Presentation Title: Scalable Surrogate Detection in Large-Scale Simulations

Committee:

• Gabrielle Allen
• Jian Zhang
• Q. Jim Chen

Date: Apr 8, 2011
Time: 10:00 AM
Location:256 Coates

Abstract:
Simulation has become a useful approach in scientific computing and engineering for its ability to model real natural or human systems. In particular, for complex systems such as hurricanes, wildfire disasters, and real-time road traffic, simulation methods are able to provide researchers, engineers and decision makers predicted values in order to help them to take appropriate actions. For large-scale problems, the simulations usually take a lot of time on supercomputers, thus making real-time predictions more difficult. Approximation models that mimic the behavior of simulation models but are computationally cheaper, namely "surrogate models", are desired in such scenarios. In the thesis, a framework for scalable surrogate detection in large-scale simulations is presented with the basic idea of "using functions to represent functions". The following issues are discussed in the thesis: i) the data mining approaches to detecting and optimizing the surrogate models; ii) a scalable and automated workflow of constructing surrogate models from large-scale simulations; and iii) the system design and implementation with the application of storm surge simulations in the occurrence of hurricanes.


All are invited.