CSDMS 2013 annual meeting poster Mary Hill
Randall Hanson, US Geological Survey San Diego California, United States. email@example.com
Leon Kauffman, US Geological Survey West Trenton New Jersey, United States. firstname.lastname@example.org
Martyn Clark, National Center for Atmospheric Research Boulder Colorado, United States. email@example.com
Dmitri Kavetski, University of Adelaide Adelaide , Australia. firstname.lastname@example.org
Ming Ye, Florida State University Tallahassee Florida, United States. email@example.com
When we build models we create worlds that we hope will inform us about the world in which we live. We hope models will help us understand processes, causes and effects; avoid difficulties; benefit human endeavors; and accommodate and nurture the ecology which has its own beauty and importance, and upon which human existence and our economy depend. Here we discuss how models can be used to achieve these goals by considering the importance of transparency (revealed importance) and refutability (tested hypotheses). We consider models with substantial execution times (for our example one model run requires 20 minutes) and transparency and refutability available using computationally frugal methods. Challenges of using these methods include model nonlinearity; non-Gaussian errors and uncertainties in observations, parameters, and predictions; and integrating information from multiple data types and expert judgment. A synthetic test case illustrates the importance of transparency and refutability in model development. The test case represents transport of an environmental tracer (cfc) and contaminant (pce) in a groundwater system with large-scale heterogeneities. Transparency is served by identifying important and unimportant parameters and observations. The frugal methods identified consistently important and unimportant parameters for three sets parameters for which sum of squared weighted residuals (SOSWR; dimensionless; constructed with error-based weighting) varies between 5606 and 92. Observations important to the parameter values are largely consistent, but the order varies for results using different parameter values because of model nonlinearity. For each set of parameters these results required 17 model runs. Refutability is served by estimating parameter values that minimize SOSWR and evaluating resulting model fit and parameter values. The computationally frugal parameter-estimation method reduced SOSWR from 5606 to 92, displayed no evidence of local minima, and required about 100 model runs each of the 10 times it was executed. The similar important parameters and observations for different parameter sets and performance of parameter estimation suggest the utility of the computationally frugal methods even for models as nonlinear as the one considered here. The value of the kinds of insights gained in this work is highlighted by the 10,000s to 1,000,000s of model runs being conducted in many studies to obtain them.
|Each bar is stacked to indicate the contribution|
of five types of observations. CSS required 17 parallelizable model runs.
|* Please acknowledge the original contributors when you are using this material. If there are any copyright issues, please let us know and we will respond as soon as possible.|