Auxiliary outputs¶

These outputs, which are generally idenfified by the mtt::aux:: prefix, represent additional information that the model may provide. They are not conventional trainable outputs, in the sense that they do not correspond to training targets. For example, such quantities might be the internal representation of the model, uncertainty estimates, or non-trainable quantities.

The following auxiliary outputs are currently supported by one or more architectures in the library:

mtt::aux::{target}_last_layer_features: The representation
of the model at the last layer, before the final linear transformation to produce target target. If the model produces multiple targets, the corresponding representations might be different. This output differs from the features output which is the same for all targets of a model.
features: A common representation of the model for all targets. Generally, this will correspond to the last representation before the decoder(s), or heads, of the model.
mtt::aux::{target}_uncertainty and mtt::aux::{target}_ensemble: Auxiliary outputs related to uncertainty estimation. For the energy output, mtt::aux::energy_ensemble is instead named energy_uncertainty. For the moment, these are only accessible through the LLPR module, which itself requires the use of the mtt::aux::{target}_last_layer_features output.

The following table shows the architectures that support each of the auxiliary outputs:

Auxiliary output	SOAP-BPNN	PET	GAP	NanoPET
`mtt::aux::{target}_last_layer_features`	Yes	Yes	No	Yes
`features`	Yes	Yes	No	Yes

The following tables show the metadata that will be provided for each of the auxiliary outputs:

mtt::aux::{target}_last_layer_features¶

Metadata for last-layer features¶
Metadata	Names	Description
keys	`"_"` if invariant or non-symmetrized, `["o3_lambda", "o3_sigma"]` if equivariant	if invariant or non-symmetrized, a single dimension named `"_"`, with a single entry set to `0`, or, if equivariant, two dimensions named `"o3_lambda"` and `"o3_sigma"` with the corresponding O3 equivariant indices.
samples	`["system", "atom"]` or `["system"]`	if a `per_atom` output is requested, the sample names will be `["system", "atom"]`, otherwise they will be `["system"]`. `"system"` ranges from 0 to the number of systems given as input to the model. `"atom"` ranges between 0 and the number of atoms/particles in the corresponding system.
components	No components if invariant or non-symmetrized, `["o3_mu"]` if equivariant	Nothing if invariant or non-symmetrized, the O3 equivariant `mu` number if equivariant.
properties	`"properties"`	the last-layer features have a single property dimension named `"properties"`, with entries ranging from 0 to the number of features in the last layer.

features¶

See the feature output in metatensor.torch.atomistic.