April 23-25

Log in

Abstract Details

files Add files


Author: Eric C Howell
Requested Type: Poster
Submitted: 2018-02-28 11:54:31

Co-authors: J.D. Hanson

Contact Info:
Tech-X Corporation
5621 Arapahoe Ave, Ste A
Boulder, Co   80303
United States

Abstract Text:
A non-parametric Gaussian process regression (GPR) model is developed in the 3-D equilibrium reconstruction code V3FIT. Gaussian processes generalize multivariate Gaussian distributions to a continuos domain. The properties of a Gaussian process are determined by it’s covariance kernel. For example Gaussian processes defined by squared exponential kernel are normal distributions over all smooth functions. Gaussian process regression assumes that the unknown profile belongs to a particular Gaussian process, and Bayesian analysis is used to determine the give the best fit to measured data.
The implementation in V3FIT uses a hybrid representation where Gaussian processes regression is used to infer select equilibrium profiles and standard parametric techniques are used to infer the remaining profiles. A quasi-Newton iteration is used to simultaneously converge on the optimal set of model parameters and kernel hyper-parameters.
This implementation is tested using both synthetic and experimental data. The synthetic case uses GPR to infer the temperature profile from Thomson measurements. This case is also used to illustrate the behavior of covered profile on the Gaussian process kernel hyper-parameters. Two experimental reconstructions are also presented. A CTH reconstruction is presented where Gaussian processes are used to infer the Soft X-ray emissivity profiles using a two-color diagnostic. A single helicity MST reconstruction is also presented where the Gaussian process is used to infer the temperature profile from Thomson data. The experimental hybrid reconstructions are benchmarked against fully parametric reconstructions, and good agreement between the two methods is observed.

This work was supported by Auburn University and the U.S. DOE through Award Numbers DE-FG02-03ER54692