Speaker
Sam Vinko
Description
See the full Abstract at http://ocs.ciemat.es/EPS2018ABS/pdf/I4.212.pdf
Addressing the Inverse Problem Instability in Plasma Physics Modelling
using Stochastic Machine Learning Optimization
Sam M. Vinko, M.F. Kasim, T. Galligan, J. Topp-Mugglestone, G. Gregori
Department of Physics, Clarendon Laboratory, University of Oxford, Oxford, UK
Our understanding of the behaviour of matter in extreme conditions has greatly benefited
from the advent of novel laser and free-electron laser facilities, and the growing availability of
high-performance supercomputing. Large-scale plasma experiments are now commonly mod-
elled via increasingly detailed simulations, where the agreement between experiment and sim-
ulation enables the extraction of physical quantities and the understanding of novel underlying
processes. However, simulations with large parameter spaces suffer from the inverse problem
instability, where very similar simulated outputs can map back to very different sets of input pa-
rameters. While this provides a fundamental problem for interpreting the results from integrated
experiments, the effect is seldom comprehensively explored due to the intractably large num-
ber of simulations required to fill the parameter space. Here we show how this problem can be
addressed using stochastic machine learning optimization together with Markov Chain Monte
Carlo techniques. We apply our approach to extract physical information from three common
experimental diagnostics: x-ray emission spectroscopy, inelastic x-ray scattering and proton ra-
diography. We find that all three suffer from inverse instabilities, rendering the extraction of
physical information from some experimental measurements impossible even when excellent
agreement with a simulation can be found. Our method provides a way to quantify the uncer-
tainty due to the unstable nature of reverse physical models, and we describe an approach to
experimental design that can mitigate its impact.
