Parameters

MACE

mace:

type: dict

argument path: mace

MACE model

sel:

type: str | int

argument path: mace/sel

r_max:

type: float, optional, default: 5.0

argument path: mace/r_max

distance cutoff (in Ang)

num_radial_basis:

type: int, optional, default: 8

argument path: mace/num_radial_basis

number of radial basis functions

num_cutoff_basis:

type: int, optional, default: 5

argument path: mace/num_cutoff_basis

number of basis functions for smooth cutoff

max_ell:

type: int, optional, default: 3

argument path: mace/max_ell

highest ell of spherical harmonics

interaction:

type: str, optional, default: RealAgnosticResidualInteractionBlock

argument path: mace/interaction

name of interaction block

num_interactions:

type: int, optional, default: 2

argument path: mace/num_interactions

number of interactions

hidden_irreps:

type: str, optional, default: 128x0e + 128x1o

argument path: mace/hidden_irreps

hidden irreps

pair_repulsion:

type: bool, optional, default: False

argument path: mace/pair_repulsion

use pair repulsion term with ZBL potential

distance_transform:

type: str, optional, default: None

argument path: mace/distance_transform

distance transform

correlation:

type: int, optional, default: 3

argument path: mace/correlation

correlation order at each layer

gate:

type: str, optional, default: silu

argument path: mace/gate

non linearity for last readout

MLP_irreps:

type: str, optional, default: 16x0e

argument path: mace/MLP_irreps

hidden irreps of the MLP in last readout

radial_type:

type: str, optional, default: bessel

argument path: mace/radial_type

type of radial basis functions

radial_MLP:

type: list[int], optional, default: [64, 64, 64]

argument path: mace/radial_MLP

width of the radial MLP

std:

type: float, optional, default: 1

argument path: mace/std

Standard deviation of force components in the training set

precision:

type: str, optional, default: float32

argument path: mace/precision

Precision of the model, float32 or float64

NequIP

nequip:

type: dict

argument path: nequip

Nequip model

sel:

type: str | int

argument path: nequip/sel

Maximum number of neighbor atoms.

r_max:

type: float, optional, default: 6.0

argument path: nequip/r_max

distance cutoff (in Ang)

num_layers:

type: int, optional, default: 4

argument path: nequip/num_layers

number of interaction blocks, we find 3-5 to work best

l_max:

type: int, optional, default: 2

argument path: nequip/l_max

the maximum irrep order (rotation order) for the network’s features, l=1 is a good default, l=2 is more accurate but slower

num_features:

type: int, optional, default: 32

argument path: nequip/num_features

the multiplicity of the features, 32 is a good default for accurate network, if you want to be more accurate, go larger, if you want to be faster, go lower

nonlinearity_type:

type: str, optional, default: gate

argument path: nequip/nonlinearity_type

may be ‘gate’ or ‘norm’, ‘gate’ is recommended

parity:

type: bool, optional, default: True

argument path: nequip/parity

whether to include features with odd mirror parityy; often turning parity off gives equally good results but faster networks, so do consider this

num_basis:

type: int, optional, default: 8

argument path: nequip/num_basis

number of basis functions used in the radial basis, 8 usually works best

BesselBasis_trainable:

type: bool, optional, default: True

argument path: nequip/BesselBasis_trainable

set true to train the bessel weights

PolynomialCutoff_p:

type: int, optional, default: 6

argument path: nequip/PolynomialCutoff_p

p-exponent used in polynomial cutoff function, smaller p corresponds to stronger decay with distance

invariant_layers:

type: int, optional, default: 2

argument path: nequip/invariant_layers

number of radial layers, usually 1-3 works best, smaller is faster

invariant_neurons:

type: int, optional, default: 64

argument path: nequip/invariant_neurons

number of hidden neurons in radial function, smaller is faster

use_sc:

type: bool, optional, default: True

argument path: nequip/use_sc

use self-connection or not, usually gives big improvement

irreps_edge_sh:

type: str, optional, default: 0e + 1e

argument path: nequip/irreps_edge_sh

irreps for the chemical embedding of species

feature_irreps_hidden:

type: str, optional, default: 32x0o + 32x0e + 32x1o + 32x1e

argument path: nequip/feature_irreps_hidden

irreps used for hidden features, here we go up to lmax=1, with even and odd parities; for more accurate but slower networks, use l=2 or higher, smaller number of features is faster

chemical_embedding_irreps_out:

type: str, optional, default: 32x0e

argument path: nequip/chemical_embedding_irreps_out

irreps of the spherical harmonics used for edges. If a single integer, indicates the full SH up to L_max=that_integer

conv_to_output_hidden_irreps_out:

type: str, optional, default: 16x0e

argument path: nequip/conv_to_output_hidden_irreps_out

irreps used in hidden layer of output block

precision:

type: str, optional, default: float32

argument path: nequip/precision

Precision of the model, float32 or float64