pod.build_fixed_kernel_gpr

build_fixed_kernel_gpr(kernel, x_mean=None, x_std=None)

Builds a Kriging model that reuses already-optimised hyperparameters.

Used wherever a fitted Kriging surrogate is refitted to resampled data (PoD bootstrap confidence bounds, the GUI bootstrap convergence trace). The optimiser is disabled, so only the posterior is recomputed, and the settings match the production model: normalize_y=True and only numerical jitter on the diagonal, because the noise variance is already part of kernel (its WhiteKernel term).

[M25] Sec. 2i re-estimates the selected model’s parameters on every bootstrap resample. [digiqual] freezes the Kriging hyperparameters instead, because re-optimising them 1000 times is too slow. The bootstrap interval therefore omits hyperparameter uncertainty.

The input scaling should be frozen too, so that the length scales keep the meaning they had in the full-data fit. Pass the full-data x_mean and x_std, or pass the model’s model_params_ dictionary as kernel.

Parameters

Name	Type	Description	Default
kernel	Kernel | dict	A fitted kernel (gpr.kernel_), or the model_params_ dict of a Kriging model from fit_all_robust_mean_models (keys kernel, x_mean, x_std).	required
x_mean	np.ndarray	Full-data input means.	None
x_std	np.ndarray	Full-data input standard deviations.	None

Returns

Name	Type	Description
ScaledGaussianProcessRegressor	ScaledGaussianProcessRegressor	An unfitted regressor ready for .fit().