Quantifying hybrid lead halide perovskites from multiple viewpoints: a human-guided machine learning approach

Abstract Hybrid organic-inorganic lead halide perovskites (HOIP) have garnered immense interest due to their outstanding semiconducting properties, despite concerns regarding toxicity and stability. This study presents a comprehensive survey of all known HOIP structures, analyzing 3028 crystal structure extracted from the Cambridge Structural Database (CSD), Inorganic Crystal Structure Database (ICSD), and Crystallographic Open Database (COD). Special attention has been paid to the subset of 2-periodic frameworks (~ 1400 structures), including single- to multi-layered networks based on (100)-cut structures sliced from the parent 3-periodic perovskite haloplumbate aristotype structure. Employing a combination of human expertise and machine learning techniques, we systematically classify these structures based on topological, crystallographic, and electronic characteristics. Our findings establish correlations between composition, dimensionality, and bandgap properties, revealing that the fraction of low-bandgap materials increases with framework periodicity. An updated Layer Shift Factor calculation algorithm has been introduced and allowed for refined classification of 2-p HOIP stacking arrangements. It has been demonstrated that stacking configurations transition continuously between Ruddlesden-Popper and Dion-Jacobson phases rather than being strictly discrete. Additionally, a machine learning model trained on 560 unique compounds successfully predicts stacking modes based on cation chemistry. This work provides the most extensive curated dataset of HOIP to date, offering valuable insights into structural trends and guiding future material design for semiconducting applications.