Charge-aware knowledge distillation of foundation interatomic potentials

Knowledge distillation compresses foundation interatomic potentials into fast students by transferring energies, forces, and sometimes Hessians. Per-atom charge is generally not part of what is transferred, so the charge of a charge-aware foundation model is discarded or left latent in the student. Here we introduce charge-aware distillation, in which an explicit per-atom charge field, referenced to a named partition convention (DDEC6), is distilled jointly with energies and forces under an enforced total charge. A lightweight moment-tensor student reproduces the charge-equilibrated teacher QET's per-atom charges to 1.4-5.7 me (milli-electrons) across five materials, an order of magnitude inside the teacher's 36.4 me deviation from DFT. A closed-form charge-equilibration head helps only when the total charge varies, and even there a scalar total-charge input closes the gap. Fine-tuning the teacher on DFT charges extends the recipe to spinel-oxide/water interfaces, which the student preserves. Exported to LAMMPS, the students carry the charge field to 50-67 million atoms on one 256 GB CPU node, $10^{3}\times$ the teacher's reach at matched memory. They deliver charge-resolved molten-salt MD that is verified against the teacher on every frame.