sigml.models package

Submodules

sigml.models.EF_Model module

sigml.models.EF_Model.get_standard_ef_model(ave_neighbor_count, em_dim=32, mul=16, interaction_layers=2, radial_layers=2, radial_neurons=64, lmax=2, weight_path=None, cutoff=4.0, device='cpu')

Get a standard \(E_f\) model

Parameters:

  • ave_neighbor_count (int) – The average number of neighbors per atom

  • em_dim (int) – Embedding dimension, the atomic one-hot representations for each atom are converted to from \(\mathbb{R}^{N_{atoms}\times 118}\) to \(\mathbb{R}^{N_{atoms}\times em\_dim}\) before being fed into network

  • mul (int) – Multiplicity of neurons within equivariant layers of the neural network. Higher means larger network

  • interaction_layers (int) – Number of convolutional interaction layers

  • radial_layers (int) – Number of radial layers for encoding of atomic positions into spherical harmonics representation

  • radial_neurons (int) – Number of neurons for each radial layer

  • lmax (int) – Maximum irrep order to consider when building equivariant layers

  • weight_path (str, optional, default = None) – The path to the weights of the model if model has been previously trained and saved

  • cutoff (float, optional, default = 4.0) – The cutoff radius of the model when considering neighbors for the graph neural network

  • device (str, optional, default = "cpu") – The device to build the model on

Returns:

model – The standard \(E_f\) model

Return type:

sigml.models.EF_Model.EF_Model

sigml.models.Sig_iws_Model module

sigml.models.Sig_iws_Model.get_standard_full_sig_model(ave_neighbor_count, leg_lmax, em_dim=64, mul=64, interaction_layers=2, radial_layers=2, radial_neurons=64, lmax=2, orbital_count=5, cutoff=4.0, weight_path=None, device='cpu')

Get a standard \(\Sigma(i\omega)\) model

Parameters:

  • ave_neighbor_count (int) – The average number of neighbors per atom

  • leg_lmax (int) – The maximum angular momentum number \(l_max\) used for the Legendre expansion

  • em_dim (int) – Embedding dimension, the atomic one-hot representations for each atom are converted to from \(\mathbb{R}^{N_{atoms}\times 118\) to \(\mathbb{R}^{N_{atoms}\times em\_dim}\) before being fed into network

  • mul (int) – Multiplicity of neurons within equivariant layers of the neural network. Higher means larger network

  • interaction_layers (int) – Number of convolutional interaction layers

  • radial_layers (int) – Number of radial layers for encoding of atomic positions into spherical harmonics representation

  • radial_neurons (int) – Number of neurons for each radial layer

  • lmax (int) – Maximum irrep order to consider when building equivariant layers

  • orbital_count (int) – Number of orbitals within correlated subspace for each atoms. Currently only supports atoms of the same orbital count

  • cutoff (float, optional, default = 4.0) – The cutoff radius of the model

  • weight_path (str, optional, default = None) – The path to the weights of the model if model has been previously trained and saved

  • device (str, optional, default = "cpu") – The device to build the model on

Returns:

model – The standard \(\Sigma(i\omega)\) model

Return type:

CrystalSelfEnergyNetwork

sigml.models.Sinf_Model module

sigml.models.Sinf_Model.get_standard_sinf_model(ave_neighbor_count, em_dim=16, mul=32, interaction_layers=2, radial_layers=2, radial_neurons=64, lmax=2, orbital_count=5, cutoff=4.0, weight_path=None, device='cpu')

Get a standard \(\Sigma_{\infty}\) model

Parameters:

  • ave_neighbor_count (int) – The average number of neighbors per atom

  • em_dim (int) – Embedding dimension, the atomic one-hot representations for each atom are converted to from \(\mathbb{R}^{N_{atoms}\times 118\) to \(\mathbb{R}^{N_{atoms}\times em_dim\) before being fed into network

  • mul (int) – Multiplicity of neurons within equivariant layers of the neural network. Higher means larger network

  • interaction_layers (int) – Number of convolutional interaction layers

  • radial_layers (int) – Number of radial layers for encoding of atomic positions into spherical harmonics representation

  • radial_neurons (int) – Number of neurons for each radial layer

  • lmax (int) – Maximum irrep order to consider when building equivariant layers

  • orbital_count (int) – Number of orbitals within correlated subspace for each atoms. Currently only supports atoms of the same orbital count

  • cutoff (float, optional, default = 4.0) – The cutoff radius of the model

  • weight_path (str, optional, default = None) – The path to the weights of the model if model has been previously trained and saved

  • device (str, optional, default = "cpu") – The device to build the model on

Returns:

model – The standard \(\Sigma_{\infty}\) model

Return type:

Sinf_Model

sigml.models.network module

class sigml.models.network.Network(*args: Any, **kwargs: Any)

Bases: Module

Equivariant neural network class

Parameters:

  • irreps_in (e3nn.o3.Irreps or None) – representation of the input features can be set to None if nodes don’t have input features

  • irreps_hidden (e3nn.o3.Irreps) – representation of the hidden features

  • irreps_out (e3nn.o3.Irreps) – representation of the output features

  • irreps_node_attr (e3nn.o3.Irreps or None) – representation of the nodes attributes can be set to None if nodes don’t have attributes

  • irreps_edge_attr (e3nn.o3.Irreps) – representation of the edge attributes the edge attributes are \(h(r) Y(\vec r / r)\) where \(h\) is a smooth function that goes to zero at max_radius and \(Y\) are the spherical harmonics polynomials

  • layers (int) – number of gates (non linearities)

  • max_radius (float) – maximum radius for the convolution

  • number_of_basis (int) – number of basis on which the edge length are projected

  • radial_layers (int) – number of hidden layers in the radial fully connected network

  • radial_neurons (int) – number of neurons in the hidden layers of the radial fully connected network

  • num_neighbors (float) – typical number of nodes at a distance max_radius

  • num_nodes (float) – typical number of nodes in a graph

forward(data: torch_geometric.data.Data | Dict[str, torch.Tensor]) → torch.Tensor

Evaluate the network

Parameters:

  • data (torch_geometric.data.Data or dict)

  • containing (data object)

  • (atoms) (- pos the position of the nodes)

  • nodes (- z the attributes of the)

  • optional

  • nodes

  • type (for instance the atom)

  • optional

  • belong (- batch the graph to which the node)

  • optional

Returns:

x – The output features of the nodes

Return type:

torch.Tensor

Module contents