| name | guide-for-model-training |
| description | General guideline for MLFF model training. Describes the decision flow for choosing between pre-trained model validation, fine-tuning on available data, and launching a full PFD workflow. |
| metadata | {"dependent_skills":["machine-learning-force-field","dft-calculation","deepmd","abacus"],"tags":["model","fine-tune","machine-learning-force-field"]} |
Model training is a dynamic, iterative process. Follow this decision flow and revisit earlier steps if results are unsatisfactory.
Decision Flow
Stage 1 — Check the Pre-trained Model
- Always try the pre-trained model first.
- Query the available domain datasets. If not sure, ALWAYS ask user.
- Query the selected domain datasets for the target system. If it is available, validate the pre-trained model for it.
- If validation passes, stop here — no training is needed.
Stage 2 — Fine-tune on Available Data
- If dataset available and pre-trained model does not perform adequately, fine-tune it using the existing dataset.
- Prefer this over more complex workflows whenever data is available.
- Evaluate the fine-tuned model; if results are sufficient, stop here.
Stage 3 — PFD Workflow (Last Resort)
- If no available dataset or fine-tuned model is insufficient, launch the PFD workflow.
- PFD iteratively generates new data via MD exploration, labels it, and retrains until convergence.
- See
pfd-finetuning guide for full details.
General Notes
- Always evaluate the model after each step before proceeding to the next.
- Prefer the simplest effective strategy — avoid PFD if fine-tuning suffices.
- Each step may require multiple rounds of interaction and parameter adjustment.