Abstract Details
Abstracts
Author: Mark Cianciosa
Requested Type: Poster
Submitted: 2018-03-01 15:58:10
Co-authors: K.H. Law, E.H. Martin, A. Zafar, D.L. Green
Contact Info:
Oak Ridge National Laboratory
PO BOX 2008 MS6169
Oak Ridge, TN 37831
United States
Abstract Text:
Physical models are often developed such that basic physics is used to predict observable outcomes (A -> B). Often researchers desire to do the reverse, determine the underlying physics from the known outcomes. However, the inverse of the physics model (B -> A) isn't always available. Inverse methods determine these unknown quantities by searching parameter space for a set of optimal input parameters. Minimizing the differences between known and model outcomes, determines the most probable parameters given the evidence. However, searches in parameter space can be computationally costly. Machine learning methods can significantly reduce this computational cost by producing a direct reduced model of the inverse representation (B -> A). Using a physical model, a training set can be produced by sampling a large range of parameter space offline allowing rapid, even real time inversion online. This presentation will show the results of training neural nets and general linear models to produce an inverse model for spectral diagnostic measurements. Using uncertainty quantification methods, the inherent uncertainty in these reduced model is demonstrated.
Comments: