RADII: Radius-Resolved Benchmark of Nanoparticle Structures for Generative Models in Materials Science

RADII is a radius-resolved benchmark of ~75,000 nanoparticle structures (33–11,298 atoms across ten materials) designed to measure the extrapolation frontier of graph generative models for materials science. Each structure is a spherical truncation of a published crystal lattice; the benchmark treats radius as a continuous scaling knob and provides leakage-free in-distribution and out-of-distribution splits. This archive contains the seed inputs for the benchmark: the ten primitive unit cell CIFs and per-(material, radius) base XYZ structures. The full ~12 GB benchmark (with rotation augmentation) is reproduced locally by running the generation pipeline shipped in the GitHub repository. Paper: "How Far Can You Grow? Characterizing the Extrapolation Frontier of Graph Generative Models for Materials Science" (KDD '26). Code & generation pipeline: github.com/KurbanIntelligenceLab/RADII (MIT license). To reproduce the full benchmark from this archive: python -m create_radii.create_radii --raw-data radii_raw.zip --output radii. Materials included: Ag, Au, CH₃NH₃PbI₃, Fe₂O₃, MoS₂, PbS, SnO₂, SrTiO₃, TiO₂, ZnO.