DFT dataset and qNEP models for CsGeBr3, CsSnBr3, PbTiO3

Charge-aware machine-learned interatomic potentials and their training data for the three perovskites PbTiO3, CsGeBr3 and CsSnBr3. Files Two files per system, named <system>_<file>: <system>_train.xyz – training set in extended XYZ format, with reference energies, forces and virials. <system>_nep.txt – the NEP potential, usable directly with GPUMD or calorine. Systems PbTiO3 – 345 structures, 5–320 atoms per structure, reference method VASP with r2SCAN. CsGeBr3 – 110 structures, 5–160 atoms per structure, reference method FHI-aims with r2SCAN50. CsSnBr3 – 103 structures, 5–320 atoms per structure, reference method FHI-aims with r2SCAN50. Potentials All three are fourth-generation NEPs with charge support (charge_mode 1, lambda_q 0.1), three species, radial and angular cutoffs of 8 and 4 Å, trained for 300 000 generations. The two halides additionally use a ZBL repulsive term.