sup3r.utilities.loss_metrics.GeothermalPositiveTemperatureGradientLoss

class GeothermalPositiveTemperatureGradientLoss(depths=range(0, 8000, 1000), temperature_prefix='t')

Bases: Sup3rLoss

Positive geothermal gradient loss

This loss applies the positive-gradient regularization described in [1]. It penalizes negative vertical temperature gradients so predicted temperature increases with depth. Temperature features are expected in C and named <temperature_prefix>_<depth>m (e.g. “t_2000m”).

References

[1]

Aljubran, M. J., and Horne, R. N., “Thermal Earth model for the conterminous United States using an interpolative physics-informed graph neural network,” Geothermal Energy, vol. 12, no. 1, article 25, 2024. doi:10.1186/s40517-024-00304-7.

Initialize the positive temperature-gradient loss

Parameters:

• depths (iterable of int, optional) – Depth levels in m used to assemble temperature channels. At least two depths are required.

• temperature_prefix (str, optional) – Prefix used for temperature channels in C. Expected feature names are <temperature_prefix>_<depth>m.

Methods

call(__, x_gen)	Evaluate the positive temperature-gradient loss
from_config(config)
get_config()

Attributes

LOSS_METRIC
dtype

call(__, x_gen)

Evaluate the positive temperature-gradient loss

Parameters:

• x_true (tf.tensor) – Ground truth data (unused).

• x_gen (tf.tensor) – Synthetic generator output used to compute vertical temperature gradients. The feature axis must contain temperature channels in C. Shape must be either: (n_observations, spatial_1, spatial_2, features) or (n_observations, spatial_1, spatial_2, temporal, features)

Returns:

tf.tensor – 0D tensor loss value

__call__(y_true, y_pred, sample_weight=None)

Call self as a function.