- 22 views
- 23 downloads
Machine Learning: A novel approach to infer plasma parameters from probe measurements
-
- Author / Creator
- Olowookere, Akinola
-
The research reported in this thesis is multidisciplinary in nature. It presents the use
of kinetic simulations, state of the art data analysis, and machine learning techniques
to infer plasma parameters and satellite parameter from Langmuir probe measurement.
Physical parameters such as plasma state variables, are generally inferred from
probe measurements using analytic or empirical formulas derived by applying different
approximations to physical theories. The objective of this work is to develop more
accurate techniques to infer physical parameters from probe measurements under
more realistic plasma conditions with quantifiable uncertainties and with a potential
for incremental improvements by adding more complex physical processes. Threedimensional
particle-in-cell simulations are used to calculate the current collected
by fixed-bias spherical Langmuir probes relative to the satellite under conditions of
increasing realism, starting with an isolated probe attached to a guard, to probes
attached to a satellite. The advantage of fixed bias probes considered in cases studied
throughout this thesis, is their higher temporal and spatial resolutions, compared
to the more standard mode operation where bias voltages are swept in time. The
calculated currents and the plasma and satellite parameters assumed in the simulations
are then used to build a solution library or synthetic data set, to construct
regression models to infer parameters of interest such as the satellite potential, the
density, the plasma flow velocity, and the ratio between the plasma density and the
square root of the temperature. The solution library is randomly split into two disjoint
sets; one to train models, and the other to validate inferences, and assess the
skill of the models and quantify their uncertainties. Different approaches are used to train models, including radial basis function (RBF) regressions, deep neural networks,
and combinations of these methods with analytic estimates using the boosting
ensemble learning technique, and affine transformations. In the combined approach,
the parameters of interest are first estimated using analytic expressions, followed by
regressions to reduce inference errors. Simple affine transformations are also applied
to improve the accuracy of analytic inferences, when the Pearson correlation coefficient
R with known values is high. In each case, models’ inference skills are assessed
using different metrics to quantify discrepancies compared with known values in synthetic
data sets. The models show excellent performance with the maximum relative
error ranging from 7% to 12% for the density and the ratio of density and the square
root of temperature, and a maximum absolute error in the range of 0.2 V to 0.4 V for
the floating potential in all the cases considered. The models are applied to in situ
data, and the inferences are compared to in-situ data from satellites, with which they
show excellent qualitative agreement. Finally, the agreements between the models
inferences and the synthetic data values indicate that the approaches used in this
thesis are promising with the advantage of producing uncertainty margins that are
specifically related to the inference techniques used. -
- Subjects / Keywords
-
- Graduation date
- Fall 2022
-
- Type of Item
- Thesis
-
- Degree
- Doctor of Philosophy
-
- License
- This thesis is made available by the University of Alberta Library with permission of the copyright owner solely for non-commercial purposes. This thesis, or any portion thereof, may not otherwise be copied or reproduced without the written consent of the copyright owner, except to the extent permitted by Canadian copyright law.