smact.property_prediction.roost module

ROOST (Representation Learning from Stoichiometry) predictor implementation for composition-only property prediction.

ROOST-based property predictor implementation.

This module provides the RoostPropertyPredictor class for predicting material properties from chemical composition using pre-trained ROOST (Representation Learning from Stoichiometry) models.

class smact.property_prediction.roost.predictor.RoostPropertyPredictor(property_name: str, fidelity: str | None = None, model_name: str | None = None, model_path: str | Path | None = None, device: str = 'cpu', elem_embedding: str = 'matscholar200', batch_size: int = 128, **kwargs: Any)

Bases: BasePropertyPredictor

Property predictor using pre-trained ROOST models.

ROOST (Representation Learning from Stoichiometry) predicts material properties from composition alone using message passing on stoichiometric graphs. This class loads pre-trained checkpoints and performs inference.

Example

>>> predictor = RoostPropertyPredictor(property_name="band_gap")
>>> predictions = predictor.predict(["NaCl", "TiO2"])

>>> # With fidelity selection
>>> predictor = RoostPropertyPredictor(property_name="band_gap", fidelity="hse06")

>>> # With uncertainty quantification
>>> result = predictor.predict(["NaCl"], return_uncertainty=True)
>>> print(result.uncertainties)

available_properties

Class attribute listing all supported properties.

Type:

ClassVar[list[str]]

elem_embedding

Element embedding type used by the model.

batch_size

Batch size for inference.

available_properties: ClassVar[list[str]] = ['band_gap']

classmethod from_checkpoint(checkpoint_path: str | Path, property_name: str, device: str = 'cpu', **kwargs: Any) → RoostPropertyPredictor

Create predictor from a local checkpoint directory.

Parameters:

• checkpoint_path – Path to checkpoint directory containing model.json, model.pt, and state.pt files.

• property_name – Name of the property.

• device – Device to run on (“cpu” or “cuda”).

• **kwargs – Additional arguments passed to constructor.

Returns:

Initialised RoostPropertyPredictor.

predict(compositions: str | list[str], return_uncertainty: bool = False, validate_smact: bool = True) → ndarray | PredictionResult

Predict property values for given compositions.

Parameters:

• compositions – Single composition string or list of composition strings (e.g., “NaCl”, [“TiO2”, “GaN”]).

• return_uncertainty – If True, return PredictionResult with uncertainty estimates (requires robust model).

• validate_smact – Whether to validate compositions using SMACT. Set to False to skip validation.

Returns:

Property predictions as numpy array, or PredictionResult object if return_uncertainty is True.

property supported_properties: list[str]

List of properties supported by ROOST predictor.