README.md

MACE Training Scripts and Data

This directory contains example scripts and input datasets required to train MACE models from scratch.
All hyperparameters are specified directly in the job (.job) files. Job scripts and input data are organized by dataset.

Model Types

• *invariant.job
  Invariant MACE models (max_L = 0) using model="MACE" for predicting energies
  (splitting, HOMO, LUMO, or gap).

• *embedding_invariant.job
  Invariant MACE models (max_L = 0) using model="MACE" with charge and spin embeddings
  (additional hyperparameter: --embedding_specs).

  For energy prediction targets, we additionally tested equivariant MACE models with
  max_L = 2 in the Supplementary Information.

• MACE_dipole*.job
  Equivariant MACE models (max_L = 2) using model="AtomicDipolesMACE" for predicting the dipole moment magnitude.

Data Splits

• Each dataset contains a 1-extended_xyz/ directory with MACE input files in extended XYZ format.
• To prepare extended XYZ files for MACE, all energy values are converted to eV.
• 10-fold cross-validation (CV) splits are provided (folds 0–9, for example, 0_train.xyz and 0_test.xyz).
• For Octa-MK, 10-fold CV splits are provided separately for HOMO_LUMO_gap and splitting.
  In addition, a train–validation split (as defined in the reference paper) is provided, together with the corresponding job files.
  See the --train_file and --test_file in the job scripts under Octa-MK.