Обучение ML-потенциалов DeePMD-kit

★ 7.5 · general

deepmd-train is a Claude Code skill that trains DeePMD-kit machine learning interatomic potentials on DFT data — energies, forces, and stresses — for use in molecular dynamics simulations. It supports the DPA-3 architecture (recommended) and the legacy se_e2_a descriptor, as well as fine-tuning from pretrained models using the --finetune flag with reduced learning rate and step count. The workflow is orchestrated via catgo_workflow_engine: a dp train task produces a checkpoint, followed by a dp freeze task that exports the production-ready frozen_model.pb file. Prerequisites include deepmd-kit ≥ 3.0, a CUDA-capable GPU, and training data in dpdata format — the companion data/dpdata skill handles conversion from VASP or Quantum ESPRESSO outputs. Aimed at computational chemists and materials scientists who need a reliable ML potential without manually wiring the training pipeline.