Band gaps and densities of states of 1,784 halide perovskites (SIESTA-PBE/DZP), with orbital cluster-expansion features and analysis code

Density-functional reference data and code supporting the manuscript Band gaps of halide perovskites from an orbital cluster-expansion descriptor: mechanism, a localized failure in the iodides, and a hybridization channel. Contains the 1,784 SIESTA-PBE/DZP supercell inputs, the primary result table with band gaps, band edges and convergence flags for all 1,784 structures (322 of them metallic), the density of states of every structure on a common grid (1784 x 801), the OCE feature matrix (1,784 x 3,088), the fitted-model outputs including the penalty-selection variance study, and the scripts that regenerate every table and figure of the paper. The 1,784-structure pool is shared with an earlier paper of the same author, where it was used as a candidate pool for an active-learning loop on cohesive energies at a smaller basis; no band gaps were computed there. The gaps, the densities of states and the double-zeta-plus-polarization labels in this deposit are new. Raw SIESTA run directories (~3 GB) and the eigenvalue archive (~660 MB) are not deposited; everything derived from them is included at full resolution. Provenance of the structures. The 1,784 structures were not taken from a public structure database. They were constructed for this work from the cubic Pm-3m (alpha) archetype for A in {Cs, K}, B in {Pb, Sn, Ge} and X in {F, Cl, Br, I}, including halide-mixed compositions, by code/build_dataset.py; code/gen_inputs.py then converts the pool into the SIESTA inputs. data/structures_source.json is the pool before conversion. Provenance of the lattice constants that seed the construction. Each endmember composition is seeded with a lattice constant and mixed compositions are linearly interpolated (Vegard) between the seeds. The seed table is not a curated experimental table, and its provenance is not uniform. Where the ABX3 composition has a known cubic perovskite phase, the seed is the published experimental constant: CsPbCl3 and CsPbBr3 from Møller, Nature 1958, 182, 1436 (a = 5.605 and 5.874 Å); CsGeCl3, CsGeBr3 and CsGeI3 from Thiele, Rotter and Schmidt, Z. Anorg. Allg. Chem. 1987, 545, 148 (a = 5.434 and 5.636 Å for Cl and Br; the CsGeI3 seed is close to but not identical with the reported H-phase constant); CsSnBr3 from Mori and Saito, J. Phys. C: Solid State Phys. 1986, 19, 2391 (a = 5.804 Å); and the reference value for CsPbI3 from Trots and Myagkota, J. Phys. Chem. Solids 2008, 69, 2520 (a = 6.29 Å). CsSnI3 and CsSnCl3 match the standard tabulated cubic constants (6.219 and 5.560 Å), whose primary measurement has not been traced for this release. The three fluorides and all twelve K-based compositions have no known cubic perovskite phase — for the K series the Goldschmidt tolerance factor lies far below the perovskite window — so no experimental constant exists to cite and those seeds are estimates. Of the 24 entries, 7 rest on a measurement, 1 is approximate against one, and 16 are estimates; each entry of the LATTICE table in code/build_dataset.py carries its own tag. Two seeds deviate from the measured value and are kept as used. CsPbI3 is seeded at 6.39 Å against an experimental 6.29 Å (+1.6%; 6.39 is the value a PBE relaxation returns), and CsPbF3 at 5.10 Å against 4.80 Å (+6.3%; CsPbF3 is R-3c at ambient conditions, not cubic). All 1,784 structures and every result in the paper were generated at these geometries, so they are reported here rather than silently corrected. They are inputs to the descriptor, not reference data: no quantity reported in the paper is a lattice constant, and the labels are computed at exactly these geometries, so the fits are internally consistent. Correction of record. Versions of build_dataset.py before 31 July 2026 attributed the seed table to “Cs/Pb halides — Stoumpos 2013, JACS; Sn/Ge — Heyns 1990 / Schueller 2018”. That attribution was wrong on every count and has been replaced by the sources above. There is no Stoumpos 2013 JACS paper; Schueller et al., Inorg. Chem. 2018, 57, 695 covers FASnI3 and FAPbBr3, which contain neither Cs nor Ge; and no Heyns 1990 paper on these structures could be located.