Skip to main content

emulators.gpr

emulators.gpr

Gaussian Process Regression (GPR) emulator.

A flexible non-parametric emulator built on GPflow, using an automatic relevance determination (ARD) squared-exponential kernel. It captures nonlinear responses and gives excellent calibrated uncertainty, at higher computational cost than the linear/Bayes-linear emulators; best for small-to-medium training sets.

Classes

Name Description
GPR Gaussian Process Regression emulator implemented in GPflow.

GPR

emulators.gpr.GPR(x=None, y=None, test_fraction=0.25)

Gaussian Process Regression emulator implemented in GPflow.

Uses an ARD squared-exponential kernel to model nonlinear parameter-output relationships with well-calibrated predictive uncertainty. More expensive than the linear/Bayes-linear emulators, so it is best suited to small-to-medium training sets where prediction quality matters most.

Example

import numpy as np, pandas as pd
from historymatching.emulators.gpr import GPR
x = pd.DataFrame({'beta': np.random.rand(40), 'gamma': np.random.rand(40)})
y = pd.DataFrame({'peak': np.sin(3 * x['beta']) + x['gamma'] ** 2})
em = GPR(x, y)
em.train()
pred = em.predict(x)

Methods

Name Description
get_hyperparameters Return GPR hyperparameters as a JSON-serializable dict.
predict Predict an output using the trained emulator.
print_emulator_description Display detailed specifications (for example, emulator coefficients)
train Fits a GPR model.
get_hyperparameters
emulators.gpr.GPR.get_hyperparameters()

Return GPR hyperparameters as a JSON-serializable dict.

predict
emulators.gpr.GPR.predict(x)

Predict an output using the trained emulator.

print_emulator_description
emulators.gpr.GPR.print_emulator_description()

Display detailed specifications (for example, emulator coefficients) for the trained emulator.

train
emulators.gpr.GPR.train()

Fits a GPR model.

Inputs are normalized to [0, 1] and outputs are standardized (zero mean, unit variance) before training. This ensures: - Kernel lengthscales are comparable across parameters with very different physical scales (e.g. 0.0005–0.006 vs 1.0–3.0). - The optimizer isn’t confused by large output offsets (e.g. birth weight ~ 3000 g with a signal of a few grams).

Predictions are un-standardized automatically in predict().