B1 The MACE potential and the two MD engines

A molecular-dynamics (MD) engine steps positions and velocities forward one step at a time, giving a trajectory; a MACE machine-learned interatomic potential supplies the forces, here for diamond-structure silicon. MACE is a model family, not one universal checkpoint; a model suited to one chemical domain may not suit another.

The lesson pins the MACE-MP-0a small checkpoint. ALCHEMI Toolkit loads the original file, LAMMPS a separate ML-IAP export. Hashes differ because formats differ; the export needs its own validation.

model.toml
 1[model]
 2format = 1
 3id = "mace-mp-0a-small"
 4status = "locked"
 5family = "MACE-MP-0a"
 6variant = "small"
 7filename = "2023-12-10-mace-128-L0_energy_epoch-249.model"
 8url = "https://github.com/ACEsuit/mace-foundations/releases/download/mace_mp_0/2023-12-10-mace-128-L0_energy_epoch-249.model"
 9release_tag = "mace_mp_0"
10published_size_bytes = 32581838
11sha256 = "2ddb079cee0e131eaaf6912ba581b394551ead283e95c99cfe78c605d10b5736"
12license = "MIT"
13license_url = "https://github.com/ACEsuit/mace-foundations/blob/main/LICENSE"

Check both hashes in the full model identity file before preparing either engine.

Other families (MACE-MPA, MACE-OMAT, MACE-OFF) differ in training data and use; listing them does not mean the pinned ALCHEMI image or ML-IAP export can run them without new preparation and tests. See the MACE foundation-model documentation when choosing a model for a new system.

  • ALCHEMI Toolkit builds atomic data structures, calls the model and can advance a batch of independent systems: one program and GPU allocation, no forces exchanged.

  • LAMMPS is a compiled MD engine; ML-IAP calls MACE, Kokkos runs supported work on the GPU.

  • First comparison: one silicon system, one GPU, both engines. Later, ALCHEMI batches systems and LAMMPS runs separate client processes.

One MACE checkpoint feeds ALCHEMI directly and LAMMPS through an ML-IAP export.

Note

The checkpoint stays outside the SIF, so the image can be built, inspected and reused without a large embedded model or assumed redistribution rights. The model is selected explicitly and mounted read-only at run time. Changing it needs a new identity and ML-IAP export, not just a filename edit.