xnn.dnn.models.hdnnp.NNP1G#

class xnn.dnn.models.hdnnp.NNP1G(species, hidden=(30, 30), activation='tanh', coordinates='distances', length=1.0, symmetrize=False, max_permutations=5040, cutoff=1.0)[source]#

Bases: InteratomicPotential

First-generation neural network potential: one network for the whole system.

The low-dimensional potentials that preceded the HDNNP map a fixed set of coordinates of a fixed system directly to its energy with a single feed-forward network (Blank et al. 1995, Lorenz et al. 2004; Fig. 1 and eq. 1 of Behler and Parrinello 2007). Here the coordinates are the n (n - 1) / 2 interatomic distances of a system of n atoms in a fixed order (or a function of them), so the energy is invariant to translations and rotations. As in those potentials the network is tied to its number of atoms and their order; symmetrize=True makes it invariant under permutations of like atoms by averaging the network over all of them (exact, at one evaluation per permutation).

Parameters:

  • species (sequence of int) – Atomic number of every atom, in the order the structures list them.

  • hidden (sequence of int, optional) – Hidden-layer widths, by default (30, 30).

  • activation (str, optional) – Hidden-layer activation, by default "tanh" (the output is linear).

  • coordinates (str, optional) – Network inputs: "distances" (default) r_ij, "inverse_distances" 1 / r_ij or "exponential" exp(-r_ij / length).

  • length (float, optional) – Length scale of the "exponential" coordinates, by default 1.0.

  • symmetrize (bool, optional) – Average over every permutation of like atoms, by default False.

  • max_permutations (int, optional) – Largest permutation group accepted with symmetrize, by default 5040.

  • cutoff (float, optional) – Neighbour-list radius of the data pipeline; the model reads positions only. By default 1.0.

Variables:

  • n_atoms (int) – Number of atoms of the system.

  • net (torch.nn.Sequential) – The network.

  • input_std (input_mean,) – Standardization of the inputs (fit_scaling()).

  • energy_scale (energy_shift,) – E = energy_shift + energy_scale * net(x).

inputs(data)[source]#

The coordinates of every structure, (B, n_atoms (n_atoms - 1) / 2).

Parameters:

data (AtomicGraph)

Return type:

Tensor

forward(data)[source]#

Energy of each structure; the per-atom energies are its equal shares.

Parameters:

data (AtomicGraph) – Batched structures of the model’s atoms.

Returns:

node_energy and energy.

Return type:

dict[str, Tensor]

fit_scaling(graphs)[source]#

Standardize the inputs and the energy on training structures.

The input statistics are taken over every permuted copy, so they are the same for like-atom pairs and keep a symmetrized model invariant.

Parameters:

graphs (iterable of AtomicGraph) – Training structures with energies.

Return type:

None

classmethod from_config(cfg)[source]#

Build from cfg.extra: species (one entry per atom, in order), hidden, activation, coordinates, length, symmetrize, max_permutations; cfg.cutoff sizes the data pipeline’s graph.

Return type:

NNP1G