials with abundant industrial applications. Despite sustained research, only 235 different zeolite frameworks have been realized out of millions of hypothetical ones predicted by computational enumeration. Structure-property relationships in zeolite synthesis are very complex and only marginally understood. Here, we apply structure and graph-based unsupervised machine learning to gain insight on zeolite frameworks and how they relate to experimentally observed polymorphism and phase transformations. We begin by describing zeolite structures using the Smooth Overlap of Atomic Positions method, which clusters crystals with similar cages and density in a way consistent with traditional hand-selected composite building units. To also account for topological differences, zeolite crystals are represented as multigraphs and compared by isomorphism tests. We find that fourteen different pairs and one trio of known frameworks are graph isomorphic. Based on experimental interzeolite conversions and occurrence of competing phases, we propose that the availability of kinetic-controlled transformations between metastable zeolite frameworks is related to their similarity in the graph space. When this description is applied to enumerated structures, over 3,400 hypothetical structures are found to be isomorphic to known frameworks, and thus might be realized from their experimental counterparts. Using a continuous similarity metric, the space of known zeolites shows additional overlaps with experimentally observed phase transformations. Hence, graph-based similarity approaches suggest a venue for realizing novel zeolites from existing ones by providing a relationship between pairwise structure similarity and experimental transformations.