MACE 0.3.16 · Si practical teaching examples

Unzip and enter mace-practical-wave260.
Run bash run.sh to recreate 7 figures and 5 CSV tables from real reference data.
This does not rerun DFT or training.

Within allocated compute resources:
  bash calculate.sh eval          Native evaluation + descriptors, 8 test structures
  bash calculate.sh foundation    Read-only MACE-MPA-0 baseline inference
  bash calculate.sh multihead     Select PBE/PBEsol heads with native CLI
  bash calculate.sh convert       Native device conversion, writes a CPU copy
  bash calculate.sh ase           Fixed-cell BFGS + 20-step NVE
  bash calculate.sh lammps        Stock LAMMPS fix external callback, single MPI rank
  bash plot-training.sh           Native training-curve plot

Submit GPU training through Slurm:
  bash train.sh scratch64-s13
  bash train.sh finetune64
  bash train.sh multihead64
  bash train.sh distributed64
Other configurations: scratch64-s29, scratch64-s47, scratch32-s13,
scratch32-stress100. Each helper creates a fresh work directory and preserves old output.
After the training job finishes:
  bash convert-new.sh work/train-scratch64-s13/models/scratch64-s13.model
The conversion job writes work/trained-cpu.model. Keep the original GPU model.

96 labels = 48 independent Si geometries x real PBE/PBEsol VASP single points.
16 extra native DFT finite-difference controls were not training labels.
32 train / 8 validation / 8 held-out test geometries per functional, same split
across heads. Small near-equilibrium structures only: not a production potential
for defects, phase transitions, melting, bond breaking or unknown compositions.

Units: total free energy F in eV/configuration, forces eV/Angstrom,
stress eV/Angstrom^3 with ASE positive tension, virial=-volume*stress in eV.
The original scratch32-s13 model has held-out stress RMSE5.374GPa. The separately
trained stress-weight100 model gives0.432GPa. Original results remain unchanged.
Finetuning improves E/F here but worsens stress; all metrics are supplied.

Foundation model is not duplicated: readonly cluster model path and SHA are in
MANIFEST.json. The own derived finetune checkpoint is supplied under the upstream
MIT license. Official GPL LAMMPS Python binding sources and license are retained;
the shared library comes from the installed lammps/30Sep2026-8173m module.
No POTCAR, WAVECAR, full OUTCAR/XML, licensed executables or private runtime logs.
