Energy Offset

class graph_pes.models.offsets.EnergyOffset(offsets)

Bases: GraphPESModel

A model that predicts the total energy as a sum of per-species offsets:

\[E(\mathcal{G}) = \sum_i \varepsilon_{Z_i}\]

where \(\varepsilon_{Z_i}\) is the energy offset for atomic species \(Z_i\).

With our chemistry hat on, this is equivalent to a PerfectGas model: there is no contribution to the total energy from the interactions between atoms. Hence, this model generates zero for both force and stress predictions.

Parameters:

offsets (PerElementParameter) – The energy offsets for each atomic species.

class graph_pes.models.FixedOffset(**final_values)

Bases: EnergyOffset

An EnergyOffset model with pre-defined and fixed energy offsets for each element. These do not change during training. Any element not specified in the final_values argument will be assigned an energy offset of zero.

Parameters:

final_values (float) – A dictionary mapping element symbols to fixed energy offset values.

Examples

>>> model = FixedOffset(H=-1.3, C=-13.0)

class graph_pes.models.LearnableOffset(**initial_values)

Bases: EnergyOffset

An EnergyOffset model with learnable energy offsets for each element.

During pre-fitting, for each element in the training data not specified by the user, the model will estimate the energy offset from the data using ridge regression (see guess_per_element_mean_and_var()).

Parameters:

initial_values (float) – A dictionary of initial energy offsets for each atomic species. Leave this empty to guess the offsets from the training data.

Examples

Estimate all relevant energy offsets from the training data:

>>> model = LearnableOffset()
>>> # estimates offsets from data
>>> model.pre_fit_all_components(training_data)

Specify some initial values for the energy offsets:

>>> model = LearnableOffset(H=0.0, C=-3.0)
>>> # estimate offsets for elements that aren't C or H
>>> model.pre_fit_all_components(training_data)