nvalchemi.dynamics.mep.IDPPModel#

class nvalchemi.dynamics.mep.IDPPModel[source]#

Evaluate the image-dependent pair potential for reaction-path images.

Linear Cartesian interpolation can place atoms unnaturally close together or produce uneven changes in interatomic distances. IDPP provides a better initial guess for reaction-path calculations, such as NEB, by relaxing the interior images toward smoothly varying pair distances before evaluating them with a physical potential [1].

For image fraction \(s \in [0, 1]\), the target distance for atoms \(i\) and \(j\) is interpolated between the endpoint distances:

\[d_{ij}^{\mathrm{target}}(s) = (1-s)d_{ij}^{(0)} + s d_{ij}^{(1)}.\]

The artificial IDPP energy for an image is

\[E_{\mathrm{IDPP}} = \sum_{i<j} \frac{\left(d_{ij} - d_{ij}^{\mathrm{target}}\right)^2}{d_{ij}^4},\]

which weights close atom pairs more strongly. Forces are the negative coordinate gradient, \(\mathbf{F}_i = -\partial E_{\mathrm{IDPP}} / \partial \mathbf{r}_i\).

Call prepare_idpp_targets() on each grouped path batch before evaluation. Pair displacements use minimum-image wrapping for periodic images. IDPP energies are optimization objectives, not physical potential energies.

References

compute_embeddings(data, **kwargs)[source]#

Raise because the analytic IDPP model has no embeddings.

Parameters:
Return type:

AtomicData | Batch

direct_derivative_keys()[source]#

Return outputs computed analytically by the IDPP model.

Returns:

The analytic force output.

Return type:

set[str]

property embedding_shapes: dict[str, tuple[int, ...]]#

Return no embedding shapes because IDPP has no learned features.

extra_repr()#

Format the model config for nn.Module.__repr__.

Parameters:

self (Any)

Return type:

str

forward(data, **kwargs)[source]#

Compute active IDPP outputs for every image.

Parameters:
Returns:

Active per-image energies and per-atom forces.

Return type:

ModelOutputs

Raises:
  • TypeError – If data is not a Batch.

  • KeyError – If data has not been prepared with IDPP target distances.

  • RuntimeError – If any distinct atom pair is coincident within machine precision.