Training models

Training models is done by the respective packages, not through appa. appa does provide tools to extract single-atom energies, in case you do not want to include these in the training set but supply them separately.

Usage: appa convert extract-isolated [OPTIONS] XYZ_FILE OUT_XYZ

  Extract isolated-atom reference energies from an XYZ file and write a new
  XYZ file without the isolated-atom configurations.

Options:
  --help  Show this message and exit.

GRACE provides input defaults with the gracemaker -t tool. For MACE, you can use an input yaml such as

name: mace
model: ScaleShiftMACE
train_file: mytrain.xyz
valid_file: myvalid.xyz
seed: 1
r_max: 5.0
num_interactions: 2
hidden_irreps: 64x0e + 64x1o 
energy_weight: 1.0
forces_weight: 10.0
energy_key: DFT_energy
forces_key: DFT_forces
lr: 0.01
scaling: rms_forces_scaling
batch_size: 3
max_num_epochs: 500
patience: 10
eval_interval: 1
ema: true
ema_decay: 0.99
swa: true
start_swa: 300
amsgrad: true
default_dtype: float32
device: cuda
E0s:
  1: -1.20657996  # adapt single-atom energies
  8: -1.59725568
  79: -0.19135585
enable_cueq: true
restart_latest: true

Then submit a job with

mace_run_train --config input.yaml

The NequIP repo contains an example config.