Multimodel ranking and inference in ground water modeling.

Area of Science:

• Hydrogeology
• Environmental Science
• Statistical Modeling

Background:

• Hydrogeologic uncertainty necessitates evaluating multiple plausible models for groundwater system characterization.
• Existing model selection criteria often favor parsimony, potentially leading to underfitting and an oversimplified representation of reality.
• Multiple models typically provide acceptable fits to observational data, underscoring the need for multimodel inference.

Purpose of the Study:

• To advocate for the use of Kullback-Leibler (K-L) information as a rigorous and practical approach for multimodel inference in hydrogeology.
• To highlight the statistical underpinnings that make K-L information preferable to other model selection criteria.
• To demonstrate the application of K-L information for groundwater model evaluation through a computer-generated example.

Main Methods:

• Utilizing Kullback-Leibler (K-L) information for quantitative model comparison and selection.
• Employing a computer-generated example to illustrate the practical implementation of K-L based multimodel inference.
• Analyzing statistical foundations to contrast K-L information with traditional model selection approaches.

Main Results:

• Kullback-Leibler (K-L) information provides a computationally simple and interpretable method for model inference.
• K-L information inherently favors parsimonious models while offering a more realistic measure of predictive precision.
• The K-L approach treats models as approximations, aligning with the expectation that more data reveals greater system detail.

Conclusions:

• Multimodel inference using Kullback-Leibler (K-L) information is a statistically sound and practical approach for hydrogeologic modeling.
• K-L information offers advantages over traditional criteria by providing a more realistic assessment of uncertainty and model fit.
• The proposed method enhances the evaluation of groundwater systems by embracing model uncertainty and iterative refinement with new data.