q734738781/mace-screening-and-relaxation

mace-screening-and-relaxation

Overview

Use this skill to run cheap MACE screening or short MACE MD sampling on a structure batch before spending VASP resources.

Quick Start

1. Prepare a clean structure batch under input_dir.
2. Choose mace_relax_batch for geometry cleanup, mace_sp_batch for static ranking, or mace_md_batch for ASE-backed MD sampling.
3. Choose model, default_dtype, dispersion policy, and any shortlist criterion needed for the next VASP or ML step before dispatch.
4. Keep output_root outside input_dir.
5. Use the collected outputs and batch-state files to decide which candidates advance to VASP.

Allowed tools

• mace_relax_batch
• mace_sp_batch
• mace_md_batch

Workflow

1. Choose relax vs single-point deliberately

• Use lightweight local filesystem/Python checks for paths and batch shape before launching managed MACE. Submit one intentional managed batch rather than probing remote execution for setup questions local inspection can answer.
• mace_relax_batch needs a model; it can also toggle head, dispersion, and relax_lattice.
• mace_sp_batch is for energy evaluation only and does not relax geometry.
• mace_md_batch is for trajectory generation and thermal sampling, not a replacement for a converged relaxation.
• Do not compare relax and SP outputs as if they were the same screening stage.
• For geometry optimization with mace_relax_batch, keep default_dtype=float64 by default. Only switch to float32 when the user explicitly wants a cheaper, lower-rigor screening pass and the numerical looseness is acceptable.

2. Configure MACE MD through md_config

• The tool schema intentionally exposes only one free-form md_config object for MD controls. Do not expect the tool schema to list every ASE parameter.
• Keep calculator-level basics (model, head, dispersion, default_dtype) in the top-level tool arguments unless you need advanced calculator switches.
• Inside md_config, use optional groups named calculator, dynamics, thermostat, barostat, and output.
• Use only the keys needed by the chosen method; do not fill thermostat/barostat keys just because they exist in a template.
• The remote runner supplies defaults for omitted keys: NVT, 300 K, 1 fs, 1000 steps, Bussi thermostat, trajectory/log every 10 steps.

Common md_config templates:

{
  "dynamics": {"ensemble": "nve", "temperature_K": 300, "timestep_fs": 1.0, "steps": 1000}
}

{
  "dynamics": {"ensemble": "nvt", "temperature_K": 300, "timestep_fs": 1.0, "steps": 1000},
  "thermostat": {"type": "langevin", "friction_per_fs": 0.01}
}

{
  "dynamics": {"ensemble": "nvt", "temperature_K": 300, "timestep_fs": 1.0, "steps": 1000},
  "thermostat": {"type": "nhc", "tau_fs": 100, "tchain": 3, "tloop": 1}
}

{
  "dynamics": {"ensemble": "npt", "temperature_K": 300, "timestep_fs": 1.0, "steps": 1000},
  "thermostat": {"tau_fs": 100},
  "barostat": {"type": "isotropic_mtk", "pressure_bar": 1.01325, "pdamp_fs": 1000}
}

ASE mapping:

• dynamics.ensemble="nve" maps to VelocityVerlet.
• NVT thermostat.type maps as bussi -> Bussi, nhc -> NoseHooverChainNVT, langevin -> Langevin, and berendsen -> NVTBerendsen.
• NPT barostat.type maps as isotropic_mtk -> IsotropicMTKNPT, full_mtk -> MTKNPT, and berendsen -> NPTBerendsen.
• For NPT Berendsen, include barostat.compressibility_bar_inv; it is system-specific.
• For output cadence, set md_config.output with traj_interval, log_interval, log_stress, or overwrite only when defaults are unsuitable.

3. Keep input and output trees separate

• Both tools reject output_root inside input_dir.
• The runtime stages a temporary batch tree under output_root, dispatches remotely, collects outputs, then removes the staging tree.

4. Use collected evidence, not launch success alone

• Returned metadata includes batch_state_rel, collected stdout/stderr/status files, and any batch_summary_rel.
• On dispatch failure, the tool still tries to collect partial outputs; inspect those before deciding to rerun.

5. Use this skill while the workflow artifact is still materials-side

• Use this skill for structure batches, candidate ranking, geometry cleanup, and materials-side post-analysis before expensive reference calculations.
• If the screening run produces a shortlist that should become a training dataset or an active-learning update, hand off that artifact to the ML skills as the next step.

Method-critical defaults

• For adsorption-energy screening on slabs, do not silently inherit the tool default for dispersion; choose it explicitly.
• Keep the dispersion setting consistent across clean slab, gas-phase reference, and adsorbed structures when the comparison depends on relative adsorption energies.
• Unless the user explicitly asks for a no-dispersion baseline, prefer enabling dispersion when surface-adsorbate interactions or ranking sensitivity may depend on it.
• Always report whether dispersion was enabled.
• If a screening stage is intended only as a cheap geometry triage rather than an energy-ranking stage, say so explicitly.
• Treat default_dtype=float64 as the conservative default for geometry relaxation. If you deliberately downgrade to float32 for speed, say so explicitly in the run summary.
• For MACE MD, default to dynamics.ensemble="nvt" with thermostat.type="bussi" for generic sampling unless the scientific question requires energy conservation (nve) or pressure control (npt).
• For MACE MD, keep default_dtype=float32 by default for throughput. Use float64 only when explicitly checking numerical sensitivity.
• For NPT, only use structures with a real periodic cell and periodic boundary conditions in all three directions.
• For NPT Berendsen, set barostat.compressibility_bar_inv explicitly because it is system-specific; do not invent it silently for solids.
• Prefer barostat.type="isotropic_mtk" for generic NPT unless anisotropic cell fluctuations are part of the question. Use full_mtk deliberately and report it.
• Keep timestep_fs, steps, ensemble, thermostat, barostat, target pressure, target temperature, and total time visible in summaries.

Output Contract

Return:

• chosen MACE stage (relax or sp)
• for MD, chosen ensemble plus grouped calculator, dynamics, thermostat, barostat, and output controls
• output_root_rel
• batch_state_rel
• shortlist or keep/drop rule for downstream VASP handoff

References

• Use vasp-input-preparation only after a MACE shortlist exists; do not send the whole raw candidate pool forward by default.
• When the loop is ready for dataset building or retraining, hand off to mace-dataset-curation and active-learning-relabel-loop.