DFT dataset and qNEP model for Cs(Ge,Sn)Br3

Density functional theory (DFT) dataset and charge-aware machine-learned interatomic potential for the halide perovskites CsGeBr3, CsSnBr3 and their mixed compositions Cs(Ge,Sn)Br3. The training set contains 348 structures (32,935 atoms) computed with FHI-aims using the hybrid meta-GGA r²SCAN50 functional. Energies and forces are given for all structures, stresses for 112 and Born effective charges for 11. The potential is a charge-aware neuroevolution potential (qNEP) based on NEP4 (`charge_mode 1`), trained with GPUMD 4.8. It reproduces the training data with root-mean-square errors of 1.7 meV/atom for energies, 60.3 meV/Å for forces, 13.0 meV/atom for virials and 0.071 e for Born effective charges. Files: train.xyz (training set), nep.txt (potential, for GPUMD or calorine), control.in (example FHI-aims input) and README.md.