MACE-MP0 Refinement Examples

tricor-generated supercells refined with MACE-MP0, the universal machine-learning interatomic potential trained on the Materials Project DFT dataset.

The pipeline per (material, regime) is:

1. Voronoi tile          (cell.generate(num_steps=0))
2. Orientation refine    (cell.refine_initial_orientations)
3. Pre-MACE cleanup      (bond_relax + enforce_hard_core)
4. MACE+wall relax       (ASE LBFGS, or Langevin MD at
                          the melting point for liquid)

A per-pair soft wall (scripts/_wall_calculator.py) is added below each per-pair minimum distance to suppress MACE’s near-overlap basins; it is silent above the minimum.

Install

uv pip install torch "mace-torch>=0.3" matplotlib

The medium-mpa-0 model (~76 MB) downloads automatically on first use. Each regime page links a standalone .py reproducer.

Materials

• Copper
• Silicon
• Carbon
• Silicon dioxide
• Strontium titanate