Curated Data Set, Structure–Property Relationships, and Machine Learning Prediction of Band Gaps for One-Dimensional Bismuth(III) and Antimony(III) Halides

Abstract Vast structural diversity and remarkable optoelectronic properties of organic–inorganic and all-inorganic metal halides continue to stimulate huge research interest. While the outstanding performance of lead halide perovskites has defined the field of perovskite photovoltaics, concerns regarding lead toxicity and long-term stability have driven the search for lead-free alternatives. Low-dimensional bismuth(III) and antimony(III) halides have emerged as promising candidates; however, their rational design remains hindered by the lack of established structure–property relationships and structure-predictive possibilities. In an effort to fill this knowledge gap, we report here an open, curated data set of one-dimensional (1D) bismuth(III) and antimony(III) halides featuring edge-sharing {MX4}− chain-type anions (M = Bi, Sb; X = Cl, Br, I), including experimentally determined optical band gap values. Leveraging crystallographically accessible structural parameters, we constructed two descriptor spaces for compounds with an α-{MX4}− anion type: a primary space based on M–X bond lengths and X–M–X bond angles (full geometric descriptor space) and a reduced space comprising average bond lengths and distortion parameters of building units [MX6] (reduced distortion descriptor space). Both spaces additionally incorporate descriptors capturing the interchain X···X contacts, anion composition, and crystallographic experiment temperatures. Given the limited data volume, we deliberately employed classical machine learning (ML) algorithms with cross-validation and stability analysis. The resulting models achieve prediction errors of 0.10 and 0.11 eV for the two descriptor spaces, approaching a typical experimental uncertainty of optical band gap determination. The feature-importance analysis reveals that geometric descriptors of the inorganic anionic framework dominate over noncovalent interchain interaction descriptors in governing the band gap. To promote reproducibility and further development in the field, we released a complete data set of group 15 1D-halometallates, trained models, and supporting materials in an open-access repository under the MIT license.