NPT and D3 dispersion¶

Experimental

NPT and D3 carry an "experimental: not yet qualified" line in the run report until their qualification studies are recorded.

NPT¶

ensemble="NPT" in fm.md (and --ensemble NPT on the command line) is ASE 3.27's LangevinBAOAB barostat on the graph engine:

run = fm.md("NbCl5.cif", model="mh1", head="matpes_r2scan", ensemble="NPT",
            temperature_K=300, pressure_GPa=0.0, cell_mode="hydrostatic",
            steps=200_000, repeat=(3, 3, 3), dispersion="d3bj", output="npt/", seed=1)

cell_mode is "hydrostatic" or "full". For other barostats, pass an explicit integrator:

Record	ASE class
`fm.integrators.LangevinBAOAB(...)`	`LangevinBAOAB`
`fm.integrators.MTK(..., cell_mode="hydrostatic" or "full")`	`IsotropicMTKNPT`, `MTKNPT`
`fm.integrators.Berendsen(..., pressure_GPa=...)`	`NPTBerendsen`, `Inhomogeneous_NPTBerendsen` (equilibration only, non-canonical)

Cell divergence. A vapour at low pressure expands without bound. LangevinBAOAB stops the run with CellDivergedError (restart and metadata written first) once the volume leaves [V0/20, 20 V0] or a cell vector is longer than 1000 Å. Change the limits with max_volume_ratio and max_cell_length_A.

Stress convention. The force field returns the potential (virial) stress as a full 3x3 tensor in ASE's sign convention; the integrator adds the kinetic part, as ASE's get_stress(include_ideal_gas=True) does.

D3(BJ) dispersion¶

FastMACE has its own GPU D3(BJ) term. It runs inside the captured step in MD, in relaxations and in quenches, and needs no extra package.

run = fm.md(..., dispersion="d3bj")

"d3bj" is shorthand. FastMACE expands it to a complete fm.D3 record before planning: Becke-Johnson damping with the parameter set of the head's training functional (pbe for PBE-trained models, r2scan for matpes_r2scan), 40 bohr pair and coordination cutoffs, no three-body term, float64. The expanded record is printed in the report and recorded in metadata.json.

"d3bj" is refused when:

the training reference already contains dispersion (MACE-OFF, OMOL-0, MH-1's omol head), since D3 would count it twice;
the training functional is unknown (a model loaded from a path or URL). Pass an explicit record then:

run = fm.md(..., model=model, dispersion=fm.D3(functional="pbe"))

PolarMACE (POLAR-1) with D3 is refused with any record.

An explicit fm.D3 is used as given; a functional that differs from the training functional is a report warning. In TOML, a [dispersion] table holds the record's fields.

When to use it. The molecular chlorides need D3 in NPT: without it, NbCl5 swelled by 20 % at 300 K and evaporated at 600 K in our tests.