DFT dataset and NEP model for MA,Ge,Sn,I3

MA(Ge,Sn)I3 neuroevolution potential and DFT training data Training data and machine-learned interatomic potential for MAGeI3, MASnI3 and their mixed compositions. * `train.xyz` Full training set, 668 structures in extended XYZ format. Reference energies, forces, virials and Born effective charges from FHI-aims.* `nep.txt` The NEP potential, usable directly with GPUMD or calorine.* `control.in` Representative FHI-aims input for a single-point reference calculation. Reference calculations FHI-aims with the hybrid meta-GGA r²SCAN50 (`libxc HYB_MGGA_XC_R2SCAN50`), `intermediate` species defaults, scalar-relativistic `atomic_zora`, `sc_accuracy_rho 1e-6`. The included k_grid is provided as an example and was generated using a k-point density parameter of 5.6, which corresponds to a reciprocal-space sampling of approximately 0.18 Å⁻¹ for the example cell. For other structures, the k_grid should be regenerated from the chosen k-point density. NEP Potential Fourth-generation NEP with charge support (`nep4_zbl_charge1`), six species (C, N, H, Ge, Sn, I), radial and angular cutoffs of 8 and 4 Å, a ZBL repulsive term with `use_typewise_cutoff_zbl 0.7`, and `lambda_q 0.1`.Fitted to the complete `train.xyz` for 300000 generations.Root-mean-square errors against the training data are 4.3 meV/atom for energies, 0.126 eV/Å for forces, 0.020 eV/atom for virials and 0.072 e for Born effective charges.