mlff-path-optimization

name	mlff-path-optimization
description	Use this skill for managed MLFF NEB optimization after a complete locally interpolated fixed-image path has been validated.
license	project-local
allowed-tools	ls read_file write_file edit_file execute get_avail_remote_task get_remote_task_spec remote_submission remote_submission_batch

mlff-path-optimization

Overview

Execute one validated fixed-image MLFF path per stage without remote interpolation.

Quick Start

1. Finish endpoint validation, atom remapping, image-count selection, interpolation, and overlap QC locally.
2. Copy one contiguous 00.vasp through NN.vasp tree into stage/input/path/.
3. Query the selected backend directly with get_remote_task_spec(task_name="mlff_neb", template_overrides={"backend": "<enabled-backend>"}, detail="full"), then use its resolved defaults and concrete convergence schema.
4. Submit one path with remote_submission; submit two or more independent same-config paths with one remote_submission_batch.
5. Inspect the batch summary, per-path summary, energy CSV/profile, and final images. Use receipt recovery only after a returned failure.

Allowed tools

• ls
• read_file
• write_file
• edit_file
• execute
• get_avail_remote_task
• get_remote_task_spec
• remote_submission
• remote_submission_batch

Workflow

1. Require a complete local path

• Endpoint-only input is invalid. The tree must include at least one intermediate image.
• Numbering starts at 00, is contiguous and consistently zero-padded, and every image has identical atom order, cell, PBC, and constraints.
• Rework any overlap/short-distance warning before submission; remote launch cannot repair interpolation.

2. Keep one path per stage

• The only canonical tree is stage/input/path/*.vasp.
• Several independent paths become several first-level stages under a batch root. Do not put multiple path directories in one stage.
• Remote AutoNEB insertion is intentionally unsupported; interpolation stays in local preparation.

3. Choose the optimization episode

• Start with fixed-image plain NEB. Enable climbing only as an explicit refinement decision when the band already localizes the saddle sufficiently.
• Keep coarse and climbing-image refinement in separate stage copies if their results must be compared or audited.

4. Validate collected evidence

• Check convergence, maximum projected NEB force, barrier, endpoint energy difference, highest-energy image, profile shape, and per-task errors.
• Preserve final image files and use pathway analysis before treating the highest image as a transition state.

Method-critical defaults

• MACE is the registered default with float64; every deployment-enabled backend returned by get_remote_task_spec uses the same default fmax=0.05 eV/Angstrom, 300-step plain-mode contract with climbing disabled.
• Keep model/head/dispersion/precision fixed across all images and comparison paths.
• A climbing-image run is refinement, not the default first rough optimization.

Output Contract

Return:

• local image-tree/QC provenance, stage path, backend/model, fmax, steps, and climb choice;
• work_dir_rel and receipt/context identifiers;
• output/batch_summary.json, per-path summary.json, energy CSV/profile, and final-image paths;
• convergence status and the required downstream barrier/TS validation.

References

• MACE path-specific notes: references/mace.md
• Use neb-prepare before this skill and neb-analysis after collection.