OrbMol-v2: Learnable Per-Atom Electrostatics Model

OrbMol-v2 is an extension of the OrbMol architecture, trained on the Open Molecules 2025 (OMol25) and OPoly26 datasets. It incorporates learnable per-atom electrostatics, including latent charge prediction and a long-range Coulomb energy module. The model is conditioned on system charge and spin, and is designed for molecular simulations.