geopfa.extrapolation.build_combined_kernel

build_combined_kernel(X_stdized, Y_stdized, use_matern=True, use_rbf=True, bias=True, white=True, longscale=True, lower_frac=0.02, upper_frac=0.2)[source]

Construct a composite kernel consisting of optional components: RBF, Matern 3/2, long-scale RBF, Bias, and White noise kernels.

Parameters:

  • X_stdized (numpy.ndarray) – Standardized training coordinates.

  • Y_stdized (numpy.ndarray) – Standardized training targets.

  • use_matern (bool, optional) – Whether to include a Matern 3/2 component.

  • use_rbf (bool, optional) – Whether to include a short/medium-scale RBF component.

  • bias (bool, optional) – Whether to include a constant Bias kernel.

  • white (bool, optional) – Whether to include a White noise kernel.

  • longscale (bool, optional) – Whether to include a very smooth long-scale RBF background kernel.

  • lower_frac (float, optional) – Lower bound fraction (of global radius) for ARD lengthscales.

  • upper_frac (float, optional) – Upper bound fraction (of global radius) for ARD lengthscales.

Returns:

  • kernel (GPy.kern.Kern) – Combined kernel created by summing enabled components.

  • info (dict) – Diagnostic information for each component, including learned bounds and initialization values.

Notes

  • Behavior is identical to the original implementation.

  • The long-scale RBF kernel is fixed to very smooth lengthscales and tiny variance, acting as a gentle background component.