DeePMD-kit — инференс и валидация DP-моделей

★ 7.5 · general

deepmd-inference is a Claude Code skill that runs DeePMD-kit inference to predict energies, forces, and stresses using a trained DP model stored as a frozen file (.pb or .savedmodel). It covers three key workflows: model accuracy evaluation with `dp test`, which reports RMSE for energy, forces, and virial against DFT reference data; single-structure prediction via the Python API using `DeepPot.eval`; and geometry optimization through an ASE calculator with BFGS. Model compression via `dp compress` is also documented, delivering 3–10× faster inference with minimal accuracy loss. The skill is aimed at researchers validating a DP model before production molecular dynamics runs or quickly assessing individual atomic configurations.