The importance of uncertainty quantification in model reproducibility

Area of Science:

• Computational science
• Statistical modeling
• Climate science

Background:

• Computer models often have high-dimensional inputs and long computation times.
• Deterministic models still face significant uncertainties, impacting validation and reproducibility.
• Effective uncertainty quantification is essential for reliable scientific computing.

Purpose of the Study:

• To present a statistical framework, history matching, for global parameter search in computer models.
• To integrate uncertainty quantification into the model validation and reproducibility process.
• To utilize Gaussian process emulators for efficient model prediction and uncertainty analysis.

Main Methods:

• Employing history matching for global parameter search by comparing model output to observed data.
• Utilizing Gaussian process (GP) emulators for rapid predictions and calculation of output uncertainty distributions.
• Identifying parameter sets that yield acceptable agreement between model outputs and observed data within uncertainty bounds.

Main Results:

• History matching effectively identifies parameter regions consistent with observational data.
• Gaussian process emulators provide fast, uncertainty-aware predictions of model behavior.
• The framework was demonstrated on a toy model and a climate model, showing its applicability.

Conclusions:

• History matching provides a robust statistical framework for computer model validation and reproducibility.
• The integration of GP emulators significantly speeds up the uncertainty quantification process.
• Variability in model predictions derived from identified parameter settings is crucial for model development and inter-model comparisons.