Skip to content

mliport

mliport runs machine-learned interatomic potentials — MACE, DPA, GRACE, UMA and NEP — from one command line, one terminal UI and one Python API. You hand it a structure and a checkpoint; it relaxes the structure, runs molecular dynamics, finds a reaction barrier with NEB, searches low-energy configurations of an adsorbate on a slab, or analyzes a trajectory. Results come back in files anyone who has used VASP can open: OUTCAR, OSZICAR, CONTCAR and XDATCAR, next to JSON records of exactly what was run.

It is a workflow layer. It is not a model, not a trainer and not a DFT code: the files are VASP-shaped, but there is no SCF, no k-point mesh and no POTCAR, and an absolute energy only means something inside the one model that produced it. Whether the potential is right for your chemistry is up to the checkpoint you pick and the checks you run on it.

What it does for you

  • One workflow, five backends. Single point, relaxation (fixed or variable cell), NVE/NVT molecular dynamics, NEB and CI-NEB, configuration search, batch sweeps and a job queue behave the same whichever model sits underneath.
  • No silent guessing. There is no default backend. A misspelled option, or one that belongs to another backend, stops the run instead of being ignored. A DPA checkpoint with several heads asks which one you mean.
  • A paper trail. Every run records the model file and its hash, the resolved configuration (and where each value came from), the device that was really used and the mliport version.
  • Analysis that admits when the data is too short. MSD, transport, RDF, VACF, GEMDAT electrolyte analysis and Arrhenius fits report an explicit "insufficient sampling" instead of a number you should not trust.

Try it

git clone https://github.com/hydrogen1222/mliport
cd mliport
./scripts/install_mliport.sh --engines mace --device auto

.venv-mace/bin/mliport sp examples/structures/li10gep2s12_primitive.vasp \
  --model /path/to/mace-omat-0-medium.model --model-type mace --task bulk

The quickstart walks through this line by line, including where to get the checkpoint.

Where to go next

If you want to… Read
install it and run a first calculation Quickstart
understand environments, Python versions and GPUs Installation
check a model on a system with a known answer Test systems
pick a model, or see which ones were validated Models, Backends
set options through the CLI, an INCAR file or settings.ini Configuration
run a relaxation, MD, NEB, a configuration search or a batch Workflows
analyze a trajectory Analysis
know what each output file contains Output files
see what has actually been tested Validation
fix an error Troubleshooting
move over from the old mlipx name Migration
cite mliport and the models it runs Citations
work on mliport itself Development