gcnwm's picture
Align model bundle docs with supplementary archive
19f29cb
# Curated FYP Active-Learning Checkpoints
This supplementary `model/` folder stores a curated export of the strongest
final-iteration checkpoints from the MACE and NequIP active-learning study on
rMD17 ethanol. Each bundle is selected by the best final forces MAE across the
three repeated experiment seeds for its architecture/strategy family.
## Included bundles
| Bundle | Architecture | Strategy | Selected seed | Final forces MAE (meV/Å) | Final energy MAE (meV) | Files |
| --- | --- | --- | ---: | ---: | ---: | ---: |
| mace-random | MACE | RANDOM | 1 | 9.45 | 1.79 | 2 |
| mace-qbc | MACE | QBC | 3 | 8.61 | 1.69 | 8 |
| nequip-random | NequIP | RANDOM | 2 | 10.66 | 4.24 | 7 |
| nequip-qbc | NequIP | QBC | 3 | 9.53 | 6.52 | 21 |
| mace-passive | MACE | Passive control | 3 | 10.19 | 2.05 | 3 |
| nequip-passive | NequIP | Passive control | 1 | 10.13 | 2.55 | 2 |
## Notes
- `mace-qbc/` contains the full four-member committee for the selected QBC seed.
- `nequip-qbc/` contains one raw/package/compiled triplet per selected committee member.
- Per-bundle `metadata.json` files record the selected seed, final metrics, and
original relative source paths from the thesis workspace.
## Passive controls
The passive bundles are same-budget 550-label controls from the frozen v4 split. They are included to support the thesis comparison against active QBC/random methods; they are not active-learning acquisition runs.