ml-committee-uncertainty

What problem does it solve?

This Skill quantifies epistemic uncertainty in a MACE machine-learning interatomic potential by measuring disagreement across a committee of independently trained models, so you can identify out-of-distribution structures that should be verified or labeled with DFT.

Core Features & Use Cases

• Committee-based uncertainty (energy and forces): Computes energy standard deviation (meV/atom) and force variance/std (meV/Å) from multiple MACE checkpoints.
• Automatic DFT flagging: Flags structures whose uncertainty exceeds configurable energy/force thresholds and saves them for labelling.
• Active-learning ready outputs: Writes per-structure summaries (uncertainty_summary.json), flagged structure files, and uncertainty distributions for analysis.
• Use Case: Run uncertainty screening on an MD trajectory to find when your system drifts beyond the MLIP training distribution, then send the flagged frames to a DFT labelling workflow.

Quick Start

Run committee inference on your trajectory by providing the input structures file and at least three MACE checkpoint paths to produce flagged structures for DFT verification.