| name | mlff-screening-and-relaxation |
| description | Use this skill for MLFF single-point screening, ranking, and geometry relaxation when choosing among enabled MACE, FairChem UMA, MatterSim, or ORB-v3 backends. |
| 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-screening-and-relaxation
Overview
Prepare deterministic SP/relax stages, choose an enabled backend, and keep backend settings separate from operation settings.
Quick Start
- Read
remote-stage-layouts, then put uniquely named structures directly under a clean input/.
- Call
get_avail_remote_task, then query the intended backend directly with detail="full", for example get_remote_task_spec(task_name="mlff_sp", template_overrides={"backend": "mattersim"}, detail="full"); do not query {} first when selecting a non-default backend.
- Pass only nested
backend, backend_config, and task_config overrides returned by that query.
- Submit one stage with
remote_submission; submit two or more independent same-config stages with one remote_submission_batch. Leave submission_config.resources and .machine unset.
- Inspect
output/batch_summary.json and per-input summaries. 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. Select operation before backend
- Use
mlff_sp for energies/forces on unchanged geometries and mlff_relax when positions or the cell must be optimized.
- Query the concrete backend schema before overrides. Do not reuse MACE keys after switching to UMA, MatterSim, or ORB-v3.
- Empty overrides use the administrator-enabled default backend. If no default is effective, choose one of the returned
available_backends explicitly.
2. Build one deterministic stage
- Put files directly under
input/; do not submit recursive project trees.
- Preserve atom constraints and source provenance. Reject filename-stem collisions such as
case.xyz plus case.vasp in one stage.
- For UMA, set shared physical metadata in
backend_config.defaults and exceptions in backend_config.items, keyed by exact filenames relative to input/.
3. Group for model reuse
- One stage initializes one selected model and processes its structures sequentially.
- For similarly sized short relaxations, 30-50 structures per stage is an empirical starting point, not a probe requirement or hard limit. Use smaller groups for heterogeneous costs.
- SP can use substantially larger groups. Split only into complete, independently runnable stages.
4. Interpret as MLFF evidence
- Verify backend, model/checkpoint, provider version, precision/device, method settings, per-input errors, and actual outputs.
- Use a consistent backend/model/domain task across structures being ranked unless a difference is scientifically intentional and reported.
- Treat MLFF screening as triage unless the user explicitly accepts ML-potential accuracy for the conclusion.
Method-critical defaults
- MACE is the initial deployment default:
model=mh-1, head=omat_pbe, dispersion=false, and float64 for SP/relax. State any change that affects comparability.
- Choose dispersion explicitly for adsorption-energy comparisons and apply the same choice to clean slab, adsorbate, and references.
- Keep
relax_cell=false unless periodic-cell optimization is intended; cell relaxation requires a valid fully periodic cell.
- For UMA
omol, set charge and multiplicity-style spin explicitly. Non-omol UMA tasks require charge and spin zero.
- MatterSim-v1 is a bulk-material model; do not present surface, interface, or long-range-interaction results as quantitative without validation.
- ORB-v3 defaults to
precision=float32-high; change precision only as a declared numerical choice.
Output Contract
Return:
- operation, backend, model/checkpoint, and method-critical overrides;
- stage path or batch root with first-level stage count;
work_dir_rel, remote_context_id, submission_hash, and receipt_rel when present;
output/batch_summary.json, per-input error count, and shortlist/keep-drop rule;
- the boundary between MLFF screening and any downstream reference calculation.
References
- MACE fields and model semantics:
references/mace.md
- FairChem UMA domain and charge/spin semantics:
references/fairchem_uma.md
- MatterSim scope and model identity:
references/mattersim.md
- ORB-v3 precision and confidence metadata:
references/orb_v3.md