Skip to content

Force-Field / MLIP Training

ALIGNN-FF is trained with the same train_alignn.py script but using the atomwise_alignn model. Instead of id_prop.csv, the script reads id_prop.json with entries for jid, energy, forces, and stress.

Energy is per atom

Energy values in id_prop.json must be stored per atom, not per structure.

Train from scratch

train_alignn.py \
  --root_dir "alignn/examples/sample_data_ff" \
  --config "alignn/examples/sample_data_ff/config_example_atomwise.json" \
  --output_dir="temp"

Fine-tune a pretrained ALIGNN-FF

Pass --restart_model_path with a checkpoint that uses the same model configuration:

train_alignn.py \
  --root_dir "alignn/examples/sample_data_ff" \
  --restart_model_path "temp/best_model.pt" \
  --config "alignn/examples/sample_data_ff/config_example_atomwise.json" \
  --output_dir="temp1"

Building an id_prop.json from VASP runs

A reference notebook compiles a directory of vasprun.xml files into an id_prop.json:

→ Colab: make_id_prop

Joint training with extra targets (v2024.10.30+)

From v2024.10.30, you can jointly train energy + forces + stress along with atomwise non-gradient properties (charges, magnetic moments) and optional fingerprint features. See Multi-output regression and the end-to-end configs under alignn/examples/.

The Silicon example is the best starting point before training a new force-field on your own data:

→ Colab: Train ALIGNN-FF on Mlearn (Silicon)

See also