Use this skill for MACE-based rapid screening and relaxation loops before DFT, including candidate pruning and handoff criteria.
Use this skill to run cheap MACE screening on a structure batch before spending VASP resources.
input_dir.mace_relax_batch for geometry cleanup or mace_sp_batch for static ranking.output_root outside input_dir.mace_relax_batchmace_sp_batchmace_relax_batch needs a model; it can also toggle , , and .headdispersionrelax_latticemace_sp_batch is for energy evaluation only and does not relax geometry.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.output_root inside input_dir.output_root, dispatches remotely, collects outputs, then removes the staging tree.batch_state_rel, collected stdout/stderr/status files, and any batch_summary_rel.dispersion; choose it explicitly.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.Return:
relax or sp)output_root_relbatch_state_relvasp-input-preparation only after a MACE shortlist exists; do not send the whole raw candidate pool forward by default.mace-dataset-curation and active-learning-relabel-loop.