[[CSDMS meeting abstract presentation::Numerical models are critical to integrating knowledge and data for environmental systems and understanding future consequences of management decisions, weather variability, climate change, and so on. To attain the transparency and refutability needed to understand predictions and uncertainty and use models wisely, this clinic presents a strategy that emphasizes fundamental questions about model adequacy, sensitivity analysis, and uncertainty evaluation, and consistent use of carefully designed metrics. Emphasizing fundamental questions reveals practical similarities in methods with widely varying theoretical foundations and computational demands. In a field where models take seconds to months for one forward run, a credible strategy must include frugal methods for those in Kansas who can only afford 10s to 100s of highly parallelizable model runs in addition to demanding methods for those in Oz who can afford to do 10,000s to 1,000,000s of model runs. Advanced computing power notwithstanding, people may be in Kansas because they have chosen complex, high-dimensional models, want quick insight into individual models, and/or need systematic comparison of many alternative models. This class will briefly review the fundamental questions, demonstrate relations between existing theoretical approaches, and address challenges and limitations. Students will be able to examine a model constructed using FUSE and compare results from computationally frugal method evaluations conducted in class and demanding methods for which results are provided.Notice:
During the clinic you will have the opportunity to run an exercise on your laptop. The exercise uses R, which is freely downloadable. The clinic is only an hour, so it will really be necessary to have downloaded and installed R prior to arriving. Do this as follows
go to http://cran.cnr.berkeley.edu/
Install version 2.15.3
Linux, Mac, or Windows versions are available.
You can install with or without administrative privileges.
The Sobol’ results take 6,000,000 model runs and about 12 hours, so can not be run in class. They are provided in the file:
Each line presents average results for a bootstrapped Sobol’ sample for a portion of the full parameter space. The averages for the entire range of parameters is on the line with grid index=101
The R script using this file to create plots; it does not do the runs. ]]