MACE

class mlptrain.potentials.mace.mace.MACE(name: str, system: System, foundation: str | None = None)

Bases: MLPotential

__init__(name: str, system: System, foundation: str | None = None) → None

MACE machine learning potential

Parameters:

foundation –

(str) Either the shortcut of the foundation model used in fine-tunning like “medium_off” for MACE-OFF(M), “medium” for MACE-MP-0(M), or the path to the foundation model.

Here, naive fine-tuning is default without any other argument specified.

To initiate multi-head fine-tuning, specify mace_params[‘pt_train’]=/path/to/replay/dataset For MACE-MP models, the replay dataset is provided by MACE through setting mace_params[‘pt_train’]=’mp’ Some Replay datasets could be accessed here: https://github.com/ACEsuit/mace-foundations/releases More details on https://github.com/ACEsuit/mace/tree/main?tab=readme-ov-file#pretrained-foundation-models

property args: Namespace

Namespace containing mostly default MACE parameters

property ase_calculator: ASECalculator

ASE calculator for MACE potential

property batch_size: int

Batch size of the training set

property filename: str

Name of the file where potential is stored

property get_E0s

property requires_atomic_energies: bool

Does this potential need E_0s for each atom to be specified

property requires_non_zero_box_size: bool

MACE cannot use a zero size box

property valid_fraction: float

Fraction of the whole dataset to be used as validation set