neb-calculation

name	neb-calculation
description	Use this skill for the execution stage of NEB and dimer workflows, especially the detailed run protocol for plain-NEB to CI-NEB refinement or NEB/frequency/dimer refinement.

neb-calculation

Overview

Use this skill once the pathway inputs are already prepared and the question is how to run them robustly. It covers the two common execution branches: plain-NEB -> CI-NEB and NEB or TS guess -> frequency/mode guess -> dimer. For image generation and dimer input assembly, use neb-prepare. For post-run barrier interpretation and QC, use neb-analysis.

Quick Start

1. Decide whether the task is a NEB refinement workflow or a dimer-refinement workflow.
2. Confirm the prepared root, execution branch, climb policy, dtype, output root, and whether the run is coarse convergence or refinement.
3. Dispatch the prepared calculation root through remote_submission or remote_submission_batch, normally with task_name="vasp_execute_neb" for VASP pathway work.
4. For MLFF pathway optimization, use mlff-path-optimization and submit the complete local path with task_name="mlff_neb".
5. Keep coarse convergence and refinement as separate episodes instead of mixing them into one opaque run root.

Allowed tools

• get_avail_remote_task
• get_remote_task_spec
• remote_submission
• remote_submission_batch

Workflow

1. Default NEB route: coarse plain-NEB, then CI-NEB

• Do not use managed NEB execution as a diagnostic for missing preparation. If image trees, endpoint provenance, or dimer modes are uncertain, return to neb-prepare first.
• A NEB tree that carries an overlap/short-distance preparation warning is very likely to contain abnormal interpolation. In CatMaster tools, short_distance_count > 0 means at least one image has a minimum interatomic distance below the configured threshold, default 0.8 Å; verify or remediate the image tree before deciding to submit.
• Start from a prepared NEB root with climbing image disabled.
• Run a coarse plain NEB first to localize the band and reduce gross path noise.
• Once the band is reasonably converged, restart from those coarse-converged images with climbing image enabled for CI-NEB refinement.
• Do not present a direct first-shot climbing-image run as the default; use it only when the saddle is already well localized and the task has a reason to skip the coarse stage.

2. Keep refinement episodes explicit

• Keep the coarse plain-NEB run and the CI-NEB refinement run in separate output roots.
• Do not change image count, endpoint definition, and convergence strategy all at once if you expect to compare outcomes scientifically.
• Record which stage produced which artifact; “NEB completed” is not enough.

3. Dimer route: NEB or TS guess, then frequency-guided mode, then dimer

• Use this route when the goal is TS refinement rather than only a band profile.
• A common pattern is: coarse NEB or TS guess -> frequency/mode estimate -> dimer refinement.
• The practical role of the frequency/mode step is to produce a chemically meaningful reaction-direction guess before launching the dimer.
• If the initial mode guess is poor, fix the mode selection rather than blindly rerunning the same dimer job.

4. Dispatch details for VASP pathway jobs

• Prefer task_name="vasp_execute_neb" for NEB or dimer-style VASP runs so they use the dedicated submission preset rather than the generic VASP one.
• Report the prepared root, task name, and output root together.
• Treat launch success as only one checkpoint; it does not prove the pathway is physically meaningful.

5. Managed MLFF NEB

• Use task_name="mlff_neb"; for a non-default backend, query get_remote_task_spec(task_name="mlff_neb", template_overrides={"backend": "<enabled-backend>"}, detail="full") before setting backend or convergence overrides.
• Copy one locally interpolated flat image tree into stage/input/path/, with contiguous 00.vasp through NN.vasp files. Endpoint-only input is invalid.
• Use remote_submission_batch for multiple paths, one complete path stage per first-level child.
• Keep float64 by default for MACE pathway optimization and use fixed-image plain mode. Remote AutoNEB interpolation is unsupported.
• Keep climb as an explicit refinement decision rather than an implicit default.

Method-critical defaults

• plain-NEB -> CI-NEB is the default robust VASP barrier workflow.
• Do not enable climbing image in the first rough-convergence stage unless there is a clear reason.
• If the workflow takes the dimer branch, keep the mode-generation evidence explicit; dimer runs without a credible reaction direction are weakly interpretable.
• Keep default_dtype=float64 as the default for MACE geometry/path optimization; only use float32 when the run is explicitly exploratory.
• For pathway work, separate preparation, execution, and analysis artifacts cleanly.
• Before dispatch, verify endpoint ordering and image-distance QC; launch success does not repair an overlapped band.

Output Contract

Return:

• which branch was run (plain_neb, cineb_refinement, dimer_refinement, mace_neb, or similar)
• execution root and submission evidence
• whether the run was coarse convergence or refinement
• any follow-up action the next stage should take

References

• Use neb-prepare before this skill if the image tree or dimer inputs are not ready yet.
• Use neb-analysis after collection to interpret the barrier, profile shape, and common pitfalls.
• Use mlff-path-optimization for the complete operation-specific managed MLFF protocol.