quantity.property_prediction

This module contains methods to learn a molecular properties from features of a neural network used for potential energy prediction.

Pre-processing

Utility functions for initialization of the dataset and the model as well as a masked loss function for masking virtual atoms that are added to conserve array shapes.

build_dataset(targets, graph_dataset)	Builds dataset in format that is used for data loading and throughout property predictions.
init_model(prediction_model)	Initializes a model that returns predictions for a single observation.
init_loss_fn(error_fn)	Returns a loss function to optimize model parameters.

Post-processing

Property prediction at the atom level is often highly class-imbalanced. To account for this imbalance, the following functions evaluate prediction accuracy for each atom species:

per_species_results(species, ...)	Sorts per-atom results by species and returns a per-species mean.
per_species_box_errors(dataset, per_atom_errors)	Computes for each snapshot in the provided graph dataset, the per-species error.