Abstract:
|
Uncertainty quantification (UQ) and Bayesian calibration help obtain reliable predictions in a multi-scale model. These models often must be calibrated to experimental bench-scale data. The calibration framework usually requires a stochastic emulator trained to limited output from a computationally expensive model, and a discrepancy term between the model and reality. Here we present several different advancements in statistical methodology in UQ and calibration. These include calibration of very large models to bench-scale data, accounting for extrapolation uncertainty due to functional inputs, an intrusive dynamic discrepancy approach for upscaling of uncertainty, a large number of potential parameters, and a an input space of potential functions. We used flexible and computationally approaches for the emulator and/or discrepancy, including a Bayesian Smoothing Spline (BSS) ANOVA Gaussian Process. The methodology presented may have far-reaching impact in many areas of science where multiscale modeling is used. This work will include applications to carbon capture systems and plasma physics.
|
ASA Meetings Department
732 North Washington Street, Alexandria, VA 22314
(703) 684-1221 • meetings@amstat.org
Copyright © American Statistical Association.