VARIABLE SCREENING METHODS IN META-MODEL DESIGN FOR A LARGE STRUCTURAL DYNAMICS SIMULATION
- Amanda L.
- John F.
- François M.
- Scott W.
This research addresses the problem of analyzing the nonlinear transient response of a large degree-of-freedom structural dynamics simulation. A threaded joint assembly's response to impulse loading has been studied. Twelve parameters relating to the input level, preloads of the joint and friction between components are thought to influence the acceleration response of the structure. Due to the high cost of physical testing and large amount of computation time to run analytical models (three hours on 504 processors of LANL's ASCI supercomputer, Blue Mountain to simulate three milliseconds of response) a fast-running metamodel is being developed. In this case, a metamodel is a statistically developed surrogate to the physics-based finite element model and can be evaluated in minutes on a single processor desktop computer. An unreasonable number of runs is required (312>500,000) to generate a three level full factorial design with 12 parameters for metamodel creation. Some manner of down-selecting or variable screening is needed in order to determine which of the parameters most affect the response and should be retained in subsequent models. A significant effects methodology has been used, involving a design of experiments technique, in which all parameters were lirst included in the model and then eliminated on the basis of statistical contributions associated with each parameter. Bayesian variable screening techniques, in which probabilities of effects are generated and updated, have also been explored. Encouraging results have been obtained, as the two methods yield similar sets of statistically significant parameters. Both methods have been compared to general sensitivity analysis (GSA). The resulting compact metarnodel can then be explored at more levels to appropriately capture the underlying physics of the threaded assembly with a much smaller set of simulations (only 64 runs required for a three level model with quadratic and some third order terms).
- Research Organization:
- Los Alamos National Laboratory (LANL), Los Alamos, NM (United States)
- Sponsoring Organization:
- USDOE
- OSTI ID:
- 975784
- Report Number(s):
- LA-UR-01-5619; TRN: US201018%%871
- Resource Relation:
- Conference: "Submitted to: The 20th International Model Analysis Conference, Los Angeles, CA Feb. 4-7, 2002"
- Country of Publication:
- United States
- Language:
- English
Similar Records
Simplified predictive models for CO2 sequestration performance assessment
Massively parallel simulation with DOE's ASCI supercomputers : an overview of the Los Alamos Crestone project