ml-mlip-automl

Files (1)

• SKILL.md

SKILL.md (first lines)

# LLM-Coupled MLIP Hyperparameter Search (AutoML)

## Goal
To automate the discovery of optimal training hyperparameters (learning rate, batch size, backbone freezing) by coupling the agent to an iterative execution loop. Instead of conducting exhaustive grid searches, the LLM actively interprets validation MAE trajectories per epoch and strategically deduces the next optimal parameter permutation to evaluate.

## The Agentic Workflow

Instead of calling a monolithic training tool, this SKILL dictates how you (the Agent) should orchestrate a search using the individual foundation fine-tuning SKILLs (`ml-mace-finetune`, `ml-matgl-finetune`, `ml-fairchem-finetune`).

### **Execution Protocol**

1. **Initial Anchor**: Select a conservative parameter set based on the dataset size (e.g., small dataset: freeze backbone, lower LR).
2. **Execute Fine-Tuning**: Create a dedicated subdirectory for the run (e.g., `search_run_lr1e-3_frozen`) and execute the respective MLIP's `prepare_*_data.py` and training runner.
3. **Parse and Evaluate**: Read the resulting `training_history.json` directly into your context. Analyze the validation loss and MAE curves (slope, signs of overfitting/forgetting).
4. **Iterate**: Deduce the next logical parameter adjustment based on the evaluation (e.g. "Validation MAE plateaued early. Action: Unfreeze backbone & lower LR to 1e-4").
5. **Terminate**: Halt and report the best configuration once a performance target is met or degradation occurs repeatedly.

## Hyperparameter Search Spaces

Different MLIP frameworks expose different tunable parameters. When deducing your next move, restrict your search space to the available features:

| Feature | `ml-fairchem-finetune` | `ml-matgl-finetune` | `ml-mace-finetune` |
|:--------|:-----------------------|:--------------------|:-------------------|
| **Freeze backbone** | `--freeze-backbone` | `--freeze-backbone` | `--freeze-backbone` |
| **Re-init head** | N/A (Auto) | `--reinit-head` | `--reinit-head` |
| **Learning Rate** | `--lr` (Default: 4e-4) | `--lr` (Default: 1e-3) | `--lr` (Default: 1e-4) |
| **Batch Size** | `--batch-size` | `--batch-size` | `--batch-size` |
| **Scheduler** | Cosine with Warmup | Cosine or Plateau | Exponential or Plateau |
| **Force Loss Ratio** | `--force-weight` | `--force-weight` | `--forces-weight` |
| **Energy Loss Ratio** | `--energy-weight` | `--energy-weight` | `--energy-weight` |
| **Loss Criterion** | Huber/L1/L2 | Huber/L1/L2 | universal/weighted |

> [!TIP]
> **Constraints & Best Practices**
> - **Key hyperparameters to explore**: Learning rate, backbone freezing, force loss ratio, energy loss ratio, and loss criterion.
> - **Batch Size**: Should generally be set to around 8.
> - **Trial Count**: Spend around 10 trials on different sets of hyperparameters.
> - **Epoch Count during Search**: Use only ~10 epochs per trial on a small dataset to quickly gauge convergence.
> - **Dataset Size**: If the dataset is already small (smaller than 1000 structures), we don't need to test on a subset. Just use the full dataset.
> - **Final Training**: Afterwards, use the best set of hyperparameters discovered to train on the full dataset for many epochs with early stopping.

## Invoking the Search (User Prompt Template)

To initiate this workflow, users should provide an anchor dataset and instruction similar to:

> "I have a dataset at `private_data/my_dataset.json`. Please orchestrate an LLM-driven hyperparameter search over learning rates, batch sizes (~8), freezing strategies, and loss ratios. Execute one run at a time for ~10 epochs, read the `training_history.json`, explain your logic for the next permutation, and stop after ~10 trials to find the best combination. Finally, train the best model for many epochs with early stopping."

---

**Author:** Bowen Deng
**Contact:** [GitHub @learningmatter-mit](https://github.com/learningmatter-mit)

From learningmatter-mit/AtomisticSkills, MIT.

More from learningmatter-mit/AtomisticSkills