Bayesian modelling of nuclear fusion experiments

Bayesian probability theory as a general framework for scientific modelling and inference is introduced and applied to nuclear fusion experiments in order to provide consistent inference solutions given multiple heterogeneous data sets. Fusion plasmas are complex physical systems, in which charged particles are confined by the electromagnetic force. The physics parameters of the plasmas involve various independent measurements from sophisticated scientific instruments. Owing to the complexity of the experiments and the fusion plasmas, so far, no physics model can predict major physical phenomena, like transport, sufficiently well. Hence, generic, non-parametric Gaussian processes are used to model physics parameters such as plasma current density and pressure. Multiple predictive models of scientific instruments have been developed individually, and they have been combined into a joint model with Gaussian process priors in order to perform robust and consistent inference. The joint model provides the joint posterior probability distribution of the physics parameters, hyperparameters and other unknown parameters, such as calibration factors. This thesis theoretically and experimentally shows that the joint posterior distribution intrinsically embodies Bayesian Occam's razor. Therefore, by exploring the joint posterior distribution, inference solutions can be found with optimal values of all the model parameters, based on the principle of Occam's razor. In other words, we can apply Bayesian Occam's razor to real-world problems without calculation of the model evidence, typically requiring marginalisation over a high-dimensional parameter space, which is one of the major obstacles to Bayesian model selection. Based on this foundation, several applications have been developed for consistent inference of the physics parameters of the fusion plasmas at two major fusion experiments, the JET and W7-X. The first application has been developed by modelling emission spectra and relevant atomic physics for the lithium beam emission spectroscopy system at JET to provide the edge plasma electron density profiles and their posterior uncertainties. Additionally, interferometers, Thomson scattering and spectroscopy systems have been combined, improving the consistency of the inference solutions. These joint inference applications have been developed for JET and W7-X to provide the marginal posterior distribution of the plasma density and temperature profiles. Furthermore, the full joint posterior distribution of axisymmetric plasma equilibria, given magnetic field and plasma pressure measurements, has been explored for the first time at JET. These equilibrium solutions suggest two different possible plasma equilibrium current distributions for high-confinement mode fusion plasmas: either a strong toroidal current density or a poloidal current flux hole in the edge region. The principles and methods developed in this thesis are general and applicable to all kinds of scientific problems. This new approach to model selection by exploring the joint posterior distribution contributes to the general automatisation of scientific discovery.

Author dc:creator

• Kwak, Sehyun

Advisor dc:contributor.advisor

• Svensson, Jakob

Licence dc:rights.uri

• http://rightsstatements.org/vocab/InC/1.0/

Language dc:language.iso

en

Identifier URI

http://dx.doi.org/10.14279/depositonce-10806

OAI identifier oai:identifier

oai:depositonce.tu-berlin.de:11303/11915