q734738781/mace-md-sampling

mace-md-sampling

Overview

Use this skill for MACE MD in dynamics_worker. MACE relaxation, single-point ranking, and NEB/path optimization stay in materials_worker; MACE dataset construction, training, and evaluation stay in ml_worker.

Quick Start

1. Verify the starting structure or structure batch path from the materials workflow.
2. Prepare a clean MACE MD stage with input/ and grouped controls in params/md_params.json.
3. Submit with remote_submission(task_name="mace_md_dir") for one stage, or remote_submission_batch when each first-level child is one complete MACE MD stage.
4. Inspect status.json, stdout.log, stderr.log, and output/batch_summary.json after execution.
5. Summarize trajectories with md_trajectory_summary or analyze_trajectory, and write a focused script for system-specific observables.

Allowed tools

• get_avail_remote_task
• remote_submission
• remote_submission_batch
• md_trajectory_summary
• analyze_trajectory
• execute

Workflow

1. Prepare the stage layout

• Stage directory must contain input/ with ASE-readable structures.
• Write grouped MD controls to params/md_params.json; pass params={"params_path": "params/md_params.json"} to remote_submission.
• Use the task catalog entry mace_md_dir. Do not submit MACE MD through mace_relax_dir or mace_sp_dir.

stage/
  input/
    POSCAR or *.vasp/*.cif/*.xyz
  params/
    md_params.json

2. Keep MD controls grouped

• Use one md_config object with optional calculator, dynamics, thermostat, barostat, and output groups.
• Omitted defaults are NVT, 300 K, 1 fs, 1000 steps, Bussi thermostat, and trajectory/log every 10 steps.
• Set dynamics.ensemble to nve, nvt, or npt; choose thermostat/barostat keys only when the method needs them.
• For NPT Berendsen, set barostat.compressibility_bar_inv explicitly.

3. Submit through managed MACE MD

• Use only task_name="mace_md_dir" for MACE MD execution.
• Managed MACE GPU tasks default to device="auto" so exploratory jobs can still produce results when CUDA is unavailable. After completion, inspect status.json or output/batch_summary.json and report the actual device used.
• Use device="cuda" only when the user explicitly asks for GPU validation or a GPU-required production run.

4. Analyze before interpreting

• Start from run-health evidence: completion, errors, final structure, trajectory files, log files, total time, and actual device.
• Use md_trajectory_summary or analyze_trajectory for generic frame/RDF/MSD artifacts when they match the requested observable.
• Write a focused script for residence time, adsorption/desorption, reaction labels, or region-specific observables.

Method-critical defaults

• Default to dynamics.ensemble="nvt" with thermostat.type="bussi" for generic thermal sampling unless the scientific question requires energy conservation (nve) or pressure control (npt).
• Keep calculator.default_dtype="float32" by default for MD throughput. Use float64 only when explicitly checking numerical sensitivity.
• For NPT, use only structures with a real 3D periodic cell; prefer barostat.type="isotropic_mtk" unless anisotropic cell fluctuations are part of the question.
• Keep timestep, steps, ensemble, thermostat/barostat, targets, total simulated time, dtype, device, and dispersion visible in summaries.
• Do not treat a completed short MACE MD run as converged diffusion or mechanistic evidence without a credible production window.

Output Contract

Return:

• MACE MD stage path
• staged params/md_params.json path
• selected ensemble and grouped calculator, dynamics, thermostat, barostat, and output controls
• work_dir_rel
• remote_context_id, submission_hash, and receipt_rel when present
• status.json, output/batch_summary.json, and trajectory/log summary paths
• whether the result is production evidence or exploratory sampling

References

• Local source note: references/mace_md_sampling_reference.md
• Project runner: catmaster/remote/gpu/mace_md.py
• Stage layout: skills/execution/remote-stage-layouts/SKILL.md#mace_md_dir