Short Simulations for Efficient Model Evaluation, Tuning and Calibration Part I: Strategy and Framework

Wednesday, May 6, 2015 - 07:00
Add to Calendar

This 'Short Simulations' task explores the feasibility and usefulness of short (2-10 days) simulations for the purpose of efficient and effective testing, tuning and evaluation of high-resolution models. Our investigation consists of two phases: parametric sensitivity analysis and automatic model tuning, both of which involve an extensive use of UQ techniques. The parametric sensitivity analysis provides a testbed to explore the strengths and limitations of the short simulations and to fine-tune the experimental design. Later it will also be used to down-select uncertain parameters thus further speed up the auto-tuning. In the past quarters we have identified a test problem, and established a complete framework that efficiently conducts simulations and analyzes results for the parametric sensitivity study. Details of the test problem and framework are presented in this poster. Early results from the sensitivity analysis are shown in Part II. Preliminary but promising results indicate that short simulations are effective for the detection of certain model sensitivities. This testing strategy might be useful for other tasks, e.g., quantification of the ACME model's resolution sensitivity and evaluation of candidate convection schemes. The experience obtained from short simulations is also helpful in improving our understanding of the model behavior at the process level.