---

name: interfaces-md description: "Run trained SchNetPack models through ASE calculators, ensemble uncertainty, molecular dynamics configs, TorchScript deployment, and LAMMPS interface guidance." disable-model-invocation: true metadata: disco-role: operating license: NOASSERTION

SchNetPack Interfaces and Molecular Dynamics

Use this sub-skill when the task starts from an already trained SchNetPack model and asks to run, relax, deploy, or simulate atomistic systems.

Route Here For

• ASE calculators with SpkCalculator, AtomsConverter, AseInterface, SpkEnsembleCalculator, AbsoluteUncertainty, or RelativeUncertainty.
• Single-point energy/force/stress prediction, ASE optimization, normal modes, short ASE MD, or batchwise relaxation planning.
• spkmd command construction, Hydra MD config files, checkpoint/restart handling, thermostat/barostat/integrator choices, RPMD, and CPU fallback.
• TorchScript deployment for the SchNetPack LAMMPS pair style with scripts/deploy_for_lammps.py.
• LAMMPS interface advice, including pair_style schnetpack, pair_coeff atom-type mapping, cutoff metadata, and build-risk warnings.

Route Elsewhere

• Model architecture, output modules, response-force/stress training, and model checkpoints: ../models-atomistic/SKILL.md.
• Training runs, Hydra training configs, callbacks, and checkpoint selection: ../training-configs/SKILL.md.
• Dataset creation, ASE databases, property units, and unit conversion before training: ../data-pipelines/SKILL.md.
• External LAMMPS compilation, source-tree patching, or cluster build-system changes: provide guidance only unless the user explicitly authorizes environment mutation.

Fast Decision Guide

• For a Python workflow using ASE Atoms: start with ASE interface.
• For production-style MD from CLI/config: start with MD CLI.
• For LAMMPS deployment: read LAMMPS deployment, then use scripts/deploy_for_lammps.py only after checking model compatibility.
• For failures or unsafe requests: use troubleshooting.

Safety Defaults

• Prefer device="cpu" for examples and dry runs unless the user confirms CUDA availability.
• Keep MD examples bounded; do not launch long MD, GPU workloads, dataset downloads, native tests, or LAMMPS builds by default.
• Treat SchNetPack's upstream LAMMPS patch helper as reference-only because it mutates an external LAMMPS source tree and build files.