| name | pfd-distillation |
| description | Iterative workflow for model distillation. Generates a light ML force field from a pre-trained or fine-tuned model via active learning. |
| metadata | {"dependent_skills":["machine-learning-force-field","molecular-dynamics","atomic-structure","deepmd"],"tags":["iterative","active-learning","distillation"]} |
Generate a light ML force field from a pre-trained or fine-tuned model by distillation.
Standard loop:
- Structure building
- MD exploration
- Entropy-based data curation
- Labeling using a teacher model (e.g., pre-trained DPA model or fine-tuned DPA model)
- Training a student model from scratch using the labeled data
- Convergence check
- Repeat until criteria or max iterations.
Key parameters to confirm:
- task type (fine-tune/distill), max iterations (default=1), convergence criterion (default to 5 meV/atom)
- initial structure, supercell size (default to 1x1x1), perturbation settings
- MD ensemble/temperature/time/steps/save interval (default to NVT/300K/0.5ps/)
- curation max_sel(default=30)/chunk_size(default=10)
- labeling setup (predicting head for the teacher model)
- training epochs (default: 100) and split
