Verification-Based Model Tuning
Abstract
We put forth the idea of developing a statistical model representing the relationship between model parameters and macroscopic features in meteorological quantities, with the latter quantified in terms of measures commonly used in spatial verification methods. For example, if/when a spatial verification method suggests that frontal speed was forecast incorrectly, we would like to be able to set the model parameters (e.g., diffusion rate) to remedy that problem in some optimal sense. This primary goal requires the identification of relevant model parameters, which in turn requires performing sensitivity analysis (our secondary goal). To validate the results of the variance-based sensitivity analysis, comparisons are made with a more traditional method based on adjoints. After the important model parameters are identified, an emulator will be developed. An emulator consists of a statistical model which represents the relationship between model parameters and macroscopic features (e.g., spatial structure) of forecast fields. Such an emulator (and its inverse) has both scientific value and practical utility. An example of the former is the knowledge gained from identifying the statistical relations between model parameters and forecast parameters, and model-tuning is an example of the practical utility.
Document Details
- Document Type
- Technical Report
- Publication Date
- Sep 30, 2012
- Accession Number
- ADA574125
Entities
People
- Caren Marzban
- David W. Jones
- Scott A. Sandgathe
Organizations
- University of Washington