Relaxation¶
mliport opt minimizes the forces on the atoms, and optionally the stress on
the cell, with an ASE optimizer running on whichever backend you chose.
.venv-mace/bin/mliport opt structure.cif \
--model mace-omat-0-medium.model --model-type mace \
--fmax 0.05 --max-steps 50 --output runs/opt
Options you will use¶
| Option | Meaning |
|---|---|
--fmax |
force convergence threshold in eV/Å (default 0.05) |
--max-steps |
hard step limit (default 500) |
--optimizer |
FIRE (default), BFGS or LBFGS |
--cell-opt |
relax the cell too; needs a model that provides stress |
--fix-symmetry |
preserve the crystal symmetry during the relaxation |
--model, --model-type, --task, --head, --device |
model and backend selection, as for a single point |
--cell-opt is strict about what it accepts. It needs a model with stress
support, a fully periodic structure, and a bulk task. For an isolated molecule,
partial periodicity or a model without stress it stops with an explanation
instead of quietly relaxing the positions only.
What you get¶
CONTCAR is the final structure. OUTCAR and OSZICAR hold the energy and
force trace, and mliport_results.json records whether the relaxation
converged, how many steps it took, and the final energy and forces, alongside
resolved_config.json and run.log.
Converged or not¶
If the step limit is reached before --fmax, the run exits with status 2,
keeps the partial trajectory and records the reason. That is a relaxation that
did not converge, not a successful one, and the structure it leaves should be
inspected before you use it. A backend or runtime error exits with 1, keeps
the partial results and records the error in mliport_results.json.
What a relaxed structure means¶
The result is a minimum of this model's potential energy surface, not a DFT minimum, and it is only meaningful within one model identity and one reference. Cell optimization changes the reference volume, so look at the final cell before comparing energies.