Advanced Analysis and Visualization of Space Weather Phenomena PDF

University of Colorado, Boulder
CU Scholar
Computer Science Graduate Theses & Dissertations Computer Science
Spring 1-1-2017
Advanced Analysis and Visualization of Space

Weather Phenomena
Joshua James Murphy
University of Colorado at Boulder, jomu9721@colorado.edu
Follow this and additional works at: https://scholar.colorado.edu/csci_gradetds

Part of the Computer Sciences Commons, and the Physics Commons
Recommended Citation
Murphy, Joshua James, "Advanced Analysis and Visualization of Space Weather Phenomena" (2017). Computer Science Graduate Theses
& Dissertations. 140.
https://scholar.colorado.edu/csci_gradetds/140
This Dissertation is brought to you for free and open access by Computer Science at CU Scholar. It has been accepted for inclusion in Computer
Science Graduate Theses & Dissertations by an authorized administrator of CU Scholar. For more information, please contact
cuscholaradmin@colorado.edu.
Advanced Analysis and Visualization of Space Weather
Phenomena
by
Joshua J. Murphy
B.S., University of Maine, 2008
M.S., University of Colorado Boulder, 2013
A thesis submitted to the
Faculty of the Graduate School of the
University of Colorado in partial fulfillment
of the requirements for the degree of
Doctor of Philosophy
Department of Computer Science
2017
This thesis entitled:
Advanced Analysis and Visualization of Space Weather Phenomena
written by Joshua J. Murphy
has been approved for the Department of Computer Science
Prof. Xiao-Chuan Cai
Dr. Scot Elkington
Date
The final copy of this thesis has been examined by the signatories, and we find that both the
content and the form meet acceptable presentation standards of scholarly work in the above
mentioned discipline.
iii
Murphy, Joshua J. (Ph.D., Computer Science)
Advanced Analysis and Visualization of Space Weather Phenomena
Thesis directed by Prof. Xiao-Chuan Cai
As the world becomes more technologically reliant, the more susceptible society as a whole is
to adverse interactions with the sun. This “space weather” can produce significant effects on modern
technology, from interrupting satellite service, to causing serious damage to Earth-side power grids.
These concerns have, over the past several years, prompted an out-welling of research in an attempt
to understand the processes governing, and to provide a means of forecasting, space weather events.
The research presented in this thesis couples to current work aimed at understanding Coronal Mass
Ejections (CMEs) and their influence on the evolution of Earth’s magnetic field and associated Van
Allen radiation belts. To aid in the analysis of how these solar wind transients affect Earth’s mag-
netic field, a system named Geospace/Heliosphere Observation & Simulation Tool-kit (GHOSTkit),
along with its python analysis tools, GHOSTpy, has been devised to calculate the adiabatic invari-
ants of trapped particle motion within Earth’s magnetic field. These invariants aid scientists in
ordering observations of the radiation belts, providing a more natural presentation of data, but can
be computationally expensive to calculate. The GHOSTpy system, in the phase presented here,
is aimed at providing invariant calculations based on LFM magnetic field simulation data. This
research first examines an ideal dipole application to gain understanding on system performance.
Following this, the challenges of applying the algorithms to gridded LFM MHD data is examined.
Performance profiles are then presented, followed by a real-world application of the system.
Dedication
To my wife Amanda, and sons Jarod and Kaladin. Thank you for being so patient.
v
Acknowledgements
First and foremost, I would like to acknowledge my family. My wife, Amanda, and my sons
Jarod and Kaladin. They have been an incredible support while working on this dissertation. I
would also like to acknowledge my advisors: Dr. Scot Elkington, and Dr. Xiao-Chuan Cai from
LASP and Computer Science respectively. They were both tremendous with their support.
The DoD Science, Mathematics, and Research for Transformation (SMART) scholarship for
service program made this thesis possible through funding and job prospects. I would like to
specifically thank Keisha Williams and Phillippe Reed at the US Navy Space and Naval Warfare
Systems Center (SPAWAR) Atlantic in Charleston, South Carolina for allowing me extra time
during internships to work on this dissertation. That extra time was invaluable; thank you.
Contents
Chapter
1 Space Weather 1
1.1 A Brief History of Space Weather . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1
1.2 Connecting the Sun and the Earth . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3
1.3 Radiation Belts . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5
1.3.1 Trapped Particle Motion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5
1.3.2 The Adiabatic Invariants . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7
1.4 Analysis and Visualization of Particles in the Magnetosphere . . . . . . . . . . . . . 10
1.5 Numerical Models of the Earth’s Magnetosphere . . . . . . . . . . . . . . . . . . . . 10
1.5.1 Dipole Geomagnetic Field Model . . . . . . . . . . . . . . . . . . . . . . . . . 11
1.5.2 Empirical Geomagnetic Field Models . . . . . . . . . . . . . . . . . . . . . . . 12
1.5.3 Magnetohydrodynamic (MHD) Geomagnetic Field Models . . . . . . . . . . . 14
1.6 Thesis Format . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 16
2 Observational Uncertainties and CME Transients 17
2.1 The Model Environment . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 20
2.1.1 The WSA Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 20
2.1.2 The Cone Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 21
2.1.3 The Enlil Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 23
2.2 Mitigation of Uncertainties with Ensemble Modeling . . . . . . . . . . . . . . . . . . 24

vii
2.3 Analyzing Effects of Uncertainty with CME Impact Domain Analysis . . . . . . . . 27
2.3.1 The 06 October 2013 CME Event . . . . . . . . . . . . . . . . . . . . . . . . 28
2.3.2 Defining an Impact . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 28
2.3.3 Finding and Populating the Impact Domain . . . . . . . . . . . . . . . . . . . 32
2.3.4 Solar Wind Impact Domains and Quartiles . . . . . . . . . . . . . . . . . . . 35
2.4 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 43
3 Adiabatic Invariant Methods and Error Analysis in a Dipole 45
3.1 Invariant Algorithms . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 45
3.1.1 Computing µ . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 46
3.1.2 Computing K . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 46
3.1.3 Computing L∗ . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 51
3.2 Error Response in a Dipole Field . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 55
3.2.1 Error Response due to Field Line Footprint Inaccuracies . . . . . . . . . . . . 58
4 Working with LFM Model Gridded Output 63
4.1 The LFM Grid . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 63
4.2 LFM Grid Dipole Response . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 64
4.3 Algorithm Adjustments for LFM Data . . . . . . . . . . . . . . . . . . . . . . . . . . 71
4.3.1 The New K/B Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 73
4.3.2 Numerical Noise . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 76
4.3.3 Extremely Distorted Magnetic Fields and Search Progression . . . . . . . . . 80
4.3.4 Dealing with Multiple K Instances Per Line . . . . . . . . . . . . . . . . . . . 84
4.3.5 Final Updated Drift Trajectory Search Algorithm . . . . . . . . . . . . . . . . 85
4.3.6 Accuracy and Convergence . . . . . . . . . . . . . . . . . . . . . . . . . . . . 86
4.4 Application of Methods . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 88
4.4.1 Van Allen Probes L∗ and K Comparisons . . . . . . . . . . . . . . . . . . . . 92
4.4.2 LFM vs. LFM/RCM Full Field Comparisons . . . . . . . . . . . . . . . . . . 98

viii
5 Parallel Operations and Timings 108
5.1 Serial Single Point Algorithm Timings . . . . . . . . . . . . . . . . . . . . . . . . . . 108
5.1.1 Single Point Computation - L∗ Overview . . . . . . . . . . . . . . . . . . . . 110
5.1.2 Single Point Computation - KB Line Search Routine . . . . . . . . . . . . . . 112
5.1.3 Single Point Computation - K/B Model . . . . . . . . . . . . . . . . . . . . . 114
5.2 Parallel Operations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 115
5.2.1 Application Configuration . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 115
5.2.2 Parallel Scaling Efficiency . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 115
5.2.3 Parallel Timing Profile . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 117
5.2.4 Load Balancing . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 119
6 Closing Remarks 123
Bibliography 125
Tables
Table
2.1 Enlil Variables . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 25
2.2 Impact Domain Plot Values . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 37
4.1 LFM grid resolutions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 66

Figures
Figure
1.1 The Sun Earth Connection . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3
1.2 Diagram of Magnetosphere . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4
2.1 Overview of Enlil Simulation Space . . . . . . . . . . . . . . . . . . . . . . . . . . . . 17
2.2 Overview of the Cone Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 22
2.3 WSA/Enlil Simulation of 06 October 2013 CME . . . . . . . . . . . . . . . . . . . . 29
2.4 Illustration of Impact Defintion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 30
2.5 Impact Domain Plots for max(DP ) and max( dDP

dt ) . . . . . . . . . . . . . . . . . . . 33
2.6 Impact Domain Plots for time to impact and CME Contact Duration . . . . . . . . . 36
2.7 Impact Domain Plots for Solar Wind Properties . . . . . . . . . . . . . . . . . . . . 38
2.8 Quartile Envelopes for Solar Wind Properties . . . . . . . . . . . . . . . . . . . . . . 42
3.1 Results of Re-sampling Field Lines . . . . . . . . . . . . . . . . . . . . . . . . . . . . 51
3.2 Analytic Dipole L∗ vs. Intersection λ . . . . . . . . . . . . . . . . . . . . . . . . . . . 59
3.3 λ Error profiles . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 60
3.4 Etrace Error profile for 6.6RE . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 61
3.5 Etrace Error profile for range of L∗ values . . . . . . . . . . . . . . . . . . . . . . . . 62
4.1 LFM Grid Side and Front View . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 64
4.2 LFM Grid 2D Details . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 65
4.3 L∗ Errors on LFM Grid . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 66

xi
4.4 Footprint error samples on LFM single grid . . . . . . . . . . . . . . . . . . . . . . . 68
4.5 Footprint Errors on LFM Single Res Grid . . . . . . . . . . . . . . . . . . . . . . . . 68
4.6 Footprint Error Samples on LFM Double Grid . . . . . . . . . . . . . . . . . . . . . 69
4.7 Footprint Errors on LFM Double Res Grid . . . . . . . . . . . . . . . . . . . . . . . 69
4.8 Footprint Error Samples on LFM Quad Grid . . . . . . . . . . . . . . . . . . . . . . 70
4.9 Footprint Errors on LFM Quad Res Grid . . . . . . . . . . . . . . . . . . . . . . . . 70
4.10 Footprint Error Distributions for 3 Grid Resolutions . . . . . . . . . . . . . . . . . . 71
4.11 Mean Error Effect Ratio - All Grids . . . . . . . . . . . . . . . . . . . . . . . . . . . 72
4.12 K Calculation Graphic . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 74
4.13 Dipole K/B Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 74
4.14 LFM K/B model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 75
4.15 K in Bifurcation Regions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 77
4.17 Numerical Instability in LFM Data (non-filtered) . . . . . . . . . . . . . . . . . . . . 79
4.18 Numerical Instability in LFM Data (filtered) . . . . . . . . . . . . . . . . . . . . . . 79
4.21 Search Failure in Highly Distorted Field . . . . . . . . . . . . . . . . . . . . . . . . . 82
4.22 Adaptive λ Search Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 83
4.23 Convergence Difference (1 × 10−3 ) . . . . . . . . . . . . . . . . . . . . . . . . . . . . 86
4.24 L∗ Convergence Difference 1 × 10−4 vs. 4-point . . . . . . . . . . . . . . . . . . . . . 89
4.25 Convergence Differences (4-Line) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 90
4.26 02 Oct 2013 Solar Wind Conditions . . . . . . . . . . . . . . . . . . . . . . . . . . . 91
4.27 Dst for 02 Oct 2013 Storm . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 92
4.28 Van Allen Probe A & B T89 L∗ andK Plots . . . . . . . . . . . . . . . . . . . . . . . 93
4.29 Van Allen Probe A & B Dipole L∗ andK Plot . . . . . . . . . . . . . . . . . . . . . . 94
4.30 Van Allen Probes LFM L∗ Plot . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 95

xii
4.31 Van Allen Probes LFM/RCM L∗ Plot . . . . . . . . . . . . . . . . . . . . . . . . . . 97
4.32 LFM/LFMRCM |B| Field Comparison . . . . . . . . . . . . . . . . . . . . . . . . . . 99
4.33 LFM/LFMRCM Comparison K equatorial . . . . . . . . . . . . . . . . . . . . . . . 101
4.34 LFM/LFMRCM Comparison K meridional . . . . . . . . . . . . . . . . . . . . . . . 102
4.35 LFM/LFMRCM Comparison - L∗ equatorial . . . . . . . . . . . . . . . . . . . . . . 104
4.36 LFM/LFMRCM Comparison - L∗ meridional . . . . . . . . . . . . . . . . . . . . . . 105
4.37 LFM/LFMRCM Comparison - L∗ contours . . . . . . . . . . . . . . . . . . . . . . . 106
4.38 LFM/LFMRCM Comparison . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 107
5.1 Point L∗ Histogram . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 109
5.2 Point Completion Time Histogram . . . . . . . . . . . . . . . . . . . . . . . . . . . . 109
5.3 L∗ vs. Point Completion Time . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 110
5.4 Graphical Computation Completion Time . . . . . . . . . . . . . . . . . . . . . . . . 111
5.5 GHOSTpy Performance Profile for Single L∗ Point . . . . . . . . . . . . . . . . . . . 112
5.6 GHOSTpy Performance Profile for KB Line Search . . . . . . . . . . . . . . . . . . . 113
5.7 GHOSTpy Performance Profile for K calculation . . . . . . . . . . . . . . . . . . . . 114
5.8 GHOSTpy Parallel Scalability . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 117
5.9 GHOSTpy Parallel Timing Profiles . . . . . . . . . . . . . . . . . . . . . . . . . . . . 118
5.10 GHOSTpy Parallel Comms Timing Profiles . . . . . . . . . . . . . . . . . . . . . . . 119
5.11 GHOSTpy Performance Histogram for Parallel Exectution Time . . . . . . . . . . . 121
5.12 GHOSTpy Performance Histogram for Parallel Exectution Time (Improved Binning) 122
Chapter 1
Space Weather
1.1 A Brief History of Space Weather
The effects of space weather have been observed for centuries. Records of auroral observation
date as far back as 193BC ; verified observations from independent observers in multiple locations
on the same night were seen as early as 1101AD , and indications of possible intense geomagnetic
storms arose by 1138AD . (Willis and Stephenson, 1999). In the 1600s, Galileo Galilei postulated
that the aurorae were caused by sunlight illuminating air as it escaped the Earth’s shadow. René
Descartes suggested that they were caused by ice crystals forming high in the atmosphere (Kivelson
and Russell, 1995). With the discovery of Earth’s magnetic field in 1700s, the aurorae were finally
linked to the magnetosphere (Hiorter, 1747).
On September 1, 1859, Richard Carrington observed, in what would prove to be one of
the most severe space weather events to this day, an intense white-light brightening of the sun
(Carrington, 1859). The “intensely bright and white light,” what we now refer to as a solar flare,
was followed 18 hours later by an intense magnetic storm. Baker et al. (2008), a report by the
National Academy of Science on severe space weather events and their societal impacts, succinctly
summarizes the news reports of the time. Campers in the Rocky Mountains reported “auroral light,
so bright that one could easily read common print.” Henry C. Perkins, a physician in Massachusetts,
reported “a perfect dome of alternate red and green streamers” over New England. The aurora were
seen as far south as Central America in the northern hemisphere, and as far north as Santiago Chile
in the southern hemisphere. Some telegraph operators reported fires started from induced currents
2
in telegraph lines, while others reported being able to disconnect batteries from their systems and
still send messages. By the mid 1860s, researchers Hermann Fritz in Zürich, and Elias Loomis at
Yale made convincing arguments for a connection between auroral occurrences and the sunspot
cycle (Schröder, 1998), but it wasn’t until the advent of radio astronomy in the 1930s that the
significance of these connections was appreciated, and not until after the start of the space age did
the concept of what we now call space weather emerge (Bleeker et al., 2001).
In the Early 1950s, Eugene Parker, following up on observations by Ludwig Biermann (Bier-
mann and Schlüter, 1951), postulated the idea of a supersonic magnetized plasma being continu-
ously ejected from the Sun, now known as the solar wind. This conclusion forever changed the way
scientists perceive space, and led to a much greater understanding of how the Sun and Earth are
connected magnetically.
In 1958, following the successful launch of the Explorer I and Explorer III spacecraft (Explorer
II was a launch failure), James Van Allen and his team discovered what are, today, called the Van-
Allen Radiation Belts. First announced on May 1, 1958 as “enormous intensities of geomagnetically
trapped corpuscular radiation” (from James Van Allen’s Preface to Chapter 5 of Bleeker et al.
(2001)), the radiation belts have become an extremely important aspect of space weather, as their
changes can impact a great deal of today’s space based technology (Baker et al., 2008).
Another significant advance in the theories governing space weather came with the discovery
of intermittently occurring events involving large volumes of sun-ejected plasma,called coronal mass
ejections (CMEs), in the 1970s. This discovery brought with it the recognition that CMEs were
the source of non-recurrent geomagnetic storms, not solar flares (Gosling, 1993).
The potential effects of CMEs on technology was dramatically demonstrated on March 13,
1989, when a large CME impacted the Earth (Baker et al., 2008). Some of the most notable
effects were on the power grid. The Quebec grid was tripped off-line, which caused the Provence
to experience a blackout for about 9 hours. The United States experienced more than 200 separate
power grid related effects. In October of 2003, the disruptive potential of space weather was again
made clear during what are now known as the “Halloween Storms.” These storms, following more
3
Figure 1.1: This image shows the various connections between the Sun and Earth as it relates to
space weather. On the left of the image, you can see an artists interpretation of the Sun’s magnetic
field image, with a CME being ejected from the corona. The grid patterns represent the intricate
connections necessary between model simulations of each region of space. On the right side of the
image is the Earth surrounded by its magnetosphere, being perturbed by the solar wind. Image
Credit: NASA
than a decade of technological advancements since the 1989 storm, generated a more varied array
of effects, ranging from near complete Global Positioning System (GPS) distribution to the loss of
entire satellites (Baker et al., 2008).
These storms had significantly more affect on the way of life on Earth than the relatively
mild irritations caused by the Carrington event in 1859. Were the Carrington storm to occur today,
the results could be catastrophic (Baker et al., 2008, 2013a). The need for the understanding of,
as well as the need to be able to predict, space weather is clear; to maintain our way of life in this
technological age, understanding how the sun affects our local space may be crucial.
1.2 Connecting the Sun and the Earth
The primary driver of space weather is the Sun; disgorging a continuous flow of supersonic
magnetic plasma known as solar wind. Typical solar wind travels at speeds ranging from 350 − 500
km/s, has a number density around 3 − 5 cm−3 , and contains an embedded weak magnetic field,
averaging about 5 nT, known as the Interplanetary Magnetic Field (IMF). Generally parallel to the
ecliptic plane, various solar conditions can cause the IMF to deviate from its parallel course. This
4
Figure 1.2: Various regions and currents of the magnetosphere. Image Credit: Kivelson and Russell
(1995)
deviation can then interact with the Earth’s magnetic field, generating the interactions associated
with space weather.
Earth’s internal magnetic field, first suggested to be roughly a dipole by Gilbert (1958),
interacts with the solar wind, giving rise to various currents within the near Earth region. These
currents generate what can be thought of as “external contributions” to the magnetic field. This
combination of magnetic field contributions generates the magnetosphere, and can be expressed as:
Bmag = Bint + Bext
where Bmag is the magnetic field of the magnetosphere, Bint is the Earth’s intrinsic magnetic field,
and Bext contains the external magnetic field contributions caused by the interaction with the solar
wind.
The different external contributions to the magnetosphere can further be broken down by
5
each individual current system, thus giving the expression:
Bext = BRC + BT C + BF AC + BIN T + BM P
where BRC is the magnetic field contribution of the ring current, BT C is the contribution of the
tail current, BF AC is the contribution of the field-aligned current, BIN T is the contribution from
the interconnection field, and BM P is the contribution from the magnetopause. This breakdown is
how several empirical magnetospheric models are computed (Tsyganenko, 2014) (This is discussed
in more detail in a later section). Figure 1.2 provides a graphical representation of the different
current systems that contribute to the magnetosphere.
The interaction of the IMF with the magnetosphere allows for a phenomenon known as recon-
nection (Dungey, 1961). Given a IMF field oriented southward, opposite that of the geomagnetic
field, at the magnetopause boundary on the day-side of Earth, the IMF field will merge with the
geomagnetic field lines. The solar wind will draw these field lines (now “open”, or connected to the
Sun) to the night side of Earth. The field lines will reconnect in the tail, and magnetic tension will
cause the field lines to draw back sunward.
1.3 Radiation Belts
1.3.1 Trapped Particle Motion
A trapped charged particle’s motion in an electromagnetic field is described by the Lorentz
Force equation (Northrop, 1963; Roederer and Zhang, 2014), and is given in Equation (1.1).
d~
p ~
= q~v × B (1.1)
dt
where p~ is the particles momentum, and ~v is the velocity in the “original frame of reference (OFR).”
The practical result of this is a gyration of the particle around its field line. The particle’s center of
gyration is referred to as the “guiding center,” and moves along the field line in relation to its initial
parallel velocity and any parallel force. If viewed from what Roederer and Zhang (2014) calls the
6
“guiding center system (GCS)” frame of reference, which is a frame that travels along the guiding
center with a velocity:
~vk = const, and ~v⊥ = 0
In this frame of reference, v⊥ is constant, and the particle’s gyroradius can be obtained with:
∗2
v⊥ ∗
mv⊥ p∗
ρC = = = ⊥ (1.2)
a⊥ ~
qB qB
The cyclotron period can be defined as:
2πρC 2πm
τC = ∗ = (1.3)
v⊥ qB
and the gyrofrequency can be expressed as:
2π qB
ωC = = (1.4)
τC m
When an external force is introduced perpendicular to the local magnetic field, the guiding
center of a particle can drift across field lines. In the case of an external force generated by an
electric field, this is given by Equation (1.5), and is often referred to as “E-cross-B drift.”
~ ×B
E ~
~vE = 2
(1.5)
B
Additionally, centrifugal force felt by a particle bouncing between magnetic poles leads to
curvature drift, given as:
Ä ä
~ × b̂ · ∇ b̂
p2k B
~vc = (1.6)
qγm0 B2
where γ is the relativistic correction factor, and ~b is the unit vector in the direction of B.
~ A final
drift, the grad-B drift, caused by magnetic field gradients perpendicular to the local magnetic field
is given by Equation (1.7).

7
MB ~ × ∇⊥ B
~v∇B = (1.7)
qγ B2
where M is the magnetic moment of the particle. Together, Equations (1.6) and (1.7) are known as
the gradient-curvature drift. This drift, when considered in the magnetosphere, causes positively
charged particles to drift clockwise around the Earth when viewed from the north pole, and neg-
atively charged particles to drift counter-clockwise. These two drift populations, in combination
with each other, generate the ring-currents within the magnetosphere. A more thorough treatment
of this topic can be found in Roederer and Zhang (2014) and Northrop (1963).
1.3.2 The Adiabatic Invariants
The Lorentz equations describe three modes of motion for a particle trapped in a magnetic
field: Gyro-motion, as described above, bounce motion (particle bouncing between two mirror
points in a magnetic field), and a drift motion perpendicular to the magnetic field. Each type of
motion has a quasi-conserved quantity called an adiabatic invariant. These quantities are conserved
as long as the field changes slowly in time and space compared to the relevant properties of each
motion (time and space). These three invariants are described in following sections.
1.3.2.1 The First Adiabatic Invariant
The invariant associated with the gyromotion of a charged particle in a magnetic field is given
by Equation (1.8), and is often referred to as the magnetic moment of the particle.
p2⊥ p2 sin2 α
M = = (1.8)
2m0 B 2m0 B
This invariant is conserved if the changes in the magnetic field occur on a time scale large
compared to the gyroperiod (see Equation (1.3)), or the geometry of the field changes on a scale
large compared to the gyroradius (see Equation (1.2)). If changes are not large compared to these
properties, the invariant nature may be broken, and the particle motion can only be described by
the complete Lorentz Equation.

8
The literature exhibits confusion over what symbol should represent the first adiabatic in-
variant. Walt (1994) and Kivelson and Russell (1995) use the term µ to signify the invariant, while
Roederer and Zhang (2014) uses M . Other sources utilize µ to denote the non-relativistic form
of the invariant, while reserving M for the relativistic form. In this work, µ = M and are used
interchangeably.
1.3.2.2 The Second Adiabatic Invariant
The second invariant, which is associated with the bounce motion of a particle, is related
to the fact that the conservation of the first invariant (M ) within a converging magnetic field,
particles are subject to mirroring. The mirror field for a particle is given in Equation (1.9). This
field strength can also be computed as a relationship between the local B field and the particle’s
pitch angle, α.
B p2
Bm = = (1.9)
sin2 α 2m0 µ
where α is given by:
Ç å
−1 v⊥
α = tan (1.10)
vk
Within dipole-like field geometries, mirroring leads to a bounce motion along the guiding
center of the particle, leading to a periodic motion between the two poles. The invariant associated
with this periodic motion is given in Equation (1.11).
sñ ô
B(s)
I I I
J = pk ds = p cos α(s) ds = p 1− ds = 2pI (1.11)
Bm
sñ
s0m
ô
B(s)
Z
I = 1− ds (1.12)
sm Bm
where the integral I is along the field line intersecting with the particle’s position, sm and s0m are
the mirror points where B(s) = Bm . The integral I is a purely field-geometry integral, and it is
9
conserved as long as the kinetic energy and J are conserved. In realistic magnetic field conditions,
where the field is acted upon by an external force, I will no longer be constant. Utilizing the
invariant nature of the first invariant, M , a new field line invariant, K, can be constructed (Roederer
and Zhang, 2014). The equation for K is given in (1.13).
J p Z s0m »
K = √ = I Bm = (Bm − B(s)) ds = const (1.13)
2 2mM sm
The combination of M , K, and the conservation of energy allow for the definition of a par-
ticle’s guiding field line at any longitude, φ. The entire drift shell for a particle can be determined
numerically by iteratively searching for field lines that satisfy the same values of Bm and K (Roed-
erer, 1970; Roederer and Zhang, 2014).
1.3.2.3 The Third Adiabatic Invariant
The adiabatic invariant associated with the third type of motion, drift motion, is given by
Equation (1.14).
ZZ ZZ
Φ = ~ · d~s =
B ~ · n̂ ds
B (1.14)
π π
and represents the magnetic flux contained within a particle’s drift orbit. Numerically, we can
calculate this with the flux through Earth’s polar cap, bounded by the surface intersection of the
field lines making up the particle’s drift shell (Roederer and Zhang, 2014). As the time scale for
the drift orbit of a particle is much larger than those for gyro-motion and bounce-motion, this
invariant is the most likely to not be conserved. A convenient reference value, based on Φ, is the L∗
parameter; a value that represents where a particle would be located if all external contributions to
the magnetic field are slowly relaxed (see Section 1.5.2 for a description of external contributions).
The relationship between Φ and L∗ is given by:
2
2πBE RE
L∗ = (1.15)
Φ
10
where BE is the value of B at the equatorial point on the surface of the Earth, and is proportional
to Earth’s dipole moment. RE represents the radius of Earth (Roederer and Zhang, 2014).
1.4 Analysis and Visualization of Particles in the Magnetosphere
The use of L∗ is prevalent and critical in the analysis and visualization of the radiation belts
(Hudson et al. (1997, 2015); Elkington et al. (2002, 2004); Tu et al. (2012); Baker et al. (2004) to
name a small selection of examples), yet the means of calculating the values is mostly relegated to
empirical models of the magnetosphere.
Of the L∗ calculation libraries available, IRBEM (Boscher et al., 2012) utilizes by default the
IGRF model, but can be set to use a user-defined model as long as it meets specific requirements, and
implements the method for calculating L∗ found in Roederer (1970). LANLGeoMag (Henderson,
2010) solves the guiding center equations utilizing a selection from the Tsyganenko models. LANL∗
(Koller et al., 2009; Koller and Zaharia, 2011; Yu et al., 2012) takes the unique approach of utilizing
an artificial neural network to predict L∗ on a Tsyganenko model. Min et al. (2013a,b) utilizes a
technique, similar to Roederer (1970), which uses two conserved quantities to find drift trajectories
in a Tsyganenko model, but utilizes iso-energy contour from a discrete 2-D energy space, defined
by M and K to calculate drift orbits. None of these methods deal specifically with optimized
calculations of L∗ for large gridded MHD fields. The research in this thesis intends to provide a
solution to fill the need for determining the three invariants, M , K, and L∗ , in these complex field
geometries.
1.5 Numerical Models of the Earth’s Magnetosphere
As eloquently noted in Tsyganenko (2013), the geomagnetic field is “the principle agent
connecting our planet’s ionosphere with the highly variable interplanetary medium, incessantly
disturbed by dynamical processes at the Sun.” This reality has led scientists to devise models
of how the magnetosphere behaves. The following three sections describe three different types of
models used to aid in the understanding of how the Earth’s magnetosphere behaves. The first of the
11
sections covers the first order approximation of the magnetic field; the simple dipole. The second
section covers the series of Tsyganenko empirical models, and the third section briefly discusses
MHD models.
1.5.1 Dipole Geomagnetic Field Model
The first order approximation of the Earth’s magnetosphere is the basic dipole. The dipole
magnetic field provides many advantages when developing analytic methods for more complex fields,
as the field itself and many analyses of the field have analytic solutions, providing a baseline for
error analysis of methods.
The basic equations for a dipole can be found in many introductory books on space physics.
Kivelson and Russell (1995) provides equations to produce a dipole in Cartesian coordinates:
BX = 3xyMz r−5
BY = 3yzMz r−5 (1.16)
BX = (3z 2 − r2 )Mz r−5
where Bx , By , and Bz are components of the magnetic field, hx, y, zi representing the location
vector, and r is the radial distance from center of the Earth. Mz is the z component of the Dipole
Moment. In reality, Earth exhibits a difference between the spin-axis and the magnetic axis. This
difference induces what is called a dipole tilt. The basic model provided by Kivelson and Russell
(1995) does not, however, take into account any form of dipole tilt. For a more complete solution,
the equations provided in (1.17) incorporate the requisite tilt. These equations are derived from
the Tsyganenko T96 model code ((Tsyganenko, 1995, 1996)), and generate a dipole field at the
given location hx, y, zi in the GSM coordinate system.

12
Mz ÄÄ
2 2 2
ä ä
Bx = p 2 y + z − 2x sin(ψ) − 3 z x cos(ψ)
( x + y 2 + z 2 )5
Mz
By = −3y p 2 (x × sin(ψ) + z cos(ψ)) (1.17)
( x + y 2 + z 2 )5
Mz ÄÄ
2 2 2
ä ä
Bz = p 2 x + y − 2z cos(ψ) − 3 z x sin(ψ)
( x + y 2 + z 2 )5
For these equations, Mz is again the z component of the Dipole Moment (we use 3.14 × 104 nT · RE
3)
and ψ is the dipole tilt applied to the field (in radians). The vector hx, y, zi again serves as spatial
location (in units of RE ).
1.5.2 Empirical Geomagnetic Field Models
An improvement above and beyond the first order approximation requires additional contri-
butions to the magnetic field. In empirical models, these contributions are determined by fitting
a model to a large quantity of observational spacecraft data. These models are typically fit for
individual contributions to the magnetic field. According to Tsyganenko (2013), the total magne-
tospheric magnetic field vector B can be represented as the sum of these individual contributions.
Equation (1.18) gives an example of this combination methodology
B = BI + BE (1.18)
where BI represents the internal contribution to the the magnetic field (the main geomagnetic
field), and BE represents the external contributions to the magnetic field (associated with electric
currents flowing inside and outside Earth).
The external contribution, BE , can be further broken down into the sum of contributions
from individual influences. This results in BE being defined as in equation (1.19).
BE = BRC + BT C + BF AC + BIN T + BM P (1.19)
where BRC is the magnetic field contribution of the ring current, BT C is the contribution of the
tail current, BF AC is the contribution of the field-aligned current, BIN T is the contribution from
13
the interconnection field, and BM P is the contribution from the magnetopause field (also referred
to as the shielding field). The Tsyganenko magnetic field models calculate, in varying combinations
(depending on the specific model in question), external contributions, and combine them into a
unified BE . Adding these external contributions to an internal (BI ) field will result in a model of
the magnetosphere.
The following outlines some of the more commonly used Tsyganenko models. One of the
earlier empirical models used in the field was the Tsyganenko T89 Model (Tsyganenko, 1989)). As
an input, this model utilizes the planetary K index (KP ) (Bartels et al., 1939). The KP index is
a 3-hour global index based on a weighted average of local K values. These values are, in turn
a quantification of the disturbances to the horizontal component of the Earth’s magnetic field.
The index value ranges from 0 to 9. By Tsyganenko’s reckoning (Tsyganenko, 1995), the T89
model (and its earlier T87 iteration), suffered from three serious deficiencies: it had an unstable
“de facto” magnetopause; a crude method of parameterizing, utilizing only Kp ; and inaccuracies in
the Bz values of the equatorial magnetotail. The Tsyganenko (1995) model revisions were designed
to alleviate some of these deficiencies through adding a realistic shape to the magnetopause, a
fully controlled shielding component for all current contributions, a flexible tail and ring current
representations, and a more accurate fitting of spacecraft data to the field. Tsyganenko (1996)
describes the results of the completed T96 model. The inputs for the T96 model consist of solar
wind dynamic pressure (pdyn ), IMF By , IMF Bz , and Dst .
The Tsyganenko (2002b) (T02) variant was built upon the T96 model, and updated the
system with new methods. In addition to standard solar wind inputs, the model was adjusted to
add latent terms that depend on past solar wind conditions. This model updates the ring current
method from the simple empirical ring currents of the T96 and earlier models to a more accu-
rate approximation based the electric current calculated from an observed distribution of particle
pressure and anisotropy. The new ring current model also takes into account the strong dawn-
dusk asymmetry of the ring current during times of disturbance (more on the new ring current
model can be found in Tsyganenko (2000)). The Birkeland Currents simulation for this model were
14
completely redesigned to address a major deficiency of the old methods; insufficient flexibility. A
detailed description of the mathematical structure of the new Birkeland Currents module is given
in Tsyganenko (2002a).
In Tsyganenko and Sitnov (2007), a model is presented that expands the field to include
contributions from arbitrary azimuthal and radial variations of the geomagnetic field, allowing for a
more natural way to represent the field related to the tail and ring (partial and symmetric) currents.
In Tsyganenko (2014), model improvements were made that take into account the IMF dependent
shape of the magnetopause, and provided a physically more consistent global deformation of the
equatorial current sheet due to dipole tilt. The ring current module was updated to reflect a more
realistic background magnetic field, and the entire model valid range was extended to between 40
and 50 RE . Most recently, Tsyganenko and Andreeva (2015) provided an update to the model to
include the Newell coupling function (see Equation (2.1) in Chapter 2 ) (Newell et al., 2007) as a
driver for the model.
1.5.3 Magnetohydrodynamic (MHD) Geomagnetic Field Models
MHD models take a different approach to solving the magnetosphere problem. Instead of
utilizing observational data to define a model, these models treat the system as a magnetized
plasma, governed by the magnetohydrodynamic equations for the magnetosphere. A combination
of Maxwell’s equations and magnetized fluid equations, the MHD equations can be expressed as
follows (from Kivelson and Russell (1995)):
Continuity Equation:
δρ
= − ∇ · (V )ρ (1.20)
δt
Momentum Equation:
δV ∇p (J × B)
= − (V · ∇)V − + (1.21)
δt ρ ρ
15
Pressure Equation:
δp
= − (V · ∇)p − γp∇ · V (1.22)
δt
Faraday’s Law:
δB
= ∇ × (V × B) + η∇2 B (1.23)
δt
Ampere’s Law:
J = ∇ × (B − BI ) (1.24)
where ρ is the plasma density, V is the velocity, p is the pressure, B is the magnetic field, and BI
is the Earth’s internal field. The γ term is the polytropic index, and η is the magnetic diffusivity.
The Lyon-Feddar-Mobarry (LFM) 3D MHD magnetospheric model has been in use for nearly
three decades (Lyon et al., 2004). The model was designed to satisfy several criteria: maintain
~ = 0 to within the roundoff error; put resolution where it is needed; use special methods where
∇·B
there is a high magnetic field; and use an integrated ionospheric model. The model utilizes a finite
volume technique on a stretched spherical non-orthogonal adapted grid that places the majority of
the resolution closest to Earth, and takes the upstream solar wind conditions as input. The model
outputs its results in the SM coordinate system.
The LFM model solves a somewhat different form of the MHD equations. They are cast in
the form:
δρ
= −∇ · ρ~v , (1.25)
δt
~B
~
!
δρ~v B2 B
= −∇ · (ρ~v~v + IP ) − ∇ · I − , (1.26)
δt 8π 4π
~B
~
!
δEp γ B2 B
Å Å ãã
= ∇ · ~v ρv 2 /2 + P − ~v · ∇ · I − , (1.27)
δt γ−1 8π 4π
~
δB ~
= ∇ × (~v × B), (1.28)
δt
16
where EP = ρv 2 /2 + P ~ the magnetic field.

γ−1 , ρ is plasma density, ~v , velocity, P , pressure, and B,
Lyon et al. (2004) provides in-depth detail on the construction of the model.
1.6 Thesis Format
This thesis is broken down into 5 major parts. First, in Chapter 2 covers interactions between
CMEs and the solar wind, providing for analysis of observational errors vs. modeled transients.
Chapter 3 covers ideal analytic analysis for the adiabatic invariant algorithms, covering errors
associated with algorithms used to computer L∗ . Chapter 4 describes the effects of the LFM grid
on how the algorithms perform, and how they needed to be modified to work with LFM data, then
presents a brief real-world analysis utilizing the new tools. Chapter 5 covers serial and parallel
system timings, algorithm profiles, load balancing, and efficiency analysis. Finally, Chapter 6
possible improvements that can be implemented, and future directions of research.

Chapter 2
Observational Uncertainties and CME Transients
Figure 2.1: Overview of the Enlil simulation space.
The use of complex physics-based models is relatively new in space weather forecasting,
and there is much to learn about how they operate in a production forecast environment. When
predicting the transit of a coronal mass ejection (CME) through the solar system, space weather
forecast centers often utilize the WSA/Enlil + Cone modeling suite (Parsons et al., 2011; Millward
et al., 2013; Lee et al., 2013; Steenburgh et al., 2014; Mays et al., 2015). This suite is a complex
18
system that consists of three coupled models: the Wang-Sheeley-Arge (WSA) solar model (Arge
and Pizzo, 2000), a geometric CME cone model (Zhao et al., 2002; Xie et al., 2004), and the Enlil
ambient solar wind model(Odstrcil, 2003; Odstrcil et al., 2004; Odstrcil, 2004a,b). The WSA model
takes synoptic maps as input and utilizes a hybrid empirical and physics-based approach to simulate
the solar corona and solar wind through to the inner boundary of the Enlil Model. The cone model
is used to estimate the morphological and kinematic properties of CME ejecta from coronagraph
images; these properties are injected into the inner boundary of the Enlil model at the simulation
time necessary to model the propagation of the CME through the helioshpere. Enlil is a 3-D
magnetohydrodynamic (MHD) model that simulates the morphological changes to, and provides
time dependent conditions of, the solar wind through the heliosphere (to the outer boundary of the
simulation). As a unit, this coupled system attempts to provide a physically accurate representation
of solar wind conditions as CMEs travel from the Sun to the Earth. It should be noted that the
WSA/Enlil + Cone model suite does not provide a representation of the magnetic field associated
the CME ejecta cloud, thus limiting the usefulness of |B| values within the forecast. Figure 2.1
provides a large-scale overview of the Enlil simulation space, illustrating the scale and morphology
of a CME as it propagates through the inner solar system. The small dot at the leading edge of
the CME cloud represents the grid location containing Earth.
Although this coupled-model environment has seen increasing use over the past several years
(e.g. Baker et al. (2009, 2011, 2013b); Millward et al. (2013); Lee et al. (2013); Dewey et al.
(2015); Mays et al. (2015); Cash et al. (2015)) , the predictions of CME transits still suffer from
uncertainties related to cone model fitting from incomplete data (Pulkkinen et al., 2009; Lee et al.,
2013; Mays et al., 2015). It is noted in Pulkkinen et al. (2009) that when using a single spacecraft
coronagraph source for model fitting an assumption on the structure of the CME is required. The
assumptions that generate the cone model (i.e. the CME propagates as a geometric cone) are,
at best, only rough approximations of the actual ejecta cloud geometry. Furthermore, natural
uncertainties arise when manually fitting a cone to a CME due to ambiguities in identifying the
ejecta structure. Together, these conditions lead to uncertain central origins for the modeled CMEs.
19
Eliminating the human factor from ejecta identification, by means of automated fitting, makes for
more consistent cone estimations; however, uncertainties still exist in the modeled origins due to
the nature of the machine vision algorithms used (Pulkkinen et al., 2009; Jacobs et al., 2013).
As cone parameter uncertainties must be addressed in order to accurately predict a CME
transit, methods have been developed over time to help mitigate the associated issues. One such
method is ensemble modeling; a technique whereby multiple simulations are conducted for a given
event (with CME cone parameters generated with some statistical approach) to provide a proba-
bilistic forecast for CME arrival time (Lee et al., 2013; Mays et al., 2015). While these ensemble
predictions provide a means for generating a probabilistic forecast, a more detailed understanding
of the distribution of effects due to observational uncertainties is desired.
It is posited that the error responses to these observational input uncertainties and their
propagation within the numerical models can have a significant effect on a forecast outcome. To
test this hypothesis we have devised a method, called CME Impact Domain Analysis (CME
IDA), for describing the distribution of possible forecast results when utilizing the WSA/Enlil
+ Cone modeling suite (the method should also work, with some modification, for any other
heliospheric solar wind model). Several works in the literature point to the CME central origin
(solar latitude and longitude) being one of the most critical parameters for generating an accurate
forecast(Pulkkinen et al., 2009; Millward et al., 2013; Lee et al., 2013; Mays et al., 2015), therefore,
this study places a primary focus on the CME central origin. With this focus, a domain is defined
that is inclusive of model runs, with varying central origins, that will generate a CME impact at
our target location, in this case, Earth. Called the Impact Domain, this abstract domain can be
used to analyze the statistical distribution of CME effects on the solar wind, and how they vary
with the uncertainty of input.
Utilizing a Community Coordinated Modeling Center prediction for a CME that occurred on
October 06, 2013 at 1439 UTC, this work constructs, by example, a CME IDA. This method gener-
ates IDA analysis plots to describe the statistical distribution of the impact-time, impact-duration,
and effect on various solar wind parameters. Additionally, an analysis similar to superposed epoch
20
analysis (SEA) (Chree, 1913) is used to construct time-dependent tendencies for the domain. Fol-
lowing a description of the model environment and example of the method, the information obtained
from these analyses is discussed and a process for generalizing the method is described.
2.1 The Model Environment
In an effort to understand how uncertainties affect a forecast, a simulation modeling config-
uration was chosen for this study that parallels an actual forecast environment. Steenburgh et al.
(2014) outlines the concept of operations (CONOPS) for the WSA-Enlil operational transition
used for space weather forecasting at the Space Weather Prediction Center (SWPC), and provides
a useful resource for configuring the simulated forecast environment.
The Steenburgh et al. (2014) CONOPS document, as well as those by Parsons et al. (2011) and
Mays et al. (2015) indicate that, within an operational environment, the Global Oscillation Network
Group (GONG) synoptic magnetograms are typically used as the input for the WSA model, and
the cone model is used for CME kinematic parameter insertion into the Enlil model inner boundary.
With these documents as reference, it was decided to utilize the resources of the CCMC to conduct
simulation runs of the WSA/Enlil + Cone coupled system. The GONG magnetogram synoptic
maps were used as input, and the cone model profile generated by the CCMC/SWRC for the 06
October 2013 event was used for baseline CME parameter insertion.
2.1.1 The WSA Model
In this study, the simulation of a CME begins with the Wang-Sheeley-Arge (WSA) solar
model (Arge and Pizzo, 2000). This model, which is an improvement on the original Wang-Sheeley
model (Wang and Sheeley, 1992), is a hybrid empirical and physics-based model of the sun’s corona
and solar wind. The WSA model is composed of three parts. Ground based observations of the
photospheric magnetic field provide synoptic maps as input to a magnetostatic potential field source
surface (PFSS) model (Schatten et al., 1969) that determines the coronal field to 2.5 solar radii
(R ). Next, the PFSS provides input to the Schatten Current Sheet model (SCS) (Schatten, 1971)
21
which provides a magnetic field topology of the upper corona. Finally, the model uses an empirical
velocity relationship to provide solar wind speed at the outer coronal boundary (truncated at 30R )
(Arge and Pizzo, 2000) .
For input to the Enlil model, only the innermost portion of the SCS solution is utilized,
providing Enlil with the radial magnetic field and solar wind speed at 21.5R . Densities and
temperatures at the boundary are not provided directly by WSA, but are inferred by assuming
mass flux conservation and pressure balance (Odstrcil, 2003; Lee et al., 2013).
2.1.2 The Cone Model
Within the forecast environment, one of the primary means of determining kinematic and
morphological properties of CMEs is through the application of the cone model (Millward et al.,
2013; Steenburgh et al., 2014; Mays et al., 2015). In Zhao et al. (2002) it was observed that, for
CME’s on the limb of the sun, most (but not all) coronal mass ejections propagate nearly radially
for the first couple of solar radii. If the assumption is taken that earthward facing (halo) CMEs
also exhibit this behavior, then geometric analysis of the CME images can produce a 3-dimensional
representation of the CME, modeled as a geometric cone. Zhao et al. (2002) proposed an iterative
method for fitting the cone to the observed picture. This was extended by Xie et al. (2004) into an
analytic method, which reduced the time and complexity required to calculate a cone fit.
Figure 2.2, a figure from Pulkkinen et al. (2009), provides the relationships between various
parameters of a CME as calculated via the cone model. Within this figure, the propagation di-
rection of the CME cloud is defined on the (x, y, z) coordinates, while the plane-of-sky is defined
on the (x0 , y 0 , z 0 ) coordinates. The two ellipses represent the boundary of the CME as seen in two
sequential coronagraph images. ω is the cone half-angle, and α, along with θ, defines the direction
of propagation. With the assumption that CMEs propagate radially from the Sun center, the di-
rection of propagation implies the central origin of the CME. The velocity is calculated from the
relationship x = x0 + v · ∆t, where x and x0 are determined by fitting the same cone profile to two
sequential coronagraph images, and ∆t is determined by the elapsed time between the two images.
22
utomatic Determination of the Conic CME Model Parameters
gure 1 Cone model ′ , z′ ) refers to

Figure 2.2: and Cone different
Model: “The parameters
plane (y 0 , z 0 )associated
refers to the withplane the
of skymodel.
(x0 -axisThe plane
points toward (ythe
′ -axis points
ane of sky (xobserver, toward
i.e. toward the observer,
the reader). Angle αi.e. toward
defines the reader).
the direction of the Angle α defines
propagation of the CMEthe direction of
in the (y 0 , z 0 ) plane, i.e. ′
the ′
angle between
opagation of the CME in the (y , z ) plane, i.e. the angle between the y 0 -axis and the x-axis as ′
projected into the (y 0, z0)
the y -axis and the x-axis as projec
′ ′ plane. Angle θ defines the rotation of the cone off of the (y 0 , z 0 ) plane, i.e. ′the ′angle between the
to the (y , z )x0plane. Angle
- and x-axes. defines
θ the
ω is opening thehalf-angle
rotationof of thethe cone
cone, x0 theoffinitial
of the plane,
(y , zof)the
distance i.e. the
cone front in angle betw
′
e x - and x-axes. ω iscoordinates
the opening (x, y, half-angle ofvelocity
the cone,
the rotated z) and v is the 0 the
of thexcone initial
front distance∆tofindicates
propagation. the cone front in
the time interval during which the cone front propagates from x to x.”
tated coordinates (x, y, z) and v is the velocity of the cone front 0propagation. $t indicates the time inte (Pulkkinen et al., 2009)
uring which the cone front propagates from x0 to x.
mages. The objective is attacked in two different parts. In the first part in Section 2, ima
rocessing techniques are utilized to extract transients from white-light LASCO C3 co
agraph data. In the second part in Section 3, times series of the extracted transients
sed in an inversion scheme seeking for the optimal cone model parameters. A bootst
ethod is used to provide probability distributions of the model parameters. In Section
itial validation of the new automatic procedure is carried out by means of comparisons
anually determined cone model parameters. Finally, in Section 5 the results of the pa
23
According to Millward et al. (2013), when faced with only a single coronagraph perspective,
forecasters rely on the Xie et al. (2004) method. With this method an ellipse is drawn over the
coronagraph image with the minor axis intersecting the center of the sun. A cone can then be
calculated to determine the central origin (heliospheric latitude and longitude) and the cone angle
in terms of the major and minor axis of the ellipse, its distance from origin, and the angle of the
minor axis with respect to the horizontal. Their experimental results show, as a CME halo tends
toward circular geometry, the parameters derived are extremely sensitive to the exact form of the
ellipse. The paper explores this phenomenon through multiple fittings of two different CMEs. It
was determined that the CME parameter most affected by the model uncertainties is the latitude of
the central origin, giving greater uncertainty to the origin point and direction of CME propagation.
Millward et al. (2013) further noted that the absolute direction of events is a critical parameter in
assuring accurate arrival times.
It is obvious from these works that determining exact initial parameters for a CME is not
possible from the limited information available. The assumption that a CME propagates as a
geometric cone is only a rough approximation of the actual ejecta cloud morphology, and the natural
uncertainties that arise when manually fitting a cone to a CME cause ambiguities in identifying
the cloud edge, making an exact cone fit unlikely. Still, the cone model seems to be the best choice
available for modeling a CME in near-real-time space weather prediction due to its simplicity and
speed of calculation.
2.1.3 The Enlil Model
The Enlil solar wind model (Odstrcil, 2003; Odstrcil et al., 2004; Odstrcil, 2004a,b), a 3-
D MHD model of the solar wind within the heliosphere, is used to provide, in conjunction with
the cone model, the propagation of CMEs through the solar system. Kinematic properties, as
determined by the cone model, are injected into the inner boundary of the Enlil model, and are
propagated through the simulation space, interacting with the ambient solar wind.
The Enlil model solves a set of time-dependent MHD equations in conservative form on a
24
spherical grid to propagate solar wind conditions for the time period being simulated. The sys-
tem solves eight coupled nonlinear partial differential equations utilizing a version of the Total
Variation Diminishing (TVD) technique (Harten, 1983; Harten et al., 1987), known as Total Varia-
tion Diminishing Lax-Friedrich (TVDLF)(Tóth and Odstrčil, 1996), modified to include a two-step
time-discritization (Odstrcil, 2004a). This method is second-order accurate away from shocks and
discontinuities, and provides stability that ensures non-oscillatory solutions (Odstrcil, 2003).
To facilitate a more detailed description of the analyses performed in this study, Table 2.1
provides a breakdown of the variables tracked throughout the execution of the Enlil simulation.
This information is similar to that found in Table 12 of Odstrcil (2004b), but reflects the state of the
current version of the Enlil Model (v. 2.8f). The Name column denotes the name of the variables
within the NetCDF output files, the Description column is the plain-language identification of the
variables, and the Units column provides the units associated with the values within the variable.
Variables of interest, and that are also used in analyses in this paper, include X1 (helio-
spheric position), X2 (meridional position), and X3 (azmuthal position), which define the spatial
coordinates for the simulation grid (in the HEEQ+180 coordinate system). B1, B2, and B3 are
respectively the radial, meridional, and azimuthal magnetic field vector components. Similarly V1,
V2, and V3 are the radial, meridional and azimuthal velocity vector components. D is the mass
density, T is the temperature and DP traces the evolution of the CME cloud (cloud tracer).
2.2 Mitigation of Uncertainties with Ensemble Modeling
Mays et al. (2015), Lee et al. (2013) and Millward et al. (2013) all provide in-depth detail
on the problems encountered when fitting the cone model to CMEs in coronagraph images. To
summarize, the issues can be reduced to problems of identifying features for any given CME. It
is necessary to accurately identify the ejecta boundary to obtain a proper cone fit; If the CME
boundaries are unclear, it becomes difficult to identify the proper ellipse, leading to uncertain
measurements. One method for mitigating these uncertainties is through the use of ensemble
modeling; a technique whereby multiple simulations are conducted for a given event (with the
25
Table 2.1: Variables tracked by the Enlil Model
Name Description Units

X1 Heliospheric Position m
X2 Meridional Position radians
X3 Azimuthal Position radians
B1 Radial Magnetic Field T
B2 Meridional Magnetic Field T
B3 Azimuthal Magnetic Field T
V1 Radial Velocity m/s
V2 Meridional Velocity m/s
V3 Azimuthal Velocity m/s
D Mass Density kg/m3
T Temperature K
DP Cloud Tracer -
BP Polarity Tracer -
DT Physical Time Step s
T IM E Physical Time s
N ST EP Numerical Step -
26
CME cone parameters based on some statistical approach) to provide a probabilistic forecast for
CME arrival time. The studies by Lee et al. (2013) and Mays et al. (2015) provide a look at two
different approaches to mitigating uncertainties with ensemble modeling. While the approaches
are different, they both seek to provide more realistic forecasts of arrival time through statistical
means.
The Lee et al. (2013) study utilizes an ensemble of eight WSA/Enlil + Cone runs to generate
a probabilistic forecast. A cone software routine (utilizing the Xie et al. (2004) method) developed
at the NOAA/SWPC is used to calculate cone parameters in a way designed to capture different
possible cone configurations. The authors utilized the SOlar and Heliospheric Observatory (SOHO)
Large Angle and Spectrometric COronagraph (LASCO) C3 coronagraph images for the primary
cone fit. The study also utilized Solar TErrestrial RElations Observatory (STEREO) B COR2-B
coronagraph images to constrain the radial distance of the cloud and the cone angle. For each
ensemble set member, a cone fit was conducted with a fresh initialization of the cone tool to ensure
that any previous fit parameters were not transferred to the next member. The fit was conducted in
eight different fitting sessions utilizing slightly different values for angular width and leading edge
distance (guided by the STEREO observations). For the study in question, the resulting ensemble
set generated fits with central origins varying by one to three degrees in both latitude and longitude.
The study consistently produced early CME arrival times, leading the authors to explore the effects
of background solar wind as a contributing factor. They concluded that predicting shock arrival
times accurately depends not only upon the input cone CME parameters but also on the reliable
specification of the background solar wind.
Mays et al. (2015) describes another method for real-time ensemble modeling utilizing a soft-
ware tool, developed by the CCMC for real-time CME analysis, called Stereoscopic CME Analysis
Tool (StereoCAT). This method utilizes stereoscopic triangulation to produce a cone fit from two
different fields of view (e.g. STEREO A and STEREO B). Utilizing the StereoCAT system in “en-
semble mode”, a forecaster measures CME leading edge height and angular width. The graphical
interface is reset, and the measurements are repeated a number of times. These measurements
27
are combined in all possible combinations to generate the probabilistic ensemble set, resulting in
n = m2 cone fittings, where m is the number of independent measurements made. The ensemble
set is then used to produce n WSA/Enlil + Cone model runs, resulting in a distribution of n
predicted arrival times at the location of interest. The study showed that for 47% (8 out of 17)
of the ensemble runs that generated impacts (of the 30 ensembles conducted), the observed CME
arrival time was within the spread of ensemble predictions. The conclusion was drawn that the
initial distribution of CME input parameters has significant influence on the accuracy of a CME
prediction.
With both methods, the useful information obtained is a spread of possible arrival times
based on the repeated “best guesses” of the forecasters. Both methods rely on independent cone
fittings (no fixed variables), and do not fully analyze the behavior and sensitivity of individual
cone properties (i.e. central origin, cone angle, velocity), or the sensitivity to variation within
different portions of the domain space. The CME Impact Domain Analysis (CME IDA) method
is an attempt to provide a means of examining the sensitivity of individual parameters, through
the entire impact domain, utilizing the distribution of associated effects within the domain. As a
starting point, the method is presented by examining the sensitivity of variation in central origins,
as this parameter has been identified through forecasting experience to be one of the more important
parameters (Millward et al., 2013).
2.3 Analyzing Effects of Uncertainty with CME Impact Domain Analysis
To gain a firm understanding of how uncertainties in CME central origins affect forecast
results, it is important to understand how they affect the propagation of a CME through the Enlil
model itself. Furthermore, it is necessary to understand how this propagation affects the potential
geoeffectiveness of the event. To facilitate this understanding, a method was developed for analyzing
the entire CME impact domain. The following sections outline, through example, how to construct
a CME IDA. The discussion begins with the selection of the 06 October 2013 CME event, then
explores the concept of impact definition. From there the mechanics of constructing the impact
28
domain are discussed. Finally, utilizing the discussion of solar wind - magnetospheric state variable
coupling functions presented in Newell et al. (2007), an analysis of the domain, and the products
produced, is presented.
2.3.1 The 06 October 2013 CME Event
As a starting point for this analysis, the 06 October 2013 CME event was selected. This
event has the advantage of being the only Earthward-directed CME in the simulation time-frame,
providing a relatively clean heliosphere. The event has also been studied by other researchers,
providing additional information on the nature of the event (e.g. Foster et al. (2015); Hudson et al.
(2015); Mays et al. (2015)).
Following the selection of the event, it was modeled utilizing the cone profile produced by
the CCMC/Space Weather Research Center (SWRC) at the time the CME occurred (the cone
profile is published in the CCMC Space Weather Database of Notifications, Knowledge, Information
(DONKI)). Figure 2.3 shows a snapshot of the CCMC Runs on Request (RoR) results for this
simulation. The figure presents views of the heliosphere in the ecliptic plane (a), the meridional
plane (b), and on the 1 AU shell (c). Image (d) provides a comparison between observational data
and the model results at Earth. This cone profile called for a CME ejection velocity of 790km/sec,
a cone half-angle of 25.0◦ , central origin longitude/latitude of 6.0◦ /−15.0◦ and a 21.5R crossing
time of 18:19 UTC. These values were utilized as a baseline for all model runs within the impact
domain. As an interesting side note, the Mays et al. (2015) ensemble study produced median
parameters for this event of: Velocity = 747km/s; Latitude = 1◦ ; Longitude = 2◦ ; Cone half-angle
= 16◦ . This resulted in a CME transit time of 53.0 hours, compared to the CCMC-based prediction
of about 76 hours, providing additional evidence to the uncertainty of cone fittings.
2.3.2 Defining an Impact
The definition of an impact is central to the results of two primary CME IDA analyses: time
to impact and impact duration. Depending on how an impact is defined, different information will
29
Figure 2.3: Global view of the 06 October 2013 CME event, shown at 0600 on October 10: WSA-
Enlil + Cone scaled density is shown in (a) constant Earth Latitude plane, (b) meridional plane
of Earth, and (c) the 1 AU shell. CME cone model parameters are from the CCMC/SWRC event
prediction (Velocity: 790km/s; cone half-angle: 25.0◦ ; lon/lat: 6.0◦ /−15◦ ; 21.5 R crossing: 18:19
UTC). Panel (d) compares the simulation (blue) compared to the measured (red) density values at
Earth. The dashed blue line represents the background solar wind density, and the yellow shaded
region indicates the CCMC/SWRC prediction for CME impact.
30
Impact Definition for 06 Oct 2013 CME
1.0 Norm DP
Norm dDP
dt
CME Impact Region
0.8
0.6
Normalized Trace
0.4
0.2
0.0
−0.2
1.0 Δ Density (Ymax = 11.08 N/cm3)

Δ Radial Velocity (Ymax = 46.57 km/s)
Δ |B| (Ymax = 2.10 nT)
0.8 CME Impact Region
0.6
Normalized Trace (Y/Ymax)
0.4
0.2
0.0
−0.2
−0.4
0 0 0 0 0 0 0 0 0 0
2:0 0:0 2:0 0:0 2:0 0:0 2:0 0:0 2:0 0:0
-1 -0 -1 -0 -1 -0 -1 -0 -1 -0
ct Oc
t
Oc
t ct ct Oc
t
Oc
t ct Oc
t
Oc
t
09O 10 10 11O 11O 12 12 13O 13 14
Figure 2.4: Illustration of the impact definition utilized in this study. The top panel provides
normalized traces for DP (gray line) and dDPdt (dot-dash red line). The defined impact region is
shown in the shaded area. The bottom panel provides the relationship of the impact region to the
Y
change in solar wind parameters. The plot is normalized by Ymax where Y is the original value for
the trace, and Ymax is the original maximum (provided in the plot legend).
31
be presented in these analyses. This concept is illustrated with Figure 2.4; an illustration of how
this study defines an impact. The top panel provides normalized traces for DP (the cloud tracer)
dDP
and dt (the rate of change of the cloud tracer). The defined impact region is shown in the shaded
area. The bottom panel provides the relationship of the impact region to the change in density,
Y
radial velocity, and |B|. The plot traces are normalized by Ymax where Y is the original value for
the trace, and Ymax is the original maximum value. The values for Ymax are provided in the plot’s
legend. It can be seen that the impact definition (shaded yellow region) reveals only a subset of
the information presented by the model runs. This information is reflected in the distributions
for arrival time and contact duration. This study utilizes the entire disturbance area, defined as
∆tda = tdf − tdi (where tdf is the time of last disturbance, and tdi is that of the first disturbance),
for solar wind CME IDA plots, as well as for quartile plots, so as to look at effects such as pileup
and rarefaction. The region of pileup that precedes a CME can clearly be identified in Figure 2.4
by the time-gap between the cloud tracer and the solar wind tracers. The rarefaction region can
be identified where the change in solar wind traces fall bellow zero. A more focused study of just
the impact region is possible by adjusting the ∆tda definition.
For the remainder of this paper, the impact will be defined based on the cloud tracer provided
in the Enlil output. The rate of change of the cloud tracer is used to identify the upward trending
inflection point of DP (the maximal increase). We consider the CME to impact the target when
DP 0 > 0 and DP 00 = 0. To eliminate the outlying runs that would be tagged as an impact, but have
little or no impact on the solar wind, we restrict the impact set to runs where max(DP ) > 1×10−5 .
This threshold value was chosen somewhat arbitrarily based on observed trends in the data, however
at only 0.000039% of the maximum cloud tracer value for the domain, it does reduce the impact
set to a meaningful domain. For simplicity, we define the end of CME contact to be the time at
which DP falls back below the value of DP at impact.
The authors acknowledge that there are other definitions of an impact. Mays et al. (2015), for
instance, define an impact as a sharp increase in modeled solar wind dynamic pressure. Lee et al.
(2013) define an impact as the first steep rise in the solar wind speed and magnetic field. From
32
these definitions it is clear that a standard definition of an impact is not in use. The method used
in this paper utilizes the same idea (sharp increase in a solar wind parameter), but applies it to the
Enlil variable responsible for tracing the CME ejecta cloud. It is argued that this definition, while
different from others, is suitable to demonstrate the method being described. It is not expected that
the choice of a different reasonable impact definition will change the effectiveness of the method.
2.3.3 Finding and Populating the Impact Domain
With the nature of an impact defined, it is fairly straightforward to find the boundaries of
the impact domain. As mentioned in Section 2, the impact domain is defined as the set of Enlil
runs where an impact is generated at the given target location. The free variables in this study are
those that define the central origin of the CME (Latitude and Longitude). The impact domain is
thus an abstract 2-Dimensional domain with Solar Latitude as one dimension and Solar Longitude
as the second dimension.
Looking at Figure 2.5, the nature of the impact domain can be clearly seen. The plots in
this figure are Impact Domain Plots. They are used to display information related to the domain
as a whole. In this study, the Y-axis denotes the Solar Latitude, and the X-axis denotes Solar
Longitude. The color bars indicate the values of the information being plotted. Contour lines are
used to provide a better visual perspective of the distribution within the domain. Figure 2.5 (a)
provides the distribution of max(DP) (The maximum value of the cloud tracer at Earth for each Enlil
Run) over the impact domain for the 06 October 2013 event. The values are provided in percentages
of the maximum, with the maximum (the × symbol on the plot) located at (0◦ lon, 5◦ lat), which
is essentially along the Sun-Earth line (within the resolution of the grid). Image (b) presents
max( dDP
dt ) (the maximum value of DP’ at Earth for each Enlil run) within the impact domain,
and is also plotted as a percentage of maximum, which is located at grid coordinates (−5◦ , 10◦ ).
Although the plots in Figure 2.5 were generated from the results of the IDA, they serve as a
good reference for defining the domain boundaries. Recalling that the definition of impact utilized
relies on the cloud tracer and its derivative, the extent of the impact domain can be seen with these
33
Impact Domain for October 6, 2013 CME Impact Domain for October 6, 2013 CME
% Maximum Cloud Tracer Values % Maximum dDPdt
40 100 40 100
10
% Max Cloud Tracer Values 10
90 90
80 80
20 20
Solar Latitude (HEEQ)

80 90
70 70
90
80
dDP
dt
60 60
40
70
60
% Max
0 50 0 50
70
0
50
50 6
30
40 40
30 40 20 30 20 30
−20 −20
20 20
10 10
−40 0 −40 0
−40 −20 0 20 40 −40 −20 0 20 40

Solar Longitude (HEEQ) Solar Longitude (HEEQ)
(a) (b)
Figure 2.5: Impact domain plot (a) provides the distribution of max(DP ) over the impact domain
and plot (b) shows the distribution of max( dDP ◦
dt ) (both in % of maximum). The dots located at 6.7
◦
latitude, 0 longitude represent the Sun-Earth line. The × identifies the maximum values within
the domain.
34
plots. Plot (a) shows how the lower bound on DP restricts the domain, while plot (b) provides the
distribution of “impact intensity”, where the intensity is a measure of the fastest positive changing
cloud tracer value.
To locate the impact domain boundaries, a simple linear search of the domain space can be
conducted. For this study, runs were performed at 15◦ intervals along the +X axis of the domain
until a CME miss was generated. Once a miss was generated, a search at 5◦ intervals was conducted
between the last CME hit and the first CME miss from the 15◦ search space. The first miss in the
5◦ interval space was then defined as the outer boundary in the +X direction. The domain was then
defined as a box around the origin at the distance determined from the search. Verification-runs
were conducted in the +Y, -Y, and -X directions to verify the extents of the domain. The size of
the domain was increased as necessary.
Determining the spatial resolution for the domain is dependent upon the resolution of the Enlil
Model. This study utilizes low-resolution Enlil runs, which provide a grid resolution of 256×30×60.
This translates to a 4◦ separation between cells. To ensure a meaningful result from the simulation,
and avoid excessive computational expense, a commensurate spatial resolution of 5◦ was selected
for this study.
Utilizing this resolution to define the free variables, and using outer boundaries of −50◦ and
50◦ in both the latitude and longitude directions (as determined in the domain search), a RoR
ensemble job was requested utilizing the forecast information from the 06 October 2013 CME for
static parameters. This generated a total of 441 low-resolution WSA/Enlil v2.8f runs to populate
the domain.
Populating the impact domain was a fairly straightforward task, thanks to a custom ensemble
job submission form created by CCMC for the purpose of this research (the submission form is now
available as part of the RoR system on the CCMC website). This RoR ensemble submission form
allows for the specification of fixed parameters as well as providing a method for defining what
parameters to vary, and how to vary them, between runs. It was a simple task to specify the
domain conditions within the form and submit the jobs.

35
2.3.4 Solar Wind Impact Domains and Quartiles
Presentation of information within the impact domain is handled through two styles of plots.
First, information is plotted on Impact Domain Plots (IDPs). Second, time-series statistical infor-
mation is presented, in a style similar to superposed epoch analysis, on quartile plots.
IDPs take the form of a 2-Dimensional grid representing solar longitude on the X-axis and
solar latitude on the Y-axis. Each node of the grid represents the cone model central origin for a
model run. Values at the nodes represent model and/or calculated values at the target location
(e.g. Earth), and are plotted against a color-bar to represent ranges. Contour lines are added to
provide a visual perspective to the distribution of results across the domain.
Time dependent information is plotted in time-series quartile form. Conceptually, these plots
are generated by arranging the time series data into a two-dimensional array with run number on
the Y-axis and time on the X-axis. The X-axis is epoch-aligned to impact time and the array is
sorted over the Y-axis, placing the largest values at the top of the array, and the smallest values at
the bottom of the array. The quartile time-series lines can then easily be defined as the 25%, 50%,
and 75% points on the Y-axis. The minimum value series resides at the bottom of the array, while
the maximum value series resides at the top.
2.3.4.1 Impact Domain Plots for the 06 October 2013 Event
Two major items of interest in space weather forecasting are the CME arrival time at Earth
(when to expect a storm), and the associated CME contact duration (length and strength of a
storm). These categories are ideally suited to IDPs, and their distributions for the 06 October
2013 CME are provided in Figure 2.6. Plot (a) provides the distribution of time to CME impact
at Earth. This time to impact is defined as ∆t = t1AU − t21.5R (the time from when the CME
crosses 21.5R to when it reaches 1AU; the impact time can instead be calculated in terms of
absolute arrival time based on CME onset time, if desired). The distribution indicates that the
earliest arrival is, to the resolution of the CME IDA, near the Sun-Earth line, with central origin
36
Impact Domain For October 6, 2013 CME

Time to Impact
113
109
40 12
1
105 125
121
CME Time to Impact (Hours)

88
117
20
Solar Latitude (HEEQ) 113
109
80
105
0
100
96
100
92
92
11
84
7
96
−20 88
84
80
−40 76
−40 −20 0 20 40
Solar Longitude (HEEQ)
(a)
Impact Domain For October 6, 2013 CME
CME Contact Duration
0 11
5
40 11 0 43
22 1116
16
CME Contact Duration (Hours)

38
22
11
5
11
20 32
27
0 22
16
32
11
−20 11
38
5 5
32
11
−40 27 0 0
−40 −20 0 20 40
Solar Longitude (HEEQ)
(b)
Figure 2.6: Impact Domain Plots showing the central origin uncertainty distributions for time to
impact and CME Contact Duration at Earth for the 06 October 2013 CME event. The X-axis
displays the central origin longitude and the Y-axis, latitude. Each node represents a single run’s
CME central origin. The ◦ located at 6.73◦ latitude, 0◦ longitude represents the Sun-Earth line at
the time of the event. The × identifies the minimum time to impact within the domain (a) and the
maximum CME contact duration (b).
37
Table 2.2: Minimum and Maximum values, and their central origin coordinates, for the Impact
Domain Plots
Property Central Origin Value

Min Time to Impact (0, 5) 76.0 (hours)
Max Impact Duration (10, -15) 43.2 (hours)
Min ∆ Density (-15, 20) -10.5 (N/cm3 )
Max ∆ Density (-5, 10) 19.8 (N/cm3 )
Min ∆ Velocity (10, 10) -5.9 (km/s)
Max ∆ Velocity (0, 5) 78.6 (km/s)
Min ∆|B| (0, 10) -2.3 (nT)
Max ∆|B| (-5, 10) 7.8 (nT)
Max DP (0, 5) 0.255 (unitless)
Max dDP/dt (-5, 10) 6.982e-03 (unitless)
grid coordinates of (0◦ , 5◦ ) and an impact time of 76.0 hours. It is also clear that the distribution
is not symmetric. Impact times are distorted across the domain, likely due to interactions of the
CME with the ambient solar wind. The gradient of the arrival time is steepest toward the left of
the plot, indicating that, for this event, uncertainties between 0◦ and −40◦ solar longitude will have
a more profound affect on this forecast than would uncertainties toward +40◦ longitude.
Plot (b) provides the distribution of CME contact times across the domain as it propagates
through space. Clearly the impact duration is asymmetric. The maximum occurs at central origin
(10◦ , −15◦ ), with a contact duration of 43.2 hours. The distortions in the contact duration distri-
bution are likely due to the changing morphology of the CME as it interacted with the ambient
solar wind during its transit to 1AU. The image in Figure 2.1 shows the general morphology (for
this study) of a CME in transit. An extended limb can be clearly seen along the top and another
down the left side of the CME cloud; features such as these would produce much longer contact
time should they pass directly through the target location.
The need to understand the impacts of space weather at Earth is extremely important in
space weather forecasting. Utilizing IDPs, information regarding the distribution of potential geo-
effectivness at Earth can be inferred. Going beyond time to impact and impact duration, Figure
2.7 graphically presents distributions of solar wind perturbations at Earth. Each plot represents
38
Maximum Density Disturbance Minimum Density Disturbance
Max Density Disturbance (N/cm3)
Min Density Disturbance (N/cm3)

40 19.85 40 -0.00
9. -6.56 -7
0 92 -9.19 .
12.49 17.37 88 -1.31
14.8
17.37
20 14.89 20 -2.63
5
-5.2
12.40 -3.94
0 9.92 0 -5.25
-3.
8
94
2.4
7.44 -6.56
4.9
6 7.4 -2
−20 4 4.96 −20 .6 -7.88
3
2.48 -9.19
-1.31
−40 0.00 −40 -10.50
−40 −20 0 20 40 −40 −20 0 20 40

(a) (b)
Maximum Radial Velocity Disturbance Minimum Radial Velocity Disturbance
-1.48
Max Velocity Disturbance (km/s)
Min Velocity Disturbance (km/s)

40 78.56 40 0.00
8
-2.
.4
29 22
-1
-0
. 68.74 -0.74
49.1 39 47
.7
-3
-2
58. 0 .29 .7
4
.9
9 0
68.74 2
6
20 58.92 20 -1.48
-4
-5
.
.
44
18
49.10 -2.22
0 39.29 0 -2.96
9.83
29.47 -3.70
5
−20 .6 19.65 −20 -4.44
19
9.83 -5.18
−40 0.01 −40 -5.92
−40 −20 0 20 40 −40 −20 0 20 40

(c) (d)
Maximum |B| Disturbance Minimum |B| Disturbance
40 7.82 40 -0.00
-1.15
-1.
Max |B| Disturbance (nT)
6.84 -1 44
Min |B| Disturbance (nT)
4.89 -0.29
5.86 -2 .73
.0
20 6.8 5.86 20 1 -0.58
4
4.89 -0.86
3.91
0 3.91 0 -1.15
-0
.8
0.9
2.93 6 -1.44
93
8
2.
-0
−20 1.96 −20 -1.73

. 58
1.96
0.98 -2.01
-0.29
−40 0.00 −40 -2.30
−40 −20 0 20 40 −40 −20 0 20 40

(e) (f)
Figure 2.7: Impact Domain Plots showing the central origin uncertainty distributions for the Min-
imums and Maximums of solar wind properties: (a) & (b) ∆ Density; (c) & (d) ∆ Radial Velocity;
(e) & (f) ∆ Total Magnetic Field (|B|). Values are minimum/maximum over the course of the
event.
39
either the minimum or maximum value (over the course of the solar wind disturbance, ∆tda ) of the
property being displayed. Utilizing coupling functions, such as those found in Newell et al. (2007),
it is possible to illicit understanding of the geoeffectivness a CME.
Newell et al. (2007) discusses 20 Solar Wind-Magnetosphere Coupling functions and how well
they correlate solar wind conditions to 10 magnetospheric state variables. A function is provided
that correlates to 9 of the 10 state variables (the only exception is Dst, for which the Temerin
and Li (2006) coupling function correlates best). The coupling functions can be used to infer
geoeffectiveness of the event being examined. The Newell et al. (2007) function is given as
dΦM P 2 θC
Å ã
4 8
3
= v BT sin
3 3 (2.1)
dt 2
dΦM P
where dt is the rate magnetic flux is opened at the magnetopause, v is the bulk solar wind
speed, θC is the interplanetary magnetic field (IMF) clock angle, given by tan−1 (By /Bz ) and the
perpendicular component of the magnetic field (in GSM coordinates) is given by BT = (By2 +Bz2 )1/2 .
To facilitate a discussion on possible geoeffectivness of the 06 October 2013 CME, Figure 2.7
presents the minimums and maximums for three of the solar wind properties: change in Density
over background (∆ Density), change in radial velocity over background (∆ Radial Velocity), and
change in total magnetic field over background ( ∆ |B|). The plots on the left hand side of the
figure represent the maximums, and the plots on the right present the minimums. The maximum
and minimum disturbances within the Impact Domain are denoted with the symbol ×, and the
location of the Sun-Earth line is denoted with symbol ◦.
Plots (a) and (b) display the distributions for minimum and maximum disturbances to Num-
ber Density (N/cm3 ). Though density does not figure directly into the Newell et al. (2007) coupling
equation, it does figure weakly into the Temerin and Li (2006) coupling function which holds the
best correlation to Dst. Plot (a) represents the maximum compression in density, and plot (b)
indicates the maximum rarefaction (minimum ∆ density). It can be seen from these plots that
both the maximum and minimum density disturbances occur away from the Sun-Earth line, with
the maximum rarefaction occurring at a considerable angular separation. The distribution for the
40
maximum density disturbance is slightly elongated toward the lower right of the plot, with a steeper
gradient toward the upper left. The minimum density disturbances also show the elongation trend
to the lower right, however there is a large stable central region within the distribution. Uncer-
tainties in CME central origin around and toward the upper left of the plot appear to produce
little variability from +5◦ to −20◦ longitude and −5◦ to +20◦ latitude. The magnitudes of both
the minimum and maximum density disturbances are also similar; the disturbances minimize at
-10.5 N/cm3 and maximize at 19.8 N/cm3 . Given the coupling functions, it is unlikely that the
distribution in changes of density will, by themselves, cause much change in geoeffectivness of the
CME.
The radial velocity effects (shown in plots (c) and (d)) have a much larger input into the
coupling functions, and thus provide us with a better view of a distribution of effects. The velocity
distribution peaks near the Sun-Earth line (grid location (0◦ , 5◦ )) with a ∆ Velocity of 78.6 km/s,
and minimize at grid location (10◦ , 10◦ ) with a ∆ Velocity of -5.9 km/s. The distribution of velocity
perturbations is nearly symmetrical for maximum values, but are extremely distorted to the right
for minimum values. At first glance it appears that the negative variations are extreme, however
the values of negative perturbations are an order of magnitude lower than the positive, making the
negative distribution less significant.
The Minimum and Maximum ∆|B| distributions can only provide a limited amount of infor-
mation due to the lack of internal IMF in the modeled CME, though, as is suggested in Mays et al.
(2015), it is possible to illicit some information utilizing the magnitude of the magnetic field and
looking at a representative selection of IMF clock angles. For the sake of completeness, the ∆|B|
IDA distributions are presented in plots (e) and (f), with maximum ∆|B| peaking at a central
CME origin of (−5◦ , 10◦ ) and a value of 7.8 nT. The distribution is slightly compressed to the
left, with the maximum residing away from the Sun-Earth Line. The minimum values distribution
presents itself as kidney-shaped with a large stable region near the center, with the minimum value
at central origin (0◦ , 10◦ ) with a value of -2.3 nT. The steepest gradients, and the CME origins
of greatest uncertainty appear to be toward the top and left of the plot, though considerable vari-
41
ability presents itself in other regions. The magnitude of change in the negative direction is about
one-third of the magnitude of disturbance in the positive direction.
2.3.4.2 Quartile Time-series Plots for the 06 October 2013 Event
To quantify the effects of an event, it is also necessary to examine the trends of the impact
domain over time. Figure 2.8 provides a quartile breakdown for the same three solar wind proper-
ties as with the IDPs. Plots on the left-hand side of the figure are rendered with the background
solar wind conditions intact. Plots on the right-hand side have the background conditions removed,
isolating the behavior of the disturbance. Each plot distinctly outlines the envelopes of the distur-
bance, with the min/max envelope depicted with the dashed red lines. The 25%/75% envelope is
outlined with the blue dash-dot lines. The median is provided with a solid black line. Plots (a) and
(b) provide number density information (N/cm3 ); plots (c) and (d) portray radial velocity activity
(km/s), and plots (e) and (f) depict the total magnetic field (|B|, nT).
Looking at the density envelope plots, (a) is most useful when comparing the envelopes to
observations of an event, or when determining trends of the full solar wind. With the background
conditions removed, as in plot (b), trends of changes due to the CME are more apparent. The
density pile-up region is well defined, and ranges from no pile-up in the minimum case to a pileup
of 20 N/cm3 in the maximum case. The pileup begins as early as 40 hours prior to CME impact,
with the median pileup beginning roughly 10 to 15 hours prior to impact. Looking at plot (b), we
can identify the rarefaction region from the CME passage easily as the trends drop below zero (0).
It is interesting to note that nearly 75% of all impacts converge toward zero between 15 and 20 hours
after initial impact. The earliest rarefaction trends start as early as impact, and 25% of impacts see
little to no rarefaction. The solar wind radial velocity, depicted in plots (c) and (d) indicate that
25% of starting origins produce impacts between a small loss in velocity and small gain in velocity.
The median ∆ velocity peaks around 5 km/s, leaving around 90% of the velocity variation to 50%
of the starting origins. Within the constraints of the available magnetic field information, |B| field
trends, shown in plots (e) and (f), are very similar to the Density trends. There exists a build-up
42
Density Envelope Δ Density Envelope

30
Min/Max Envelope 20
1st/3rd Quartile Envelope
Median
25
15
Plasma Density (N/cm3)
Plasma Density (N/cm3)

20 10
15 5
0
10
−5
5
−10
0
−40 −20 0 20 40 60 80 −40 −20 0 20 40 60 80
Time Since Impact (Hours) Time Since Impact (Hours)
(a) (b)
Velocity Envelope Δ Velocity Envelope
420
80
400 70
Plasma Radial Velocity (km/s)
Plasma Radial Velocity (km/s)

60
380
50
360
40
340 30
20
320
10
300
0
−40 −20 0 20 40 60 80 −40 −20 0 20 40 60 80

(c) (d)
|B| Envelope Δ|B| Envelope
8
10
6
Δ|B| field intensity (nT)
|B| field intensity (nT)
8
4
6
2
4
0
2
−2
−40 −20 0 20 40 60 80 −40 −20 0 20 40 60 80

(e) (f)
Figure 2.8: Solar Wind properties (at Earth) quartile envelopes for the 06 October 2013 CME
event: Density (a & b); Radial Velocity (c & d); Total Magnetic Field |B| (e & f). Plots (a, c, e)
provide the information with the background values intact while plots (b, d, f) provide the quartile
information with the background values removed. The red lines represent the min/max envelope;
The blue lines, the 25%/75% envelope; The solid black line represents the median.
43
trend prior to impact, and there is a convergence where 50% of runs approach zero disturbance
of the total magnetic field. The rarefaction trends are also similar, with the earliest starting near
impact, and about 25% seeing none. Without accurate magnetic field information for the CME
ejecta cloud, however, conclusions about |B| field trends must be treated with assumptions to
extract any meaningful information.
2.4 Conclusions
Examining the distribution of possible CME effects utilizing CME Impact Domain Analysis
can provide insight in to how uncertainties in CME origination conditions can affect a space weather
forecast. Through examination of distributions of impact time, impact duration, and changes in
solar wind conditions at Earth, a clear visual can be created to highlight the CME central origin
regions of greatest uncertainty in a forecast. In this study it was shown that the central origin can
have a significant impact on solar wind conditions at Earth, and this, in turn, can have a non-trivial
impact on the geoeffectiveness of the event.
The CME IDA method provides a means to explore the uncertainty in the forecast through
impact domain plots, which visualize the impact domain distribution spatialy, and through time-
dependent quartile plots. These plots can be used as a basis for additional investigations as well:
One could potentially integrate the Newell et al. (2007) function over the impact duration generate
a total flux. Plotting the information on an IDP would provide a distribution of CME “effect
magnitude”.
The quartile plots, taken as a whole, can provide information about the propagation of un-
certainties through the model framework. For the analyses presented here, the total error due to
uncertainties in CME central origins can be bound by the minimum/maximum envelope. Compar-
isons of the quartile plots to observations could provide further insights to the performance of a
given forecast.
Overall, it may take a bit of investigation to illicit full meaning of the distributions presented
in this paper, but they do provide a compelling glimpse as to how the uncertainties in central origin
44
can affect a space weather forecast. Additional work will be necessary to determine if general trends
can be inferred from Impact Domain Plots such as these.
The process of generalizing the CME IDA should begin with extending the method to analyze
the sensitivity of other dimensions in the cone parameter space (i.e. velocity, cone angle). From
there the process of generating a set of baseline analyses for a CME injected into ideal steady-state
solar wind conditions should be undertaken. Building upon the ideal case, a database can be built
by systematically conducting analyses with one type of background solar wind phenomenon at a
time. From this database, a correlation study could explore the possibility of generating an analytic
model to determine how various background conditions affect the ideal solution. At that point, the
study could be extended to explore the possibility of developing statistical models of forecasts based
on the ambient solar wind conditions and the estimated cone parameters. If successful, this would
provide a tool to more accurately bound uncertainties in all forecasts.

Chapter 3
Adiabatic Invariant Methods and Error Analysis in a Dipole
Ordering observed particles in the magnetosphere is important to understanding how pro-
cesses in the radiation belts work. The three adiabatic invariants M , J, and Φ provide and im-
portant part in this ordering. M represent the magnetic flux contained within the gyro-obit of a
particle. J represent the length of travel along a field line that a particle travels. Its companion
invariant, K, is used to define the “drift shell” in which a particle is trapped. Φ is the flux contained
within the drift shell defined by K, and is used to determine the radial distance, L∗ , that a particle
would cross the equator if all of the external components of Equation (1.18) were slowly relaxed to
zero. (Roederer, 1970; Roederer and Zhang, 2014)
This chapter focuses on the first iteration of algorithms designed to calculate the adiabatic
invariants of particles trapped within the Earth’s magnetosphere. These algorithms are designed
to function with non-gridded magnetic field models, and provide a framework for error analysis of
the gridded methods presented in Chapter 4. Section 3.1 reviews the equations for the adiabatic
invariants, and provides the basic methodology necessary for computing the solutions numerically.
Section 3.2 looks at the error responses of the algorithms with respect to the ideal solutions on a
dipole, for the algorithms where analytic solutions exist.
3.1 Invariant Algorithms
This section is broken down into three parts. Each part deals with a single invariant and the
methods necessary for calculation. Even though the first invariant is a simple analytic equation,
46
it is provided here for the sake of completeness in the discussion. Following a description of the
invariant, each subsection will look at the algorithms and methodologies required to produce the
solution on a non-gridded magnetic field model. These methods are utilized in Section 3.2 to
investigate the potential errors involved.
3.1.1 Computing µ
Recall from Chapter 1 that the first adiabatic invariant is given by:
p2⊥ p2 sin2 α
µ = =
2m0 B 2m0 B
It can be seen from this equation that µ can be solved directly, and thus does not require any
complicated algorithms to calculate.
3.1.2 Computing K
Again, recall from Chapter 1 that the second adiabatic invariant is given by J, which is given
by:
I I
J = pk ds = m v µ ds. (3.1)
Roederer and Zhang (2014) states that this integral is taken along the guiding field line for a
H R sm
complete bounce cycle ( = 2 s0m ). With the specific case of equipotential field lines, v is constant
during one bounce. Taking into account the definition of the purely field-geometric quantity I (3.2),
the equation in (3.3) can be derived.
s0m

B(s)
Z
I = 1− ds (3.2)
sm Bm
J = 2pI (3.3)
47
When dealing with a more realistic magnetic field, under the influence of external forces
perpendicular to the magnetic field lines, the integral I will no longer be adiabatically invariant
due to the fact that p is changing. To obtain a value that remains invariant, it is necessary to
combine the invariance of µ (through the magnetic field mirror strength Bm ) with (3.3) to arrive
at the equation in (3.4).
p Z s0m »
K = Bm I = Bm − B(s)ds (3.4)
sm
where Bm is calculated by:
p2 B
Bm = = (3.5)
2m0 µ sin2 α
The algorithms in this section, and the remainder of this work, will utilize Bm and the
invariant K for computational purposes.
3.1.2.1 Methodology for Calculating K
There are three basic steps involved in calculating the second adiabatic invariant. First, the
guiding field line of the particle must be traced. Second, the field line position/value pairs are
re-sampled to aid in the integration of K, and finally K must be integrated between mirror points
(points of equal Bm ).
To satisfy the first requirement, field lines, which are instantaneous streamlines within a
magnetic field, can be calculated by solving the initial value ODE given in Equation (3.6).
dL(u)
L(0) = x0 , = v (L(u)) (3.6)
du
For the prototype implementation, the ODE solver chosen is the Runge-Kutta-Fehlberg
Fehlberg (1969) 4th order method with automatic time-step adjustment in the 5th order. This
method uses the 6 equations in (3.7) to generate 4th and 5th order approximations.
48
k1 = hf (xk )
1
Å ã
k2 = hf xk + k1
4
3 9
Å ã
k3 = hf xk + k1 + k2
32 32
(3.7)
1932 7200 7296
Å ã
k4 = hf xk + k1 − k2 + k3
2197 2197 2197
439 368 845
Å ã
k5 = hf xk + k1 − 8k2 + k3 − k4
216 513 4104
8 3544 1859 11
Å ã
k6 = hf xk − k1 + 2k2 − k3 + k4 − k5
27 2565 4104 40
The actual 4th and 5th order approximations are constructed utilizing the equations below.
The 4th order solution is given in Equation (3.8) and the 5th order solution is given in Equation
(3.9)
25 1408 2197 1
xk+1 = xk + k1 + k3 + k4 − k5 (3.8)
216 2565 4101 5
16 6656 28561 9 2
zk+1 = xk + k1 + k3 + k4 − k5 + k6 (3.9)
135 12825 56430 50 55
The step size can be adjusted every time step to an optimal value (hopt ) through a comparison
of the 4th and 5th order solution to examine the truncation error. Specifying a maximum error
tolerance, τ , allows us to adjust the time-step h. This is given by Equation (3.10) and (3.11).
Ç å1 Ç å1
τh 4 τh 4
s = ≈ 0.84 (3.10)
2|zk+1 − xk+1 | |zk+1 − xk+1 |
hopt = sh (3.11)
The basic approach for this method is presented in Algorithm 1 below.
This approach does a decent job finding the proper path in a dipole, as is shown in the error
analysis in Section 3.2. It is of note that the field line trace should be terminated at the surface of
the Earth (or just below to aid in calculations) for the methods in this section. When dealing with
49
Algorithm 1 Runge-Kutta-Fehlberg Method

1: procedure RK45trace(dir = forward, ib = 0.9, max steps = 2000, x0 )
2: x ← x0
3: rad distx ← magnitude(x0 )
4: inner boundary ← ib
5: path ← empty array
6: value ← empty array
7: step ← 0
8: while rad distx > inner boundary do
9: s←0
10: while s is not close to 1.0 do
11: k1 ← from Equation (3.7)
17: x4 ← from Equation (3.8)
18: x5 ← from Equation (3.9)
19: s ← from Equation (3.10)
20: h ← sh
21: x ← x4
22: valx ← value(x)
23: rad distx ← magnitude(x)
24: step ← step + 1
25: if step > max steps then
26: break
27: append path with x
28: append value with valx
return path, value
50
gridded data, the end point may not always be at a desirable location, as different models utilize
different inner boundaries. This problem will be addressed in Chapter 4 where modifications to
methods for gridded data are presented.
With the field line path established, the second step is to re-sample the trace so that, from
the minimum magnetic field value (Bmin ), both the north and south directions of the field line have
their locations defined by equal B values. In simpler terms, for every B value in the northern trace,
there should be an identical value in the southern trace, with its location dictated by this value. To
accomplish this, the prototype code utilizes a cubic-spline interpolation method. The procedure is
given in pseudo-code in Algorithm 2.
Algorithm 2 Field Line Resample

1: procedure FieldLineResample(traceF orward, traceBackward)
2: northTrace ← traceF orward
3: southTrace ← traceBackward
4: Bmin North ← minimum of northTrace
5: Bmin South ← minimum of southTrace
6: Bmin ← minimum of Bmin North and Bmin South
7: recenter both traces on Bmin
8: for B in northTrace do
9: interpolate new location in southTrace where Bsouth = Bnorth
10: append newSouth with new location/value pair
return northT race, newSouth
Figure 3.1 provides the graphical result of re-sampling on a Tsyganenko T96 magnetic field.
This field is used for reference instead of a dipole, as the application of this process on a dipole
tends to produce symmetric results, making the re-sampling unnecessary.
Once the field line is re-sampled, it is a straightforward task to integrate for K (Equation 3.4).
This is accomplished by integrating from Bmin to Bmirror in both the north and the south direction,
then adding the resultant values together. For the purpose of the prototype implementation, this
integration is done along the field line utilizing simple trapezoidal integration. The procedure for
calculating K is given in Algorithm 3.

51
Figure 3.1: This figure illustrates the results of re-sampling magnetic field lines. The south field
trace is re-sampled to equal B values of the north trace, thus allowing for more straightforward
field line integration to determine K. This re-sampling is of a Tsygenenko T96 magnetic field (with
dipole tilt). The field line being re-sampled is indicated by the blue and red dots along the line.
The red dots signify the primary, and the blue dots signify the locations that were interpolated to
provide the equal B values. The black dots represent the Bmin on each field line. Bifurcated field
lines can be easily identified on the sunward side. The black contour lines are of constant |B|.
3.1.3 Computing L∗
According to Roederer and Zhang (2014), in a dipole, the magnetic flux encompassed by a
particle shell (Φ) can be given by the equation in (3.12).
2πk0 2
2πBE RE 1.953 2
Φ = = = GaussRE (3.12)
r L L
However, in a complex magnetic field, Φ must be numerically calculated from the flux integral
as given in (3.13).
ZZ
Φ = B · dS (3.13)
π
52
Algorithm 3 CalculateK
1: procedure calculateK(traceF orward, traceBackward)
2: K ← None
3: if field line is not bifrucated then
4: if Bmirror is in both field line traces then
5: Kf ← trap int traceF orward from Bmin to Bmirror utilizing Equation (3.4)
6: Kb ← trap int traceBackward from Bmin to Bmirror utilizing Equation (3.4)
7: K ← Kf + Kb
return K
where π is the surface of the polar cap bounded by the drift shell field line intersection of the
Earth’s surface.
Roederer and Zhang (2014) further states that, if the magnetic latitude of the drift trajectory
for the particle is known, Stokes’ Theorem (Katz, 1979) can be utilized to transform the surface
integral in (3.13) to the line integral in Equation in (3.14) to calculate Φ.
Z 2π Z π/2 Z 2π
Φ ∼ 2
= 2BE RE dφ cosλ sinλ dλ = 2
BE RE cos2 [λC (φ)]dφ (3.14)
0 λc (φ) 0
where λC (φ) is the dipole latitude of the intersection C at a given longitude φ.
Utilizing Equation (3.14) as a means of flux integration means very little on an ungridded
dipole, as the solution reduces to the analytic solution for flux. It is, however, very useful in defining
an analytic version of L∗ that is dependent on only λC . The definition of L∗ (Equation 3.15) is
the logical place to begin a derivation of equations for analyzing the numerical methods used to
compute L∗ on a field more complex than a dipole.
2
2πBE RE
L∗ = (3.15)
Φ
To perform a quality error analysis, the analytic form of L∗ , related to the field line intersec-
tion latitude (λC ), is needed. To arrive at this analytic solution, start with substituting Equation
(3.14) for Φ in Equation (3.15) to arrive at Equation (3.16).
2π
L∗ = R 2π (3.16)
0 cos2 [λC (φ)]dφ
53
When dealing with a dipole, the symmetric nature of the magnetic field reduces this to the
analytic form in Equation (3.17).
1
L∗dipole = (3.17)
cos2 [λ C]
With the analytic L∗ (λ) function defined, the surface integration method can be examined.
3.1.3.1 Methodology for Calculating L∗
The first step for calculating L∗ requires the finding of the particle drift trajectory. This relies
on the invariant nature of the second invariant, K, with a basic procedure that involves first tracing
the guiding field line for the particle in question, and calculating K for the given Bm (or vice-versa).
Next, field lines at other local-times around Earth must be found that conform to the same Bm /K
pair. For the prototype system, this is done utilizing a basic bisection method to search for the
field line in question. The procedure in Algorithm 4 describes how this is accomplished.
To find the entire drift trajectory (in a non-dipole field), more than one field line trace
is needed. In a dipole, the symmetric nature of the field means that the entire drift trajectory
can be defined by a single λC , therefore only the initial trace is needed. In more complex fields,
experimentation has shown that a minimum of 4 local times must be calculated (local times equating
to midnight, dawn, noon, and dusk). More complex field geometries will be addressed, in greater
depth, in Chapter 4.
Once the drift trajectory is known, the integration grid for the surface can be prepared,
and the integration performed. Although Stokes’ Theorem eliminates the need for a full surface
integration in many cases, there may still be times when the full surface integration is necessary.
For the initial prototype, both the Stokes’ method and the full surface integration have been
implemented. For the Stokes’ method, the boundary of the surface is integrated using a simple line
integral. For surface integration, the grid is created as a simple subsection of a spherical shell. The
grid is computed as is given in Algorithm 5.
With the grid defined, the integration can be performed. At a very basic level, a surface
54
Algorithm 4 Find New Drift Shell Field Line

1: procedure findNewDriftShellLine(Bm , K, localtime, r0 )
2: low RE ← 1.0
3: start unit ← Surface location at localtime
4: start location ← start unit × r0
5: hi RE ← Radial Distance to Outer Boundary for localtime
6: new RE ← r0
7: Bnl ← N one
8: while Bnl is N one do
9: new line ← field line originating at start location
10: Bnl ← Bm from new line where Knl = K
11: if Bnl is N one then
12: if new RE is close to low RE then
13: return N one
14: new RE ← 0.5 × new RE
15: start location ← start unit × new RE
16: while Bnl is not close to Bm do
17: step ← Optimal Step Fraction ()or 0.5 for bisection
18: if (Bnl < Bm ) or (Bnl is N one) then
19: hi RE ← new RE
20: new RE ← low RE + step × (hi RE − low RE )
21: else
22: low RE ← new RE
23: new RE ← low RE + step × (hi RE − low RE )
24: start location ← start unit × new RE
25: new line ← recompute field line with new start location
27: if Bnl is N one then
28: if new RE < r0 then
30: redux ← 0.5 × (hi RE − new RE )
31: new RE ← new RE + redux
34: else
36: redux ← 0.5 × (new RE − low RE )
37: new RE ← new RE − redux
return new line
55
Algorithm 5 Defining the Integration Grid
1: procedure definePolarGrid(RE , num divisions, base lat fun)
2: lon points ← liniear space from 0 to 360 with num divisions without endpoint)
3: grid ← num divisions × num divisions array
4: for lon in lon points do
5: lat points ← liniear space from base lat fun(lon) to 90 with num divisions
6: for lat in lat points do
7: λ ← lat in radians
8: φ ← lon in radians
9: grid[lat, lon] ← (RE , λ, φ)
return grid
integral can be broken down into two parts: the area of a section, and the value of that section.
More explicitly, a surface integral looks like:
ZZ
= f (πcell ) a(πcell )
πcell
where f (πcell ) represents the cell value, and a(πcell ) the cell area.
The algorithm designed to calculate Φ takes advantage of this, allowing for testing of multiple
different area and value functions. Algorithm 6 provides this procedure, and Algorithm 7 illustrates
an area algorithm utilizing two triangles to calculate the area of a quadrilateral. A value function
needs to have the ability, utilizing the same points arguments as the area algorithm, to return
an optimal value for the given cell. Algorithm 8 provides an example that produces a cell value
utilizing the average of the given point values.
The final step is to calculate L∗ itself. The procedure presented in Algorithm 9 combines the
other two algorithms to generate a final L∗ value. The inputs to this algorithm are the location of
the particle and its pitch angle, α.
3.2 Error Response in a Dipole Field
The first invariant is an analytic calculation, and requires no analysis of errors. The second
and third invariants, however, require numerical integration and thus should be analyzed to show
their responses to error. Unfortunately, there is no analytic solution to the second invariant. The
56
Algorithm 6 Integrate Surface

1: procedure integrateSurface(grid, valFun, areaFun)
2: Φ ← 0
3: lons ← φ in grid
4: lats ← λ in grid
5: for indexφ in range of lons do
6: if indexlφ + 1 is in range of lons then
7: i1 ← indexφ
8: i2 ← i1 + 1
9: else
10: i1 ← indexφ
11: i2 ← 0
12: for indexλ in range of lats do
13: ~ 11 ← grid[lats[i1 ], lons[i1 ]]
pt
14: ~ 12 ← grid[lats[i1 ], lons[i2 ]]
pt
15: ~ 21 ← grid[lats[i2 ], lons[i1 ]]
pt
16: ~ 22 ← grid[lats[i2 ], lons[i2 ]]
pt
17: Φ ← Φ + areaFun(pt ~ 11 , pt
~ 12 , pt
~ 21 , pt
~ 22 ) × valFun(pt
~ 11 , pt
~ 12 , pt
~ 21 , pt
~ 22 )
return Φ
Algorithm 7 Area Function using Triangles

1: procedure areaFun(pt ~ 11 , pt
~ 12 , pt
~ 21 , pt
~ 22 )
2: ~
λ11 ← λ(pt11 )
3: λ12 ← λ(pt ~ 12 )
4: λ21 ← λ(pt ~ 21 )
5: λ22 ← λ(pt ~ 22 )
6: φ11 ← φ(pt ~ 11 )
7: φ12 ← φ(pt ~ 12 )
8: φ21 ← φ(pt ~ 21 )
9: φ22 ← φ(pt ~ 22 )
10: r ← r(pt11 )
11: dλ1 ← λ12 − λ11
12: dλ2 ← λ22 − λ21
13: if φ21 < φ11 then
14: dφ1 ← (φ21 + 2π) − φ11
15: else
16: dφ1 ← φ21 − φ11
17: if φ22 < φ12 then
18: dφ2 ← (φ22 + 2π) − φ12
19: else
20: dφ2 ← φ22 − φ12
area1 ← 21 r2 cosλ11 dλ1 dφ1

21:
22: area2 ← 12 r2 cosλ12 dλ2 dφ2
return area1 + area2
57
~ · n̂
Algorithm 8 Value Function using and Average of Points for B
1: procedure valueFun(pt ~ 11 , pt
~ 12 , pt
~ 21 , pt
~ 22 )
2: B~ 11 ← value(pt ~ 11 )
3: B~ 12 ← value(pt ~ 12 )
4: B~ 21 ← value(pt ~ 21 )
5: ~
B22 ← value(pt ~ 22 )
6: ~
n̂11 ← pt11 / magnitude(pt ~ 11 )
7: ~ 12 / magnitude(pt
n̂12 ← pt ~ 12 )
8: ~
n̂21 ← pt21 / magnitude(pt ~ 21 )
9: ~ 22 / magnitude(pt
n̂22 ← pt ~ 22 )
10: ~
dot11 ← B11 · n̂11
11: dot12 ← B ~ 12 · n̂12
12: dot21 ← B ~ 21 · n̂21
13: dot22 ← B ~ 22 · n̂22
14: value ← (dot11 + dot12 + dot21 + dot22 )/4
return value
Algorithm 9 Calculate L∗
1: procedure calculateLstar(Φ, BE = 3.15 × 104 , RE = 1.0)
2π B 2
RE
2: L∗ ← Φ
E
return L∗
58
best that can be hoped for with any analysis of the second invariant involves determining the error
associated with searching for field lines that maintain the invariant, plus the error response of the
third invariant (which we can solve analytically). In the analysis of the third invariant, the fact
that it can be solved analytically on a dipole can be used to generate profiles for error response
and sensitivity. The remainder of this section demonstrates the error responses associated with the
individual methods needed in solving the third invariant.
Calculation of the third invariant involves several steps, each of which can be analyzed for
their error response. Using the method found in Roederer and Zhang (2014), the invariant nature
of the second invariant (K) can be used to iteratively find the drift trajectory for a particle in the
magnetosphere.
3.2.1 Error Response due to Field Line Footprint Inaccuracies
The Roederer and Zhang (2014) method requires that a particle’s bounce trajectory be traced
to calculate K, and the footprint of the field line must be determined (the footprint will hereafter be
referred to as the λ intersection, or simply λ). With K and its associated Bmirror value, the space
around Earth is systematically searched to find other field lines conforming to these now-known
values. Once a sufficient number of field lines are located, the drift trajectory can be calculated
and the flux through the drift shell (Φ) can be calculated. With Φ calculated, L∗ can be easily
calculated utilizing Equation (3.15). In a dipole, as is being examined here, the simplified version
of the L∗ computation given in Equation 3.17 can be used instead.
To fully understand how errors propagate in this system, it is necessary to look at each
individual piece. First, it is important to note that for a dipole, field lines can be computed
analytically with the following equation from Walt (1994):
r = R0 cos2 λ (3.18)
where λ is the latitude of the field line, and r is radial distance. R0 is the equatorial crossing radial
distance (where λ = 0)
59
Given this relationship, the surface λ intersection point can be obtained. By fixing r = 1RE
and utilizing the relationship in Equation (3.18), Equation (3.19) can be derived. Figure 3.2 shows
this analytic relationship between L∗ and λ on a Dipole.
Å ã!
−1 RE
λ = cos (3.19)
R0
Analytic L * vs. Intersection λ Profile

on a Dipole
70
60
50
Intersection λ
40
30
20
10
2 4 6 8 10 12 14
Analytic L *
Figure 3.2: Analytic Dipole L∗ vs. Intersection λ
Knowledge of this profile will allow for future analysis of the accuracy of field line tracing on
dipole gridded data (more in Chapter 4).
Given a starting K and Bm , to find field lines on additional local times, bisect over field
lines until one is found where Kn = K0 for Bmn = Bm0 . This search is conducted at intervals of
longitude until a drift trajectory is defined. This provides an integration boundary for calculating
magnetic flux (Φ). If K and Bmirror are consistent over all field lines, all iteratively discovered field
lines in the bounding set will fall on the λ given by Equation (3.19).
To better understand the potential errors involved in this method, relying on the analytic
solutions to λ intersections as a guide (see Figure 3.2 and Equation (3.19)), the error associated
60
with tracing a field line through a magnetic field can be examined. Figure 3.3 shows this profile for
a range of L∗ values. The traces in this profile are conducted utilizing a Runge-Kutta 4/5 method
with a step error tolerance of 1E −5 . The dashed line indicates the absolute error between the traced
λ intersection and the analytic solution.
λ Error vs. L * on a dipole

RK4/5 Error Tol: 1e-05
Analytic λ Traced λ Absolute Error
70
65
10−4
λ of Field Line Intersection
60
λ Error
55
50
45
40
35
2 4 6 8 10
L * for Dipole
Figure 3.3: Error Profile for obtaining λ from a field line trace vs. the analytic solution.
Understanding the error in the field line trace provides a means of determining how the field
line trace λ affects the ideal solution for L∗ in general. Noting that the analytic solution for L∗ on
a dipole is given by Equation (3.17), Equation (3.20) shows how the error in L∗ due to the field
line trace (Etrace ) can be calculated.
Etrace = L∗ (λanalytic ) − L∗ (λtrace ) (3.20)
where L∗ (λanalytic ) is the analytic solution for L∗ utilizing the analytic intersection for the field
line, and L∗ (λtrace ) is the analytic solution for L∗ utilizing the field line traced solution (or, in this
case, a perturbed variation of the base λ). The profile for Etrace is given in Figure 3.4. The green
Etrace
line shows the value of Etrace for a given λerror , and the blue line shows the ratio λerror (the Error
61
Effect Ratio), which indicates the percentage of error from λ is seen in the final L∗ calculation. For
this plot (L∗ = 6.6RE ), it is obvious that the final L∗ calculation sees about half of the error seen
in λ.
Error in L * due to Error in Field Line Tracing

(Etrace) on a dipole @ L * = 6.6 RE
Etrace Error Effect Ratio
10−1
1.4
−3
10
1.2
10−5
1.0
Error Effect Ratio

10−7 0.8
Etrace
10−9 0.6
Median: 0.5452
10−11 0.4
10−13 0.2
0.0
10−13 10−11 10−9 10−7 10−5 10−3 10−1
Error in λ
Figure 3.4: Profile for the error in L∗ associated with the field line trace. The blue line indicates
the ratio E ∗
λerror for the this L value (6.6RE ). It can be seen that for an error in λ, approximately
trace
54.5% is passed to the overall L∗ calculation. The instability seen on the left side of plot is due to
floating point truncation error.
This error, however, is dependent on the starting λ (and the resulting desired L∗ ). Figure
3.5 shows the relationship of the Error Effect Ratio compared to the base λ solution. This provides
insight on the sensitivity of the L∗ calculation to errors in λ. It can be seen in this plot that, as
the base value of λ increases, so does the effect of error on L∗ .
This makes sense, as magnetic field strength gradients are much larger near the Earth’s poles.
As λ increase (gets closer to the pole), so does the resulting L∗ value. Thus, as the dipole equitorial
radial distance increases (larger L∗ ), the closer λ gets to the pole, the faster the magnetic field
changes, and the greater the effect due to any error in λ.

62
Sensitivity of L * Calculations to Errors in λ

in a Dipole
104
103
102
Error Effect Ratio
101
100
10−1
10−2
10−3
0 20 40 60 80
Base λ
Figure 3.5: Profile of the median Error Effect Ratio for a range of base λ values.
Chapter 4
Working with LFM Model Gridded Output
4.1 The LFM Grid
Lyon et al. (2004) provides a succinct summary of the Lyon-Fedder-Mobarry (LFM) simu-
lation grid. It is described as a non-orthogonal, non-adaptive, distorted spherical grid statically
adapted to the problem. Within the grid, the r and θ coordinates are fitted to the problem. The
coordinates are derived by minimizing the weighted smoothness integral:
(∇µ)2 + (∇v)2
Z
I = dS , (4.1)
w
Cell distribution is controlled by the weight (w), causing more cells to cluster wherever w is large,
and the reverse when w is small. Weight is concentrated near the magnetopause, bowshock and
tail; regions where it is expected that larger variations in field geometries will exist. The values for
w were calculated by hand. The φ grid coordinate is not adapted, although it is stretched in such
a way that larger cells fall near the equatorial plane, with slightly smaller cells near the poles. The
grid space covers a region around Earth that encompasses +30RE to −300RE along the x axis, and
a typical cylindrical radius of 100RE . The system is oriented in the solar magnetic (SM ) coordinate
system where the +Z axis aligned to the north magnetic pole, the Sun-Earth line is contained in
the xz plane, and the y axis is perpendicular to the other two (Laundal and Richmond, 2017). The
LFM grid has a nominal inner boundary of 2.3RE .
Figure 4.1 provides a side and front view of the LFM grid. In the left hand image, the X axis
points to the top. The right hand view shows the grid looking toward the +X face. The cylindrical
64
r
α
Figure 4.1: This image provides side and front views of the LFM grid. The graphics are annotated
to illustrate the basic logical arrangement of the simulation grid.
symmetry of the grid is clearly illustrated with radial components arranged in a fan line pattern
around the central x-axis (notated by α angle in the figure). Images in Figure 4.2 provide two
views of an individual grid blade. The top figure provides an overall overview of a blade, clearly
showing the weighting of the grid cells. The bottom portion of the figure provides a zoomed view
of the inner boundary. This view illustrates the distorted spherical nature of the grid.
Online documentation indicates that the grid can be configured to arbitrary dimensions
(https://wiki.ucar.edu/display/LTR/MSETUP), but the most common resolutions used are given
in Table 4.1. Also included in this table is a reference to the number of coordinate vertices in the
grid. This number will become important in future sections when dealing with calculating the K
and L invariants.
4.2 LFM Grid Dipole Response
Before delving into the specifics of dealing with MHD data on an LFM grid, it is important
to understand how the grid itself affects the outcome of the K and L∗ algorithms. To this end,
65
Figure 4.2: Image showing the detail of one planer grid “blade”. These blades radiate from the x
axis to form the full 3-dimensional grid. The upper image shows an overview of an entire blade;
the bottom image shows a close up of the area immediately around Earth, illustrating the heavier
cell weighting of the grid.
the following dipole analysis, similar to that conducted for a non-gridded dipole in Chapter 3, is
presented, along with interpretations of potential impacts to the overall K and L∗ solutions.
Recall from Chapter 3, Figure 3.3, that without a grid, tracing a field line through a dipole
provided an end-point (footprint) accuracy at the surface of Earth on the order of 10−5 degrees.
When a dipole is placed within an LFM grid, the controlling factor for accuracy becomes the
resolution of the grid. Figure 4.3 provides, instead of footprint error, the relative L∗ error responses
for three different grid sizes: Single, Double, and Quad (see Table 4.1 for more grid resolution
detail). For an analysis of field lines along the x axis, the resolution exhibiting the most erratic
error response is, predictably, the single resolution grid. At low L values, L∗ accuracy is in the
10−2 − 10−3 range. As the radial distance increases, the accuracy improves, peaking at between
7.5 and 8.0 before falling back. The double precision grid is more stable in its error response,
and performs, overall, about the same to slightly better than the single precision grid. The quad
resolution grid response is well within the 10−4 range, though it fluctuates more than the double
66
Name Resolution Number of Vertices
Single 53 x 24 x 32 40,704
Double 53 x 48 x 64 162,816
True Double 106 x 48 x 64 325,632
Quad 106 x 96 x 128 1,302,528
Table 4.1: List of resolutions for common LFM grid sizes and their corresponding number of
coordinate vertices.
resolution grid.
L * Error Comparisons - Dipole on LFM Grids
10−3
Relative L * Error
10−4
LFM Single
LFM Double
10−5 LFM Quad
3 4 5 6 7 8 9 10
Ideal L *
Figure 4.3: Profile for L∗ error by dipole L∗ generated down the +x axis.
This view along the x-axis is insufficient, however, to understand the complete behavior of the
LFM grid. To gain a more complete understanding, it is necessary to look at sample field lines in
all directions, as the x axis has a unique geometry within the grids. To simplify this understanding,
Figures 4.4 to 4.10 provide a statistical overview of the total and relative L∗ errors associated
with tracing field lines within the different grid resolutions. Starting points for field line traces
were constructed by randomly selecting starting radial distances (between 3.0RE and 10RE ) and
starting λ (from values between − π4 and π

4 radians). These selections were then combined with 16
evenly spaced values of φ (covering the full range from 0 to 2π radians). Figures 4.4, 4.6, and 4.8
67
display the total L∗ error associated with every point in the sample set, ordered by dipole L∗ . The
three different resolution distributions were computed with the same set of sample starting points,
to provide for easier comparison between resolutions. In all, the sample set contains 2400 unique
staring points.
To determine the error, each point was used as an origin to produce a field line trace. From
this line, the footprint λc was determined and Equation (4.2) used to calculate L∗ based on the
results of the trace (L∗t ). The ideal L∗ for the starting location (L∗ideal ) was then calculated using
the same equation utilizing the starting location to supply R and λ.
R
L∗ = (4.2)
cos2 (λ)
The total error was then computed with:
ET = |L∗ideal − L∗t |. (4.3)
Relative Errors were calculated by:

|L∗ideal − L∗t |
ER = (4.4)
L∗ideal
Once the errors were calculated, the data was sorted by φ to allow for examination of error
distributions by local time. Figures 4.5 (single resolution), 4.7 (double resolution), and 4.9 (quad
resolution) provide the local-time breakdown for relative L∗ errors.
Single resolution grids show the most variation in error throughout the grid (Figure 4.5). The
variation becomes less pronounced as the grid resolution increases. The median error distribution
for the single resolution grid sees the highest errors (on the order of 10−2 ) near noon (0◦ ), with
errors decreasing as field lines move toward midnight (180◦ ) .
Double resolution grids see less variation when compared to single resolution, while still
exhibiting the highest errors on the noon side. The quad resolution grid shows similar variances to
the double resolution grid, but with an overall lower median error rate. The noon-side error also
presents as high in the quad resolution grid.

68
Dipole L * Error Sample Set by dipole L * for Single Resolution LFM Grid
100
10−1
Total L * Error
10−2
10−3
10−4
10−5
4 6 8 10 12 14
Dipole L *
Figure 4.4: This figure provides a view of the dipole-L vs total L∗ error for 2400 random samples
on a single resolution LFM grid.
Relative Dipole L * Error distribution by ϕ

For Single Resolution LFM Grid
10−2
Relative L * Error
10−3
10−4
10−5
112.5
135.0
157.5
180.0
202.5
225.0
247.5
270.0
292.5
315.0
337.5
22.5
45.0
67.5
90.0
0.0
Sample Direction (∘)
Figure 4.5: This plot provides a statistical view of dipole L∗ errors within a single resolution LFM
grid. Note the cyclical nature of the quartiles. The single resolution grid has the most variance
in error. The sample consists of 2400 particle locations assembled from 150 randomly selected λ
and r values combined with 16 evenly spaced values for φ. 0◦ represents noon. Angles increase
counter-clockwise looking down at the north pole.
69
Dipole L * Error Sample Set by diople L * for Double Resolution LFM Grid
100
10−1
Total L * Error
10−2
10−3
10−4
10−5
4 6 8 10 12 14
Dipole L *
Figure 4.6: This figure provides a view of the dipole-L vs L∗ error rate for 2400 random samples
on a double resolution LFM grid.

For Double Resolution LFM Grid
10−2
Relative L * Error
10−3
10−4
10−5
112.5
135.0
157.5
180.0
202.5
225.0
247.5
270.0
292.5
315.0
337.5
22.5
45.0
67.5
90.0
0.0
Figure 4.7: This plot provides a statistical view of relative dipole L∗ errors within a double resolution
LFM grid. 0◦ represents noon. Angles increase counter-clockwise.
Figure 4.10 breaks down the distributions by overall quartiles for each resolution. It can be
seen that (as expected), the median error decreases as the resolution increases, giving confidence
70
Dipole L * Error Sample Set by diople L * for Quad Resolution LFM Grid
100
10−1
Total L * Error
10−2
10−3
10−4
10−5
4 6 8 10 12 14
Dipole L *
Figure 4.8: This figure provides a view of the dipole-L vs L∗ error rate for 2400 random samples
on a quad resolution LFM grid.

For Quad Resolution LFM Grid
10−2
Relative L * Error
10−3
10−4
10−5
112.5
135.0
157.5
180.0
202.5
225.0
247.5
270.0
292.5
315.0
337.5
22.5
45.0
67.5
90.0
0.0
Figure 4.9: This plot provides a statistical view of dipole L∗ Errors within a quad resolution LFM
grid.
that an increased grid resolution will produce a more accurate result when calculating L∗ with
MHD simulation results.

71
Relative Dipole L * Error Distribution
For Three Different LFM Grid Resolution
10−1
10−2
L * Error
10−3
10−4
10−5
Single Double Quad
Figure 4.10: This figure provides and overview of the overall error distribution for the three grid
resolutions selected. It is clear that the grid resolution has a direct, if small, impact on the accuracy
of the field line traces. This can be directly translated into an effect on the overall L∗ calculations.
The final factor to consider is the overall propagation of errors (due to grid resolution) to the
final L∗ calculation. Figure 4.11 provides the error propagation rate from the grid based on dipole L
(i.e. the λc point of a field line). The procedure for this was first outlined in Chapter 3 utilizing an
un-gridded dipole field. Unlike Figure 3.5, which ordered data by the 1RE λc intersection, Figure
4.11 orders the data by dipole L∗ , making it valid for any radial-distance crossing point. This
change was made due to the termination of the LFM grid at around 2.3RE .
It can be seen from Figure 4.11 that the median error propagation rates are nearly identical
between grid resolutions. From this, it can be concluded that, regardless of the grid resolution, the
LFM grid geometry produces an error amplification for field line footprint errors up to nearly 50×,
depending on the traced L∗ magnitude.
4.3 Algorithm Adjustments for LFM Data
Moving from a dipole to a more complex gridded magnetic field introduces several challenges
to successfully computing L∗ for any given particle. Initial test runs utilizing LFM data revealed
72
Median Error Effect Ratio vs. Dipole L
for Dipole on LFM Single Resolution Grid
Single Res LFM
50
Double Res LFM
Quad Res LFM
40
Error Effect Ratio
30
20
10
0
4 6 8 10 12 14
Dipole L *
Figure 4.11: This image shows the Error Effect Ratios for all three grid sizes based. This confirms
that the Error transmission from from the grid is essentially the same when moving between grid
resolutions.
several issues that had to be addressed in order to begin computing satisfactory K and L∗ values.
Based on the initial testing, it was discovered that the K calculation algorithm was not robust
enough to consistently and reliably calculate K along a field line. Following intense investigation,
it was determined that two primary issues were responsible for failure of the algorithm. First, the
algorithm was making assumptions about magnetic field lines that did not hold in a complex field.
Second, numerical noise in the MHD solution of the field causes fluctuations in some field lines that
make identifying the correct K/B pair on a field line very difficult. In addition to issues with the
K algorithm, problems were also identified in the method used to locate additional field lines with
the same K/B pair. In complex field geometries, specifically in areas where field lines are swept
back toward the tail region, a linear bisection search becomes invalid. These issues are discussed in
greater detail in the following sections.

73
4.3.1 The New K/B Model
The K/B relationship within a field line is one of the most important aspects to calculating
L∗ , as the field line footprints’ trajectory dictates the outcome of the magnetic flux integral used
in calculating L∗ . Failure to calculate a proper K value for a given Bm will result in the field line
search routines finding incorrect lines, resulting in complete failure of the algorithm. Originally, the
algorithm used to calculate K had a critical flaw; it made the assumption that a field line would
be symmetric about the minimum |B| point on any given field line. The distribution of K across
the line was thus assumed to be symmetric. The process of re-sampling a line to calculate K forced
this assumption on the calculations. In complex field lines, symmetry of K does not necessarily
exist. Testing quickly revealed that a different method for calculating K would be required.
The method eventually developed fell back on the physical interpretation of K. Recall from
Equation (3.4), that K is calculated with the integral:
Z s0m »
K = Bm − B(s) ds
sm
where sm is the first mirror point for a particle moving along the field line, s0m represents the
opposite mirror point along the line, Bm is the |B| of the point corresponding to where the particle
mirrors, and B(s) represents the value of |B| at points between sm and s0m along the field line.
Figure 4.12 illustrates how K is calculated. First, K = 0 will result wherever there is a local
minimum, as there is no point for the particle to mirror on that does not first have to go through a
higher intensity magnetic field. The blue tick, located at Bm = 61.3 in the figure, is an example of
one such location. On a dipole, only one such location can exist for any given line. Continuing with
the example, K for Bm = 500 (red ticks) would be calculated by integrating Equation (3.4) where
the two marks represent sm and s‘m , K for Bm = 2500 would similarly be calculated between the
gray marks, and so on. It can be seen from this illustration that an ideal dipole field will generate
a K profile that is symmetric, and this is indeed the result, as is shown in Figure 4.13. It is clear
from this figure that both the north (left side) and south (right side) hemisphere K calculations
74
are symmetric.
K Calculation on a Dipole
L=8.0
103
102
K/B Value
101
100
B
K
10−1 Bmin = 85.6
Bm = 500
Bm = 2500
10−2
Bm = 10000
Bm = 30000
0 50 100 150 200

Trace Index
Figure 4.12: This figure illustrates how K would be calculated for an ideal dipole field line. Note
that, on a dipole line, all values of Bm are symmetric between the north (left) and south (right)
portions of the field line.
K/B Profile on a non-gridded Dipole

L=8.0
108
B
K
106
104
K/B Value
102
100
10−2
0 200 400 600 800 1000
Trace Index
Figure 4.13: This image depicts an ideal K/B model, based on a dipole field line through L=8.0.
Note that the model is symmetric about Bm .
75
The first iteration K algorithm broke down because it assumed this symmetry. In reality,
the Earth’s dipole tilt, coupled with the interaction of the solar wind, induces non-symmetric field
geometry, leading to a non-symmetric K model. Figure 4.14 provides an illustration us such a
situation. These asymmetries ultimately lead to an invalid model of K when utilizing the original
algorithm, which, in turn, results in poor K/B matching in the search algorithm.
Note that, in this figure, the primary |B| minimum is located off-center, and that the field
line exhibits multiple K = 0 zones due to several local minimums. These additional K = 0 regions
may be either a valid K = 0 region, or, more likely, the result of numerical noise in the field line
trace. Some of this noise, but not all, can be dealt with, and that will be addressed in a future
section.
K/B Profile for LFM MHD data on an LFM grid

L=8.0
103
102
K/B Value
101
100
10−1
10−2
B
K
0 50 100 150 200

Trace Index
Figure 4.14: This image depicts a non-symmetric field line and its associated K profile. Note the
multiple areas where K = 0.
Adapting to the realities of complex field geometry required a complete re-write of the K
algorithm. To ensure that all points on the K profile are properly calculated, the algorithm operates
in two passes. The first pass calculates K for every known |B| value on the field line, moving from
north to south. The second pass performs the same operation moving from south to north. The two
76
K profiles are then joined. In most locations, a non-zero value is joined with a zero value, resulting
in a non-zero value for that point (a location with a mirror point behind the trace progress will
always yield a zero K), or a zero value with a zero value, which will yield K = 0 for that point.
In some instances, where there is a local maximum in the field geometry, the algorithm may yield
two slightly different values for K. In this instance, the final result is the maximum of the two.
Pseudo-code presented in Algorithms 10 and 11 walks through this process.
Algorithm 10 Calculate KB Model

1: procedure KB Model(T race Data, T race Loc)
2: Kf , Bf ← K F orward(T race Data, T race Loc)
3: Kb , Bb ← K F orward(reverse(T race Data), reverse(T race Loc))
4: K ← max(Kf , reverse(Kb ))
5: B ← Bf return [K, B]
Algorithm 11 Calculate K Forward Through Array

1: procedure K Forward(T race Data, T race Loc)
2: Kf ← Array
3: Bf ← Array
4: IDX ← 1st Element
5: while IDX < len(T race Data) do
6: Bm ← T race Data[IDX]
7: Sm ← T race Loc[IDX]
8: 0
SIDX2 ← first index where T race Data[IDX : EN D] > Bm
9: 0 0
SIDX1 ← SIDX2 −1
10: 0
Sm ← Interpolated 0
new point between SIDX1 0
, SIDX2
R s0 »
11: Ks ← smm Bm − B(s)ds
12: Kf [IDX] ← Ks
13: Bf [IDX] ← Bm
14: IDX ← IDX + 1
return [K, B]
4.3.2 Numerical Noise
Figure 4.14 also exhibits characteristics of the second major problem that needed to be
addressed: numerical noise in the MHD model output. In this case, numerical noise is defined as
variations in areas where the grid-size is insufficiently small to produce a smooth result, or where
interpolation errors produce random fluctuations. The apparently random breaks in the Figure
77
4.14 profile, where the value of K intermittently drops to zero along an otherwise unbroken curve,
are evidence of this issue. Figure 4.15 provides a more dramatic example of this phenomenon.
Matching K to B properly in this environment is extremely difficult.
K/B Model with LFM Data (Quad Res)

Line through [8.75, 0.0, 0.0] (w/out smoothing)
4
10
103
102
101
100
K
10−1
10−2
10−3 K
B
10−4
0.0 0.2 0.4 0.6 0.8 1.0
Relative Position in Trace
Figure 4.15: This figure illustrates what happens to K when a particle is in a bifurcation region
within a field line that exhibits an excessive amount of noise. It is exceedingly difficult to match a
B value to its correct K value within this configuration.
An issue can be confirmed as a noise problem in MHD output by progressively increasing
the grid resolution, and noting if the fluctuations dissipate. For the example presented in Figure
4.17, this was accomplished through inspection of the same line in three different grid resolutions;
single, double, and quad resolution. Numerical noise due to the MHD grid size can be seen in
the progression. The blue line, representing the single resolution output, displays high-frequency
oscillations through the bifurcation region. The black line, representing the double resolution
output, also shows these oscillations, but at a lower frequency. Moving to a quad-resolution grid (red
trace) shows that the oscillations virtually disappear, thereby confirming the presence of numerical
noise. While the only means of removing the numerical oscillation at this scale is to increase the
MHD grid resolution, local jitter can be smoothed through the filtering of high frequency noise in
78
the field line. This local jitter is due to errors associated with interpolating between grid points
and can induce a large number of false K = 0 regions.

Line through [8.75, 0.0, 0.0] (w/o smoothing)
Focus on Oscillations
102
100
95
101
90
100 85
|B|
80
K
10−1
75
70
10−2
65
K
10−3 B
60
0.2 0.3 0.4 0.5 0.6 0.7 0.8
Figure 4.16: This figure provides a zoomed view of |B| with its associated K model from Figure
4.15. This view clearly shows where the numerical jitter produces unwanted K = 0 regions.
Figures 4.15 and 4.16 illustrate the effects of local jitter. Figure 4.15 shows a wide angle view
of the relationship between Bm and K. The compressed field line exhibits what, at first glance,
appears to be a fairly uniform and smooth line. However, small fluctuations about the line have
caused several false K = 0 regions to appear. This detail is revealed in Figure 4.16, with the
magnetic field trace expanded to show the smaller fluctuations.
Figure 4.18 shows the same progression with a Sav-Gol 1-D digital signal filter (Savitzky and
Golay, 1964) applied to the |B| trace. This procedure removes the smaller oscillations without
dramatically changing the original trace, leaving a smoothed version of the original. Figures 4.19
and 4.20 demonstrate this effectiveness on the K/B model calculation when the filter is applied.
The number of K = 0 locations are greatly reduced, leaving only three remaining K = 0 instances.
These regions may either be valid K = 0 zones, or be caused by noise due to the LFM grid
resolution. Looking at Figure 4.20, the K = 0 region that exhibits at Bmin is the only guaranteed
79
Numerical noise in LFM output (w/out smoothing)
Line through [8.75, 0.0, 0.0]
104
103
|B|
102
B - Single Res
B - Double Res
B - Quad Res
1
10
0.0 0.2 0.4 0.6 0.8 1.0
Figure 4.17: This image depicts a common field line over 3 different grid resolutions. These lines
have not been filtered. The blue line depicts the line on a single resolution grid, black on a double
resolution grid, and red, a quad resolution grid. The numerical instability associated with the lower
resolution grids is clearly visible.
Numerical noise in LFM output (with smoothing)

Line through [8.75, 0.0, 0.0]
4
10
103
|B|
102
B - Single Res
B - Double Res
B - Quad Res
1
10
0.0 0.2 0.4 0.6 0.8 1.0
Figure 4.18: This image depicts the same field lines as in Figure 4.17, but with a Sav-Gol filter
(Savitzky and Golay, 1964) applied to the magnetic field trace, resulting in a far smoother line.
This filtering greatly improves the chances of calculating a proper K/B model.
80
true K = 0, while the leftmost K = 0 may be evidence of a bifurcated line, or may simply be an
oscillation caused by noise. The central K = 0 is almost certainly induced by noise due to grid
resolution.

Line through [8.75, 0.0, 0.0] (w/smoothing)
104
103
102
101
B/K
100
10−1
10−2
10−3 K
B
10−4
0.0 0.2 0.4 0.6 0.8 1.0
Figure 4.19: This image shows the results of applying a Sav-Gol filter to a magnetic field trace.
This is the same trace as is depicted in Figure 4.15. The resulting K model is far more reasonable
than that of the unfiltered line.
4.3.3 Extremely Distorted Magnetic Fields and Search Progression
Searching for the correct K/B matched field lines is a particularly difficult task in regions of
extremely distorted field lines. Although L∗ is only defined in regions of stable particle trapping
(Roederer and Zhang, 2014), and any calculated L∗ within the highly distorted areas would likely
not be valid, it may still be necessary to actively move a search progression through these areas. If
the target field line is near the boundary of distorted lines, the search algorithm will likely spend
large quantities of time in these regions.
Searching in highly distorted field geometries causes the standard linear bisection search
algorithm to fail. The extreme artificial situation outlined in Figure 4.21 illustrates this problem.
In this example, an initial particle is specified with a high Bm value (red dot). An initial field
81
Line through [8.75, 0.0, 0.0] (w/smoothing)
Focus on Oscillations
100
K
B
95
101
90
85
100
|B|
80
K
75
10−1
70
10−2 65
60
0.2 0.3 0.4 0.5 0.6 0.7 0.8
Figure 4.20: This image provides a zoomed view of Figure 4.19, illustrating the effectiveness of the
Sav-Gol filter in removing excessive jitter.
line is then traced to calculate K for the particle (solid black line). Finally, the search algorithm
is applied in an attempt to find the specified line. With the standard linear search algorithm,
which was based only on φ, an effort to match the original trace fails. Looking at the dashed black
line (the search path), the reason for this becomes clear; the field geometry is shifting the target
field line away from the desired φ intersection (φc ). As the search progression jumps φ angles, the
entire premise of the bisection search becomes invalid. The search relies on successive field lines
along a path exhibiting predictable decreasing field intensity as the search progresses away from
Earth, and increasing intensity as the search moves in. With the example illustrated in this figure,
these conditions are not maintained. The relative intensity, as compared to the search path, of the
distorted field line is unknown, and thus cannot be reliably used to continue the bisection.
In an effort to alleviate some of this issue, and find a valid line within distorted regions,
an adaptive λ condition was added to the search algorithm. Utilizing λ of the original particle
starting point provides for a considerably better chance of locating the necessary field lines when
particles are physically closer to Earth, as the search tends to bisect along a path that is not as
82
Example of Line Search Failure with Linear λ = 0 Search
K = 671.60, Bm=234.24, ϕ = 90.00∘
X-Axis
−20 −15 −10 −5 0 5
20
15
Search Path 10
LShell Line
r=3.0
Y-Axis
r=5.5
r=8.0
r=10.5
r=13.0 5
Start Point
−5
Z-
Ax
is
Figure 4.21: The figure illustrates a linear search failure when utilizing a fixed φ.. The black line,
labeled “LShell Line” is the line being sought. In this situation, as the search progresses along the
fixed φ, φτ along the search path shifts away from the search φs . The inset image is an angle view
illustrating the 3 dimensional positioning of the associated field lines.
heavily distorted. This provides a greater chance for finding a valid line, and to calculate a valid
L∗ . Figure 4.22 illustrates the results of utilizing this technique for the same situation as described
above. The adaptive φ search algorithm produces a fairly good, though not perfect match for
the original particle line. As long as the K/B match is within tolerance, the found line can be
considered valid, and used for the L∗ calculation.

83
Figure 4.22: Utilizing and adaptive λ search algorithm, a much better line fit is achieved. This
search result illustrates a line that was not found with a simple fixed λ linear search. To obtain a
good L∗ , it is not necessary to find the exact line, but the K and B parameters must be correct.
This method is not ideal, as it will still break down for initial particle positions physically
located farther from Earth, where it is more difficult to avoid searching across distorted lines, or is
too close to the Earth at extreme λ positions where the field density it high. A possible solution
to this problem, not yet attempted, would change the search progression to a bisection across λ at
84
a fixed radial distance from the Earth, relative to the location in question, and in a region where
field lines are not greatly distorted. This approach should lead to searching successive lines where
search conditions are not violated, while avoiding direct interaction with regions of highly distorted
field lines. This approach is being deferred as future research.
4.3.4 Dealing with Multiple K Instances Per Line
Bifurcated field lines, or lines that exhibit multiple local Bmin conditions, provide a challenge
to finding valid field lines for the L∗ calculation. In these areas, due to the definition of K, a
given value of K can appear multiple times along the arc of the line. This phenomenon is clearly
illustrated in Figure 4.18 where any given value of K appears no less than eight times along the
arc of the line.
To calculate a correct L∗ , it is necessary to correctly identify field lines that contain the
correct [K, B] pair. This requires tracing a certain path with the bisection search algorithm. This
is a straightforward search for the best fit on non-bifurcated lines. The dilemma arises, however,
when lines are bifurcated; which [K, B] pair should be used for the search path? As most methods
for calculating L∗ on empirical models ignore the bifurcation region entirely (Min et al., 2013b;
Lejosne, 2014) , the correct answer to this question is not at all clear. As L∗ may not be conserved
in these regions (Min et al., 2013b; Roederer and Zhang, 2014; Ukhorskiy et al., 2017), and the
behavior of K for particles transiting this region is extremely complicated (Öztürk and Wolf, 2007;
McCollough et al., 2012), only L∗ calculated in non-bifurcation regions should be considered valid,
though instantaneous L∗ values will still be generated for these areas. The nature of noise in the
gridded field line traces does makes it difficult to differentiate between field lines that are truly
bifurcated, and those that only appear to be bifurcated because of the noise. Additionally, it may
be necessary for the search to progress through these uncertain regions. Due to this need, the
search algorithm must be able to navigate these areas. The approach taken with the GHOSTpy
algorithm is to follow the path of minimum gap between Bm and Bk , where Bm is the mirror point
being sought, and Bk is the |B| value associated with K on the line. In regions that appear to be
85
bifurcated, this will return multiple Bm values for a given K value. Following the path of minimum
Bgap , where Bgap is defined as
Bgap = |Bk | − |Bm |,
will naturally lead to following a path through a pseudo-bifurcated line that is likely to ignore any
remaining spurious K = 0 regions. Bgap is the primary means of navigation through the model
space, with the algorithm checking this value on every iteration to decide which bisection path to
take. The path can be thought of as “what are the search boundaries set to after each iteration.” If
a gap is negative, the current search location becomes the outer boundary. If this gap is positive, it
becomes the inner boundary. If the incorrect path is chosen the algorithm will terminate in a local
minimum region, and will not converge to the optimal match, thus yielding a potentially invalid line
match. This occurs primarily within the bifurcation region, providing an L∗ value with uncertainty
as to the overall validity of L∗ calculations near these regions.
4.3.5 Final Updated Drift Trajectory Search Algorithm
The drift trajectory search algorithm is the most important component of the overall L∗
calculation, and has undergone multiple and significant changes throughout the research process.
Algorithm 12 provides the final updated search method on which all performance metrics are based.
Some error checking routines have been removed for conciseness and ease of reference.
The algorithm starts by searching around the same radial distance of the original particle,
and at the same λ angle. The area is bisected until a valid bgap is found. If the search reaches the
inner boundary and has only found an outer gap, the boundary is moved in. If the search reaches
the outer boundary and has only found an inner gap, the boundary is moved out. If the search finds
an invalid bgap , and the outer gap is defined, the inner boundary is moved out. If bgap is invalid,
and both inner and outer bgap are unset, the outer boundary is moved in. When valid negative gaps
are obtained, the outer boundary is moved in; the inner boundary is moved out for positive gaps.
The search terminates when the line is found (in reality, when it is within the error tolerance), or
86
when the gap between lines reaches . As noted earlier, the algorithm uses the min(bgap ) path to
advance the algorithm when more than one bgap (k) is present.
4.3.6 Accuracy and Convergence
Accuracy of the final L∗ calculation is directly related to how well the search algorithm finds
the integration path for the flux calculation. The earlier part of this chapter focuses heavily on
how errors propagate through the system, but does not provide a great deal of practical insight
on accuracy within complex fields. As seen in the previous few sections, the search algorithm is
not 100% accurate in finding field line footprints. The more accurate these footprints are, the
more accurate the magnetic flux calculation (this can clearly be seen in a simple walk-through of
Equation (3.16)), and thus the more accurate the final L∗ calculation. The question that needs to
be asked then, is: How many field line traces are needed in order to obtain an accurate integration
path?
Figure 4.23
Traditional methods usually involve adding field line traces until the final flux solution con-
verges (see Roederer (1970) and Roederer and Zhang (2014)). The number of field line traces
can be reduced, however, through the application of spherical linear interpolation between points.
When applied to accurate field line traces, experimentation shows that, if field line fits are good
enough, a 4-line convergence (lines at noon, midnight, dawn and dusk) can be accurate to the
87
Algorithm 12 Drift Path Search Algorithm (λ adaptive)
1: procedure Search KB(φ0 , λpt )
2: ol ← starting point radial distance
3: λτ ← λpt
4: φτ ← φ0
5: rin ← 0.65 × ol
6: rout ← 1.85 × ol
7: rτ ← (rin + rout )/2
8: rgap ← rout − rin
9: bgapin ← None
10: bgapout ← None
11: while rgap > do
12: rgap ← rout − rin
13: loc ← [rτ , λτ , φ0 ]
14: τ ← Trace line from Loc
15: bgap ← bτ (k) − bm
16: if bgap is None then
17: if bgapin is None AND bgapout is not None then
18: rin ← rτ
19: else
20: rout ← rτ
21: rτ ← (rin + rout )/2
22: Continue to next loop
23: else if bgap < 0 then
24: if rτ == rin then
25: rin ← rin /2
26: rout ← rτ
27: rτ ← (rin + rout )/2
29: else
30: rout ← rτ
31: rτ ← (rin + rout )/2
33: else if bgap > 0 then
34: if rτ == rout then
35: rout ← rout × 1.5
36: rin ← rτ
37: rτ ← (rin + rout )/2
39: else
40: rin ← rτ
41: rτ ← (rin + rout )/2
43: else if bgap is close to 0 then return τ
return τin , τout
88
first decimal place for low L∗ particles. Figure 4.23 graphically illustrates the concept. In this
figure, the red lines represent field lines traced with the T96 model. The black dots represent the
four points used to calculate the value of L∗ , and the yellow lines represent the integration path
as determined through spherical linear interpolation. Visually, it can be seen that the boundary
is well aligned with the field line footprints. The effect can be seen more quantitatively, utilizing
LFM data, in the difference plots provided in Figure 4.24. In these images, the difference between
a 4-line convergence, and a convergence to 10−4 is presented. In these quiet solar wind condition
images, we can see that the difference between these two methods is roughly 0.025 where below
L∗ ≈ 7. The differences are greatest where field line density is high, and where drift orbits move
through heavily distorted fields.
These plots only show a single time-step comparison, and only for quiet time conditions. A
more detailed examination, shown in Figure 4.25 as a comparison between L∗ converged to 103 and
a 4-point convergence. It can be seen that, in quieter times, the correlation is quite good, though
not perfect. The biggest difference is the missing of high-disturbance events, such as the onset of
the storm shown in the figure. Some of the disturbance is captured, but not to the magnitude as
the higher level of convergence. It may be possible, through more extensive study, to develop a
better set of directions to use as the interpolation base to possibly yield a better fit.
4.4 Application of Methods
The purpose of calculating adiabatic invariants is to order geomagnetically trapped particles.
The distribution function for particles within the radiation belts is a six-dimensional problem, given
by
f = f (x, y, z, px , py , pz ) (4.5)
and completely characterizes the radiation belts for a point in time, providing the number of
particles within a volume (x + dx, y + dy, z + dz) with the momenta (px + dpx , py + dpy , pz + dpz ).
89
Figure 4.24: These plots show the differences between a 4-line convergence with respect to a full
convergence to 10−4 . The contour lines are whole integer intervals of L∗ values from within the
simulation.
90
RBSP A - 10−3 Conv.
RBSP LFM/RCM 10−3 vs. 4-Line L * Converged Values RBSP B - 10−3 Conv.
(Calculated with GHOSTpy) RBSP A - 4-line
10
RBSP B - 4-line
9
7
L * (LFM/RCM)
1
0 5 10 15 20
Hour of Day (02 Oct 2013)
Figure 4.25
This can be written equivalently in adiabatic invariant space as
f = f (M, K, L, φ1 , φ2 , φ3 ) (4.6)
where M is the first invariant, K is the second invariant, and L the third. φ1 , φ2 , and φ3 represent
their respective phases. If the phase distribution is uniform, the function can be reduced to a func-
tion of just the three invariants, thereby reducing a 6-dimensional problem down to a 3 dimensional
problem, and would be given as
f = f (M, K, L). (4.7)
With the GHOSTpy analysis system now in place, a real-world application example is the
next logical step. The event chosen for a trial analysis is a CME event that impacted Earth
on 02 Oct 2013. This storm is the immediate predecessor to the event examined in Chapter 2,
and was chosen because it falls within the active time period of the Van Allen radiation belt storm
probes, and quad-resolution LFM simulations already exist in two configurations; with and without
91
an integrated ring current model. This reality provides an opportunity to explore the difference
between two simulation methods, and see how the modeled L∗ and K compare to those originally
computed for the spacecraft observations. The plots in Figure 4.26 provide an overview of the solar
wind conditions as observed at L1. The storm is clearly identifiable in the early morning of October
2, with an initial impact time of around 0100 UT.
WIND/MFI Magnetic field magnitude (3 sec)

35.
30.
25.
20.
B (nT)
15.
10.
5.
0.
WIND/3DP ion velocity

-350.
-400.
-450.
km/s
-500.
-550.
-600.
-650.
WIND/3DP ion density, scalar
50.
Ion N (#/cm^3)
40.
30.
20.
10.
0.
WIND/3DP ion temperature, scalar
50.
40.
Ti (eV)
30.
20.
10.
00:00 03:00 06:00 09:00 12:00 15:00 18:00 21:00 00:00

2013-10-02 2013-10-03
Figure 4.26: Solar wind conditions for 02 October 2013. The CME event impact can be seen early
in the morning, near 0100 UT.
Figure 4.27 shows the Disturbance Storm Time (DST ) index associated with the 02 Oct 2013
storm. The red dot indicates the peak negative DST for the storm, and indicates the date and time
that is utilized for the full field comparison in Section 4.4.2 ahead.
92
Disturbance Storm Time (Dst)
01 Oct 2013 - 06 Oct 2013
40
20
−20
Dst
−40
−60
2013-10-02 07:16:39: Dst = -77.14

−80
2013-10-01 2013-10-02 2013-10-03 2013-10-04 2013-10-05 2013-10-06
Figure 4.27: AER Dst index for 01 Oct 2013 - 06 Oct 2013. The red dot indicates the peak negative
Dst and indicates the date that will be analyzed with full field comparisons.
4.4.1 Van Allen Probes L∗ and K Comparisons
The top plot in Figure 4.28 shows L∗ calculation from the definitive data Las Alamos Na-
tional Lab (LANL) produced for both Van Allen Probe spacecraft. L∗ values are plotted for both
spacecraft, with the lines colored by orbit. The y-axis represents L∗ calculated from the Tsyga-
nenko T89 model (see Chapter 2), and the x-axis reflects the time of day on 02 Oct 2013. The
input driver for the T89 model, Kp, is plotted as a gray dashed line. The bottom plot shows the
L∗ vs. K relationship (again, calculated from the T89 model, and colored by spacecraft orbit) for
each probe.
The plots in Figure 4.29 present the same spacecraft positions, with L∗ and K invariants
computed on an underlying dipole field, and calculated utilizing the GHOSTpy utilities developed
as a byproduct of the research in this thesis. The dipole field is overlaid on a quad resolution LFM
grid.
As would be expected with a dipole field with no external contributions, there is no distortion
to the L∗ arcs along the spacecraft orbital trajectory. The K values also present in a very regular
93
RBSP T89 L* Values (RBSP Definitive) RBSP A - orbit 1
10 RBSP A - orbit 2
RBSP A - orbit 3
9 RBSP A - orbit 4
RBSP B - orbit 1
7
RBSP B - orbit 2
8
RBSP B - orbit 3
Kp
7
6
6
L* (T89)
Kp
5 5
3 4
2
3
1
0 5 10 15 20
RBSP T89 L* vs. K (RBSP Definitive)

60
RBSP A - orbit 1
RBSP A - orbit 2
RBSP A - orbit 3
50
RBSP A - orbit 4
RBSP B - orbit 1
RBSP B - orbit 2
40 RBSP B - orbit 3
30
K
20
10
0
1 2 3 4 5 6 7 8 9 10
L* (T89)
Figure 4.28: Van Allen Probes A and B L∗ and K for observations on 02 October 2013, computed
utilizing the T89 magnetic field model. These data were retrieved from the Coordinated Data
Analysis Web (CDAWeb) data repository hosted by NASA Goddard Space Flight Center.
fashion with the largest Ks appearing at lower L∗ values, though the K values show scatter when
compared to the definitive data. This scattering is due to the combined effect of interpolation and
line trace errors within the LFM grid. In general, the trends of L∗ can K calculated with a dipole
model agree those of the T89 model, though the magnitudes vary. This is to be expected as the
94
dipole is only a first order approximation of the magnetic field.
RBSP Dipole L * Values (Calculated with GHOSTpy)

10
RBSP A - orbit 1
9 RBSP A - orbit 2
RBSP A - orbit 3
RBSP A - orbit 4
8
RBSP B - orbit 1
RBSP B - orbit 2
7 RBSP B - orbit 3
L * (Dipole)
1
0 5 10 15 20
RBSP Dipole L * vs. K (Calculated with GHOSTpy)

60
RBSP A - orbit 1
RBSP A - orbit 2
50 RBSP A - orbit 3
RBSP A - orbit 4
RBSP B - orbit 1
RBSP B - orbit 2
40
RBSP B - orbit 3
30
K
20
10
0
1 2 3 4 5 6 7 8 9 10
L * (Dipole)
Figure 4.29: Dipole L∗ and K for the Van Allen Probe positions on 03 October 2013. Calculated
utilizing GHOSTpy.
Plots for L∗ and K from LFM and LFM/RCM data are presented in Figures 4.30 and 4.31
respectively. The calculations for these plots were conducted in parallel utilizing the GHOSTpy
tools, and values were converged to 10−3 after an initial 4-point convergence. Running on the
95
Summit supercomputer, the time to completion of each set (one for LFM and another LFM/RCM)
was approximately 15 minutes on 192 cores (8 nodes). Position and epoch data were obtained
from the definitive RBSP data product, and a script was written to match satellite times to model
files. These matched file/point lists were then imported into a variant of the parallel GHOSTpy
application for computation.
RBSP LFM L * Values (Calculated with GHOSTpy)

10
RBSP A - orbit 1 RBSP B - orbit 2
9 RBSP A - orbit 2 RBSP B - orbit 3
RBSP A - orbit 3 RBSP A - 4-line
RBSP A - orbit 4 RBSP B - 4-line
8
RBSP B - orbit 1
7
L * (LFM)
1
0 5 10 15 20
RBSP LFM L * vs. K (Calculated with GHOSTpy)

60
RBSP A - orbit 1
RBSP A - orbit 2
50 RBSP A - orbit 3
RBSP A - orbit 4
RBSP B - orbit 1
RBSP B - orbit 2
40
RBSP B - orbit 3
30
K
20
10
0
1 2 3 4 5 6 7 8 9 10
L * (LFM)
Figure 4.30: LFM L∗ vs. Time

96
When comparing these plots to either the T89- or dipole-calculated versions, the most obvious
feature of the LFM calculated traces is the smoothness. The dipole L∗ , being the first order
approximation, has no variation in smoothness over course of the orbits. The T89 plots show a small
degree of variation, with L∗ curve discontinuities near the tops of the curves. These discontinuities
are a direct result of the 3-hour Kp index resolution used to drive the model (See the Kp plot in
Figure 4.28). The LFM generated plots show a great deal of variation in the traces, as would be
expected from the more dynamical MHD simulation fields. Beyond the smoothness differences, the
LFM (non RMC) trace in Figure 4.30 shows relatively large L∗ values during the first orbit when
compared to both Figure 4.28 and 4.29. This time period corresponds to the main phase of the
storm, and stronger variations are expected.
With the addition of the ring current to the LFM, by way of the RCM model, the peak L∗
values, not including the early portion of the storm, are more in line with those seen in the other
two iterations. Van Allen Probe B was outbound in its orbit at the time of CME impact (See
Figure 4.32 for a view of the magnetic field around Earth near the time of impact), and appears to
capture the onset of the event. It can be seen from roughly 0100 through 0900 that magnetosphere
is exhibiting a large amount of compression, pushing L∗ values higher. As the field relaxes over
the course of the storm, the L∗ values return to a more nominal value similar to those seen in the
dipole- and T89-calculated L∗ values.
Comparing the K values in Figure 4.30 with those in Figure 4.31, we see the majority of
variation between between L∗ = 4 and L∗ = 6. This is as expected, as the LFM/RCM version
shows more activity in those L∗ regions. Good agreement is seen with the T89 model produced
values for orbits 2 & 3, with orbit 1 being the most different. Again, this difference is due to the
magnetic field dynamics involved in the MHD simulation that exceed those seen in the relatively
simple T89 model.
As a final note on the comparisons of L∗ output, it can be seen in the plots that a 4-line
convergence with LFM MHD data, the general trend of L∗ is captured. With the addition of the
ring current in the LFM/RCM runs, the fit is even better. Although a more thorough study of
97
*
RBSP LFM/RCM L Values (Calculated with GHOSTpy)
10
RBSP A - orbit 1 RBSP B - orbit 2
9 RBSP A - orbit 2 RBSP B - orbit 3
RBSP A - orbit 3 RBSP A - 4-line
RBSP A - orbit 4 RBSP B - 4-line
8
RBSP B - orbit 1
7
L * (LFM/RCM)
1
0 5 10 15 20
RBSP LFM/RCM L * vs. K (Calculated with GHOSTpy)

60
RBSP A - orbit 1
RBSP A - orbit 2
50 RBSP A - orbit 3
RBSP A - orbit 4
RBSP B - orbit 1
RBSP B - orbit 2
40
RBSP B - orbit 3
30
K
20
10
0
1 2 3 4 5 6 7 8 9 10
L * (LFM/RCM)
Figure 4.31: LFM/RCM L∗ vs. Time
4-line convergence utilizing spherical-linear interpolation would be required, the method appears
to have promise as a short-cut in calculating the computationally intensive L∗ values.

98
4.4.2 LFM vs. LFM/RCM Full Field Comparisons
Utilizing GHOSTpy in parallel, it is possible to process a large number of points in a relatively
short period of time. This provides an opportunity for comparisons of K and L∗ values within an
entire 3-D region of the grid. This process was utilized to construct a comparison between LFM and
LFM/RCM model results. Utilizing the Python mpi4py package, an application was written on top
of GHOSTpy to handle a large quantity of points in parallel. The application takes as input the
LFM grid file containing the B field, a radial distance, used as a limit on which points to process,
and a convergence value. The convergence value can be either -1 for only 4-point convergence, or a
floating point value representing the torrence for convergence. From these inputs, the application
opens the supplied LFM data file, extracts a list of points, and applies the radial distance bounds,
selecting all points within the radius as the computation set. Each process rank independently bins
the points that it is responsible for calculating, and immediately begins processing.
For each point, Bm is set to the value of |B| at the grid location. K is then calculated from
this Bm and both values saved. A 4-point convergence is then performed and the resultant L∗
value saved. If a convergence value was initially specified, each point is subjected to a convergence
routine until the specified tolerance is reached. The resultant L∗ value is then saved. Once all
points have been calculated within a rank, a “gather-all” operation is initiated, sending all saved
data to the rank 0 process. Rank 0 then re-assembles the points onto a grid and saved to disk in a
VTK file format suitable for use with ParaView.
The time-step of 02 Oct 2013 at 07:16:30 UT was chosen for the following analyses, corre-
sponds to the peak Dst index value during the main phase of the storm. Figure 4.32 provides and
overview of the magnetic fields conditions (for this time step) around Earth for both the LFM and
LFM/RCM model output. These are provided as a reference while comparing invariant plots.
In Figures 4.33 through 4.36, equatorial and meridional cut planes are proved to illustrate
the K and L∗ configurations that exist within the time step for both the LFM and LFM/RCM
runs. Figures 4.33 and 4.34 provide views of K for pitch angles of 90◦ at all grid points. The top
99
Figure 4.32: Side by side comparison of magnetic field magnitude, in the X/Y plane, between LFM
without RCM (left) and LFM with RCM (right)
images represent the straight LFM runs, with the bottom images presenting the results utilizing
LFM/RCM runs. Contour lines indicate successive whole-value L-shells ranging from L∗ = 3 near
100
the inner boundary to L∗ = 10 at the outer boundary.
A visual inspection of the two K plots reveals little difference discernible with the naked eye.
The northward swing of the K = 0 region of Figure 4.34 appears to be sharper in the LFM/RCM
model than it does with the straight LFM, and a small bulge appears at around −8RE in the low
K region.
In the same character as the K plots, Figures 4.35 and 4.36 represent the L∗ distributions
based on the K values presented in the previous figures. Again, contour lines represent the whole-
number L∗ shells from 3 to 10. The L∗ distributions in Figure 4.35, for the LFM-calculated values,
appears to be fairly smooth and regular, outside of a discontinuity that appears along the axis, to
roughly L∗ = 8. Beyond this point the continuity of L∗ seems to break down. Similar behavior is
seen in the LFM/RCM generated plots, but with far less smoothness in the constant L∗ shells. Due
to expansion outward of the L∗ regions, as compared to the non-RCM model version, the breakup
of coherent L∗ regions begins just past L∗ = 6.
Figure 4.36 shows that the smoothness of constant L∗ shells appears to be much more regular
when moving from north to south along the x axis. The discontinuity seen in the previous figures
is evident along the z = 0 line of the x axis. The north-south breakup of shell coherence can be
seen in the extreme with the LFM/RCM results in this figure.
Figure 4.37 provides a view of the relative differences in L∗ between the two model config-
urations, with black contour lines representing straight LFM, and red depicting the results from
using LFM/RCM data. Note that these contours show a discontinuity along the +X axis. This
discontinuity appears in some straight LFM runs, and in most LFM/RCM runs, with it being more
pronounced in the LFM/RCM results. The nature of this discontinuity is currently unknown, but
should be investigated further.
Of interest in this plot is the sunward bulging tendency of the LFM/RCM results compared
to the LFM results. It can be seen, starting just past L∗ = 3, that the forward edge of the L-shell
moves radially out, with the differences being the greatest between L∗ = 5 and L∗ = 6. The
LFM/RCM results between these two L∗ values completely encompass the L∗ = 6 and L∗ = 7
101
Figure 4.33: Equatorial cut-plane views of the distribution of K for a pitch angle of 90◦ ) within
full field results generated with the LFM model (top), and the LFM with RCM model (bottom)
102
Figure 4.34: Meridional cut-plane views of the distribution of K for a pitch angle of 90◦ ) within
full field results generated with the LFM model (top), and the LFM with RCM model (bottom)
103
sunward regions computed with the straight LFM model. It can also be seen quite clearly that the
LFM/RCM constant L∗ shells are far more turbulent than those of the LFM model results. Areas of
apparent disorder appear in the LFM/RCM tail region past L∗ = 6. While some disorder appears
in the LFM-result tail region, this phenomena is far more pronounced in the model containing a
ring current.
Figure 4.38 provides a set of plots representing the actual quantitative differences between
the results obtained using the LFM vs. the LFM/RCM model data. These difference plots show
the least variation between the two exists inside of L∗ = 5 as calculated with the Non-RCM model
(The contour lines are the constant L∗ values from the straight LFM results). As distance from
Earth increases, the differences progressively increase, with the largest differences seen in the tail
region.
104
Figure 4.35: Equatorial cut-plane views of the distribution of L∗ , generated from K within the
previous figures, for full field results generated with the LFM model (top), and the LFM with RCM
model (bottom)
105
Figure 4.36: Meridional cut-plane views of the distribution of L∗ , generated from K within the
previous figures, for full field results generated with the LFM model (top), and the LFM with RCM
model (bottom)
106
Figure 4.37: Contour plots showing constant L∗ shells in an equatorial view. Black lines represent
values calculated utilizing LFM without RCM, and the red lines depict values obtained utilizing
LFM with RCM.
107
Figure 4.38: Difference plots showing the absolute L∗ differences between values computed using
LFM vs. LFM/RCM as a model base.
Chapter 5
Parallel Operations and Timings
5.1 Serial Single Point Algorithm Timings
The following timings were conducted on a 12-core, 2.4 GHz, 2009 Mac Pro with 64 GB of
RAM running macOS Sierra (10.12.4). The python subsystem was installed via the Homebrew
package manager, and consists of the most recent version 2.7 libraries available at the time of
writing. The timings consist of the averages from 500 randomly selected points (restricted to a
radial distance of 8.5 RE and bounded by λ between − π4 and π

4 ) within a quad resolution LFM
simulation, with Bm being set by the |B| value at the selected point. K is defined by this Bm , and
L∗ calculated based on these values.
Figure 5.1 provides a histogram with the distribution of final L∗ calculated values. Figure
5.2 provides an overview of the completion time for all 500 samples. More interestingly, Figure 5.3
provides the relationship between execution time and final calculated L∗ values. A clear correlation
exists, with an increase in computation time seeing very little variation in increases between 2.3RE
and 7RE . Beyond 7RE , the variation increases due to the greater complexities in outer field
lines, and the associated divergence in field line trace times. This correlation provides for a good
understanding to estimate execution time for L∗ and K computations based on the approximate
particle distance and latitude. For calculations with particles at low RE and low λ, it can be
assumed that execution time will be in the lower end of the spectrum. Any particles with high
RE or high λ starting points will result in higher execution times. This correlation, along with
the dipole analytic model for L∗ (Equation 3.17), can be utilized for initial load balancing when
109
*
Histogram of Calculated L Values
70
60
50
40
30
20
10
0
3 4 5 6 7 8 9 10
L*
Figure 5.1: This histogram shows the distribution of final calculated L∗ values within the sample
set.
Histogram of Execution Time for Calculated L * Values
100
80
60
40
20
0
10 20 30 40 50
Execution Time (s)
Figure 5.2: This histogram shows the distribution of execution time per calculation within the
sample set.
calculating a large number of particles in parallel. This will be covered in sections on parallel
operations.
110
Particle L * vs. Execution Time (4-Line Convergence)
Quad Res LFM Grid
50
40
Execution Time (s)
30
20
10
3 4 5 6 7 8 9 10
L*
Figure 5.3: This plot shows the relationship between execution time and calculated L∗ value for
the sample set.
Figure 5.4 provides a graphical representation of the effects described by Figure 5.3. Notice
that the largest execution times exist in the tail region (the left side of the left-hand plot). High-
time regions can also be seen, to a more limited effect on north and south edges of the field depicted
in the y-z plane in the right-hand image. It can also be seen from this figure that the majority of
points have a relatively low computation time, indicating that a relatively few number of particles
will have an undue influence on the load balancing of parallel computations. The values depicted in
this figure represent calculations that have had their values converged to 10−3 , thus the individual
point calculation times are larger than would be expected for a 4-point convergence.
5.1.1 Single Point Computation - L∗ Overview
Before dealing with parallel operations, the serial execution for a single particle should be
examined. Figure 5.5 provides a timing overview for the major components of a single-point 4-point
convergence L∗ calculation. The analysis is refined to a lower level in Figure 5.6 with a breakdown of
the major components of the KBSearch methods. The most computationally expensive component
111
Figure 5.4: Graphical illustration of time required to compute L∗ based on location. This example
originates from the main phase of the storm depicted in Chapter 4.
is then further refined with a breakdown of the field-line/KB model routines.
Looking at the top level overview in Figure 5.5, the “ init ” method represents the 500
sample average timing for a complete L∗ calculation converged on 4-lines. The “l star’ ’ method
represents the drift boundary magnetic flux integral plus the fixed computation for obtaining L∗ ,
“add trace” is the timing for adding a φ angle interpolation point to the drift trajectory computation.
The bar representing the “ trace from location ” method indicates the time necessary to compute
a single field line trace. Within all of the profile charts, blue bars represent cumulative time over
the course of a single point execution; the orange bars represent time per call.
The average for a 4-point convergence is skewed high because points with high Bm , high K
combinations tend to take considerably more time to execute than do points of lower Bm and lower
K. This is due to the field geometry; the former situation requires extensive searching through long
and distorted field lines. These timings make it clear that the majority of execution time is spent
searching for the proper K/B field line fit. It is important to note that “ add trace ” utilizes both
112
“l star ” and “ trace from location ” as integral parts of the computation. Logically, ‘‘ init ” is
comprised of 4 line traces and a single L∗ computation. Each search progression utilizes L∗ for
boundary calculations, and a considerably number of field line traces to complete the search. The
“ trace from location ” method represents the smallest grained timing.
Average Execution Time per Method Call for a 4-point Converged L * Calculation
For a Single Point
102
98.817 % 98.960 % 100.000 %
20.31 s
25.798 %
101
Time (%)
1.667 %
100
0.884 %
0.171 %
__trace_from_location__
add_trace
__init__
l_star
Function Call
Figure 5.5: Average execution time for calculating a single point (in serial), broken down by com-
putational component. The blue bars indicate cumulative time per point. The orange bar indicates
the execution time per call.
5.1.2 Single Point Computation - KB Line Search Routine
Focusing on the component that takes the most time to complete, Figure 5.6 provides a
timing breakdown for a single line search operation,where “ search B adaptive ” represents the
113
Average Execution Time per Method Call for KB Line Search
102
100.0000 %
5.2259 s
101 14.5726 %
100
0.4998 %
Time (%)
0.2758 %
−1
10
10−2 0.0139 %
0.0087 %
10−3
0.0003 %
__get_min_b_gap__
__trace_line__
get_raw_path
__search_B__
Function Call
Figure 5.6: Average execution for a KB search and its component methods. “ search B adaptive ”
represents the total time (avg) for a single execution of a search. The remaining bars represent
time for the given methods within a single search. Blue represents cumulative time; orange is time
per call.
execution of a single line search. Again, the timings are skewed high due to the nature of high K,
high Bm points. Looking specifically at “ search B adaptive ” removes the L∗ component from
the calculation, allowing for a more focused analysis of the search.
Within a single search the primary drivers for time expenditure are the tracing of field lines,
and calculating the minimum bgap (k) for the lines. The smallest non-system time expenditure is
“get raw path”, which is the method used in calculating the λ start position for a search.
114
5.1.3 Single Point Computation - K/B Model
Figure 5.7 represents a refinement of timing granularity, focusing on the calculation of a field
line and its associated K/B model. In this figure, once again “ init ” represents the total time to
completion for tracing the complete calculation. The method “ trace line ” actually performs the
Runge-Kutta 4/5 integration to calculate the field line; “ k mod ” calculates K values for every
point on the line. This method is executed twice per line; once forward, once backward. It is clear
from this plot that calculating K requires the vast majority of the single line trace.
Average Execution Time per Method Call for Field Line Trace and K Model
102
100.000 %
0.18 s
85.374 %
6 × 101
4 × 101 43.678 %
Time (%)
3 × 101
2 × 101
14.557 %
__trace_line__
__k_mod__
__init__
Function Call
Figure 5.7: Average execution for a Field line trace with associated K model computation. “ init ”
represents the full execution time for a single computation, with the other two being the major
sub-components. Blue bars represent cumulative time; orange is time per call.
115
5.2 Parallel Operations
Using the GHOSTpy tools to calculate invariants in parallel is a fairly straightforward task.
As alluded to in Chapter 4, an application was written on top of the GHOSTpy toolkit, utilizing
the mpi4py message passing interface, in order to process large numbers of points. It is important
to note that this application is a sample of what can be done with the system. Analyses of trapped
particles may take on many different forms, requiring different sets of particles. The application
written for this thesis calculates the invariants for particles within a specific radial distance of the
Earth’s center. Other possible applications might include finding the K = 0 surface, a task that
would require an iterative process to create a grid while calculating the invariants. Other tasks
include, as is demonstrated in Chapter 4, the calculation of invariants for spacecraft locations.
These other tasks would require a different driver application for the worker classes that have been
written, for the application outlined here, to process particles in parallel.
5.2.1 Application Configuration
The application to process particles in parallel is functionally divided into two parts: a
driver application that partitions the points for processing, and a worker module that oversees
the computations on parallel nodes. As each particle can be computed entirely independent of
all other processes, interprocess communication is, ideally, not required until nodes complete their
processing tasks. Though, there may be situations where communications may be necessary to
properly balance the computational load for optimal performance.
5.2.2 Parallel Scaling Efficiency
There are two types of efficiency that are normally computed for parallel applications; strong
scaling and weak scaling. Strong scaling is utilized when a problem size remains fixed, and the
number of processing units (processes) are increased to improve performance. This is typically
the case where a problem is processor bound (takes a long time to compute a single data unit).
116
The goal behind strong scaling is to reduce the time required to compute a fixed size solution to
a reasonable number. These problems generally incur a great deal of communications overhead,
and the efficiency analysis provides a means of finding the best performance possible for a given
problem. Strong scaling efficiency is given by:
t1
Es = (5.1)
N tN
where t1 represents the time to compute a solution for a work unit utilizing a single process, and
the time required to compute a solution for the same work unit utilizing N processes is tN .
The concept of weak scaling involves problems where the computational workload for a single
data unit is fixed to a single process, and speedup is obtained when additional data units are added,
each assigned to its own process. This efficiency of weak scaling is given by:
t1
Ew = × 100% (5.2)
tN
where t1 represents the time to complete one work unit utilizing a single process, and tN is the
length of time required to complete N work units utilizing N processes.
Due to the nature of the invariant problem (where a particle K and L∗ values can be computed
independently of any other particle), the best parallel model to use is weak scaling. The efficiency
plots in Figure 5.8 provide an overview of the weak scaling efficiency for the parallel GHOSTpy
operations running on Summit.
The efficiency plots in this figure are broken down by test run and the overall average. The
tests were conducted 3 independent times, with each run selecting its particles to calculate at
random. All tests ran the same number of points on all ranks. The nature of the application is
such that all ranks hand their computed results back to rank 0 upon completion. Rank 0 then
organizes and saves the data to disk.
From these tests, we can see that the application scales fairly well; it maintains > 60%
efficiency while running on 1,536 processes. The overall degradation in efficiency is due to the final
117
Weak Scaling Efficiency
Running on Summit
100 Run 1
Run 2
Run 3
average
90
Weak Scaling Efficiency (%)
80
70
60
100 101 102 103

Number of Cores (work units)
Figure 5.8: Parallel efficiency profiles for GHOSTpy running in parallel on the Summit supercom-
puter. The efficiency profiles were run 3 independent times, represented by the blue, orange and
green lines. The average efficiency is plotted on the heavy black line.
data gather and I/O. If these aspects were to be removed, the application would see even higher
efficiency ratings.
5.2.3 Parallel Timing Profile
The time to complete execution can be viewed in 5 different segments. These include parallel
communications time, final data I/O, invariant computations, point binning and point loading. In
the following timings, data from the efficiency analysis above was utilized. Each core self-bins an
identical set of 40 work units, loads the field data, then executes the computations with a 4-point
convergence, then returns the data to rank 0 for sorting and writing to disk. Timings are taken at
each point, and returned separately so as not to interfere with the communications timings. The
timings for I/O include the time required to re-assemble the data into its proper grid.
Figure 5.9 provides an example timing for 1,536 cores. In this test, invariant computations
make up 69.33% of the total execution time. This correlates well with the average efficiency for
1,536 cores seen in Figure 5.8. Communication time accounts for 7.84% of total execution time,
118
while final assembly of data and output I/O represent 12.08%. Binning comes in at 7.28% and
loading grid data at 3.47%.
Parallel Execution Timing Profile for 1536 cores

70 69.33 %
60
50
Execution Time (%)
40
30
20
12.08 %
10 7.84 % 7.28 %
3.47 %
0
Data Load
Compute
Final I/O
Binning
Comms
Figure 5.9: Average parallel execution timing breakdown for GHOSTpy parallel jobs running on
1,536 processes.
According to the profile in 5.8, efficiency decreases as the number of cores increase. The cause
can be inferred by looking at the change in communications and I/O requirements as the number
of processes increase. This effect is profiled in Figure 5.10. The combined communication and final
I/O timings increase at a rate commensurate with the decrease in average efficiency. This indicates
that, to improve overall efficiency, a parallel I/O scheme would need to be implemented. Utilizing
parallel I/O would completely eliminate the need for interprocess communications, and the system
would exhibit theoretical linear scaling to an unlimited number of work units. In reality, an active
load balancing scheme would need to be implemented for extremely large jobs, requiring a degree
of interprocess communications, albeit less than a full gather operation would require.
119
Timings for Quad Resolution Computations
by core count
Comms
50 I/O
Average Time per Core (s)
40
30
20
10
102 103
Number of Cores
Figure 5.10: Average parallel execution timing breakdown for GHOSTpy parallel jobs running on
1,536 processes.
5.2.4 Load Balancing
For the initial GHOSTpy parallel application, passive load balancing is utilized. This type
of load balance relies on each process to determine it own optimal set of points, without input
from any other process. With the system profiled here, execution begins with each process rank
computing its own working set of particles based on a simple modulus operation. The application
is directed via command line arguments to restrict the selection of points within the LFM grid to
a specified radial distance. Once the proper data file is loaded, the grid points are extracted and
filtered to meet the radial distance restriction. Points are then binned and distributed to worker
processes.
As not all points within the specified sphere will be calculable, there are likely to be a
relatively large number of points that will be immediately rejected by the K/B algorithm. This is
typically caused when field particles reside on field lines that do not return to the inner boundary
of the model (they terminate at the outer boundary). These particles will immediately return a
N aN value, and the system will proceed to the next point. Other non-calculable points include
120
points that reside on the inner boundary AND do not have a valid mirror point at the other pole.
This will generate a N aN K value in the K/B algorithm for the particle, and a N aN L∗ value be
immediately returned.
The final failure mode for a computation involves the initial convergence on 4 points. This
occurs if a particle has a valid initial field line trace and associated K and Bm values, but valid lines
cannot be found in all four of the convergence φ angles: 0, π2 , π, and π

4. This failure is typical on
the inner boundary and on the outer edge of computable points (where lines transition from closed
to open). This failure can take considerably more time to determine that L∗ cannot be computed,
and thus return a N aN .
These failures can cause the computational load to become unbalanced at high node counts,
as the partitioned particles can finish at vastly different times. The L∗ computation timings shown
in Figure 5.3 also contribute heavily to load imbalance. The figure clearly shows that, as the L∗
value of a particle increases, so does the complexity of calculating the invariants. Any computation
bin that contains a large number of high-L∗ particles will exhibit longer execution times. The final
load-balance is also affected by the convergence of a particle calculation. Depending on the level of
convergence desired, this can add considerable expense to the calculations. This added execution
time is directly related to magnetic field complexity of the region being searched, which in turn is
related to its distance from Earth.
Figure 5.11 shows the practical results of this imbalance. This histogram depicts a parallel
run on 1,536 cores processing 379,002 particles (all grid locations within 9.5 RE on a quad resolution
grid) with the application set to converge each particle to 10−4 . Through a simple (n mod rank)
operation, the particles were binned sequentially based on the grid data structure. No effort was
made to organize particles by physical properties. The result is a clearly skewed normal distribution
of execution times per process.
It can be inferred from the failure modes and calculation timing conditions, that when a
particle exhibits a high dipole L∗ value it will also exhibit a high computed L∗ value, and possibly
the lengthy compute-to-fail time associated. These relationships can be used for more efficient
121
Histogram of Rank Time to Completion
1536 Ranks
Running on Summit with Profiling Enabled
140
120
100
Rank Count
80
60
40
20
0
4.0 4.5 5.0 5.5 6.0 6.5
Rank Time to Completion (hours)
Figure 5.11: Normalized histogram representing initial load balance for a quad resolution LFM
grid, processing 379,002 particles on 1,536 cores, and calculating to a convergence of 1 × 10−4 .
binning of particles, and be used to reject particles, prior to calculation, that would exhibit a dipole
L∗ value that exits the simulation space. This improved binning and associated filtering provides
for a more efficient passive load balancing solution.
Figure 5.12 shows the results of modified binning based on the dipole L∗ . The dipole L∗
value was analytically calculated for each particle and placed in a bin based on its truncated
integer value. Bin contents were then evenly distributed over the available processes. It can be seen
that, by utilizing this process, the execution time has been reduced from nearly 6.5 hours down to
just over 3.75 hours.
A variant of this method, not yet attempted, would involve partitioning out bins in their
entirety starting with the smallest bins first. Currently, assignment of points is accomplished by
sequentially assigning the first point in each bin, then repeating until all bins have been emptied.
By distributing the bins from smallest to largest, emptying each bin before moving on to the next,
bunching of computationally intensive points can be avoided when working with a smaller number
of processors.
122
Histogram of Rank Time to Completion (partitioned on dipole L)
1536 Ranks
Running on Summit
140
120
100
Rank Count
80
60
40
20
0
2.00 2.25 2.50 2.75 3.00 3.25 3.50 3.75
Rank Time to Completion (hours)
Figure 5.12: Histogram of execution times with improved binning.
It is important to note that these profiles represent a solution converged down to 10−4 .
In these convergence conditions, the amount of time required to converge any given solution is
not known. The differences in convergence times lead to the normal distributions seen in these
load balance plots. To ensure an optimally balanced execution while utilizing convergence, load
balancing would need to be shifted from passive to active. An active solution can be achieved
by designating a control processor to be responsible for handing out computation assignments to
all worker processes. As each rank completes its initial set of computations, it requests a new
set from the control process. If managed correctly, the load balance in this situation would be
considerably improved, but at the expense of more communications during execution. The added
communications expense will adversely affect overall scaling efficiency if workers request additional
points for processing too often. This would have to be managed carefully.
Chapter 6
Closing Remarks
The development of this first-of-its-kind system for calculating the adiabatic invariants of
geomagnetically trapped radiation within LFM MHD fields provides an easy and extensible means
for scientists to interact with LFM simulation data. The understanding of how the invariant
calculation behaves within this complex gridded environment has led to new questions; questions
that can be answered through application of the tools within the GHOSTpy framework.
The unique 4-line convergence method outlined in this thesis is one aspect that can benefit
from further exploration. It was shown that, for a single storm, the 4-line method performed well
against the full convergence results for quiet times, and was in good agreement with official K and
L∗ values put out by the RBSP mission team. An exploration study should be conducted to examine
the accuracy of the 4-point method over a much larger sample set. Now that the GHOSTpy tools
are available, this should be a fairly straightforward exercise.
GHOSTpy was designed with expandability in mind. The system is based on a plug-in like
architecture, whereby data and field-tracing classes can be sub-classed to provide new sources of
input. So long as the interface requirements are met, any magnetic field model or field line integrator
can be used. This allows for an opportunity to explore other models, and to test different methods
of field line tracing. This architecture also separates the data interface, allowing for more robust
and parallel access to data if needed.
The system also has room for improvement. As a research tool, it should forever be updated
to improve its functionality. There is room for additional strong parallelization within the invariant
124
algorithms themselves, with the potential to reduce computation time even further. The best
scaling attempted during this research utilized 3,072 cores, and processed half a million points to
a 4-point convergence in just under 5 minutes. This was not replicated in the course of research
due to system problems with the new Summit supercomputer, but it does illustrate the potential
for future research possibilities. These tools also open the opportunity for new visualizations of the
magnetosphere, offering the ability to process entire 3-dimensional portions of a grid; video imagery
showing the evolution of L-shells is a real possibility.
The advent of this system opens the opportunity to explore a multitude of open science
questions. Some of these questions might include: How does the choice of model for invariant
calculations affect the overall distribution of particles in phase space density? Can the use of a more
realistic model, such as the LFM, change the way we understand radiation belt dynamics during
magnetic storms, or is the current understanding relatively insensitive to the model choice? How
does the ring current affect phase space density when dealing with the LFM model as a base? Why
does the 4-point convergence method seem to correlate better to the LFM/RCM storm presented in
this thesis, when compared to the LFM without RCM? Is this a coincidence for this selected storm,
or does it suggest an underlying physical process that could simplify the computation of L∗ in the
future? GOSHTpy is ideally suited for exploring these, and many more, science questions. The
extensibility of the system allows for the addition of new models to explore even more questions,
allowing for growth as model sophistication increases.

Bibliography
Arge, C. N. and Pizzo, V. J. (2000). Improvement in the prediction of solar wind conditions using
near-real time solar magnetic field updates. Journal of Geophysical Research: Space Physics,
105(A5):10465–10479.
Ayachit, U., Avila, L., Osterdahl, K., McKenzie, S., and Jordan, S. (2015). The ParaView Guide.
Kitware, paraview 4.3 edition.
Baker, D., Weigel, R., Rigler, E., McPherron, R., Vassiliadis, D., Arge, C., Siscoe, G., and Spence,
H. (2004). Sun-to-magnetosphere modeling: CISM forecast model development using linked
empirical methods. Journal of Atmospheric and Solar-Terrestrial Physics, 66(15–16):1491 – 1497.
Towards an Integrated Model of the Space Weather System.
Baker, D. N., Balstad, R., Bodau, J. M., Cameron, E., Fennell, J. F., Fisher, G. M., Forbes,
K. F., Kintner, P. M., Leffler, L. G., Lewis, W. S., Reagan, J. B., III, A. A. S., Stansell, T. A.,
and Strachan, L. (2008). Severe Space Weather Events–Understanding Societal and Economic
Impacts: A Workshop Report. The National Academies Press.
Baker, D. N., Dewey, R., Anderson, B. J., Ho, G., Korth, H., Krimigis, S., Lawrence, D. J.,
McNutt, Jr., R. L., Odstrcil, D., Raines, J. M., Schriver, D., Slavin, J. A., and Solomon, S. C.
(2015). Energetic electron flux enhancements in Mercury’s magnetosphere: An integrated view
with multi-instrument observations from MESSENGER. In EGU General Assembly Conference
Abstracts, volume 17 of EGU General Assembly Conference Abstracts, page 2517.
Baker, D. N., Li, X., Pulkkinen, A., Ngwira, C. M., Mays, M. L., Galvin, A. B., and Simunac,
K. D. C. (2013a). A major solar eruptive event in july 2012: Defining extreme space weather
scenarios. Space Weather, 11(10):585–591. 2013SW000963.
Baker, D. N., Odstrcil, D., Anderson, B. J., Arge, C. N., Benna, M., Gloeckler, G., Korth, H.,
Mayer, L. R., Raines, J. M., Schriver, D., Slavin, J. A., Solomon, S. C., Trávnı́ček, P. M., and
Zurbuchen, T. H. (2011). The space environment of mercury at the times of the second and
third MESSENGER flybys. Planetary and Space Science, 59(15):2066 – 2074. Mercury after the
MESSENGER flybys.
Baker, D. N., Odstrcil, D., Anderson, B. J., Arge, C. N., Benna, M., Gloeckler, G., Raines, J. M.,
Schriver, D., Slavin, J. A., Solomon, S. C., Killen, R. M., and Zurbuchen, T. H. (2009). Space
environment of mercury at the time of the first messenger flyby: Solar wind and interplanetary
magnetic field modeling of upstream conditions. Journal of Geophysical Research: Space Physics,
114(A10):n/a–n/a. A10101.
126
Baker, D. N., Poh, G., Odstrcil, D., Arge, C. N., Benna, M., Johnson, C. L., Korth, H., Gershman,
D. J., Ho, G. C., McClintock, W. E., Cassidy, T. A., Merkel, A., Raines, J. M., Schriver, D.,
Slavin, J. A., Solomon, S. C., Trávnı́ček, P. M., Winslow, R. M., and Zurbuchen, T. H. (2013b).
Solar wind forcing at mercury: WSA-ENLIL model results. Journal of Geophysical Research:
Space Physics, 118(1):45–57.
Bartels, J., Heck, N. H., and Johnston, H. F. (1939). The three-hour-range index measuring
geomagnetic activity. Terrestrial Magnetism and Atmospheric Electricity, 44(4):411–454.
Biermann, L. and Schlüter, A. (1951). Cosmic radiation and cosmic magnetic fields. ii. origin of
cosmic magnetic fields. Phys. Rev., 82:863–868.
Bleeker, J. A., Geiss, J., and Huber, M. C. (2001). The Century of Space Science. Springer Science
+ Business Media.
Bock, A., Pembroke, A., Mays, M. L., Rastaetter, L., Ropinski, T., and Ynnerman, A. (2015).
Visual verification of space weather ensemble simulations. In 2015 IEEE Scientific Visualization
Conference (SciVis), pages 17–24.
Borovsky, J. E. and Denton, M. H. (2010). Solar wind turbulence and shear: A superposed-epoch
analysis of corotating interaction regions at 1 AU. Journal of Geophysical Research: Space
Physics, 115(A10):n/a–n/a. A10101.
Boscher, D., Bourdarie, S., O’Brien, P., and Guild, T. (2012). Irbem-lib library.
Burkepile, J. T., Hundhausen, A. J., Stanger, A. L., St.Cyr, O. C., and Seiden, J. A. (2004). Role
of projection effects on solar coronal mass ejection properties: 1. a study of cmes associated with
limb activity. Journal of Geophysical Research: Space Physics, 109(A3):n/a–n/a.
Carrington, R. C. (1859). Description of a singular appearance seen in the sun on september 1,

1859. Monthly Notices of the Royal Astronomical Society, 20(1):13–15.
Cash, M. D., Biesecker, D. A., Pizzo, V., de Koning, C. A., Millward, G., Arge, C. N., Henney,
C. J., and Odstrcil, D. (2015). Ensemble modeling of the 23 july 2012 coronal mass ejection.
Space Weather, 13(10):611–625. 2015SW001232.
Chai, J., Zhao, Y., Guo, W., and Tang, Z. (2003). A texture method for visualization of electromag-
netic vector field. In Electrical Machines and Systems, 2003. ICEMS 2003. Sixth International
Conference on, volume 2, pages 805–808 vol.2.
Chien, D. (1995). Numerical evaluation of surface integrals in three dimensions. Mathematics of

Comp, 64:727–743.
Chree, C. (1913). Some phenomena of sunspots and of terrestrial magnetism at kew observa-
tory. Philosophical Transactions of the Royal Society of London A: Mathematical, Physical and
Engineering Sciences, 212(484-496):75–116.
Chulaki, A. (2015). ENLIL with Cone Model (CCMC Website).
Denton, M. H. and Borovsky, J. E. (2008). Superposed epoch analysis of high-speed-stream effects

at geosynchronous orbit: Hot plasma, cold plasma, and the solar wind. Journal of Geophysical
Research: Space Physics, 113(A7):n/a–n/a. A07216.
127
Dewey, R. M., Baker, D. N., Anderson, B. J., Benna, M., Johnson, C. L., Korth, H., Gershman,
D. J., Ho, G. C., McClintock, W. E., Odstrcil, D., Philpott, L. C., Raines, J. M., Schriver, D.,
Slavin, J. A., Solomon, S. C., Winslow, R. M., and Zurbuchen, T. H. (2015). Improving solar
wind modeling at mercury: Incorporating transient solar phenomena into the wsa-enlil model
with the cone extension. Journal of Geophysical Research: Space Physics, 120(7):5667–5685.
2015JA021194.
Dungey, J. W. (1961). Interplanetary magnetic field and the auroral zones. Physical Review Letters,
6(2):47–48.
Eberly, D. (2011). A fast and accurate algorithm for computing slerp. Journal of Graphics GPU
and Game Tools, 15.
Elkington, S. R., Hudson, M. K., Wiltberger, M. J., and Lyon, J. G. (2002). MHD/particle
simulations of radiation belt dynamics. Journal of Atmospheric and Solar-Terrestrial Physics,
64(5-6):607–615.
Elkington, S. R., Wiltberger, M., Chan, A. A., and Baker, D. N. (2004). Physical models of the
geospace radiation environment. Journal of Atmospheric and Solar-Terrestrial Physics, 66(15-
16):1371–1387.
Fehlberg, E. (1969). Low-order classical runge-kutta formulas with stepsize control and their ap-
plication to some heat transfer problems. Technical report, NASA.
Fei, Y., Chan, A. A., Elkington, S. R., and Wiltberger, M. J. (2006). Radial diffusion simulation of
relativistic electron transport by ULF waves in the September 1998 storm. Journal of Geophysical
Research (Space Physics), 111:A12209.
Finlay, C. C., Maus, S., Beggan, C. D., Bondar, T. N., Chambodut, A., Chernova, T. A., Chulliat,
A., Golovkov, V. P., Hamilton, B., Hamoudi, M., Holme, R., Hulot, G., Kuang, W., Langlais,
B., Lesur, V., Lowes, F. J., Lühr, H., Macmillan, S., Mandea, M., McLean, S., Manoj, C.,
Menvielle, M., Michaelis, I., Olsen, N., Rauberg, J., Rother, M., Sabaka, T. J., Tangborn, A.,
Tøffner-Clausen, L., Thébault, E., Thomson, A. W. P., Wardinski, I., Wei, Z., and Zvereva, T. I.
(2010). International geomagnetic reference field: the eleventh generation. Geophysical Journal
International, 183(3):1216–1230.
Forssell, L. K. (1994). Visualizing flow over curvilinear grid surfaces using line integral convolution.
In Proceedings of the Conference on Visualization ’94, VIS ’94, pages 240–247, Los Alamitos,
CA, USA. IEEE Computer Society Press.
Foster, J. C., Erickson, P. J., Baker, D. N., Claudepierre, S. G., Kletzing, C. A., Kurth, W.,
Reeves, G. D., Thaller, S. A., Spence, H. E., Shprits, Y. Y., and Wygant, J. R. (2014). Prompt
energization of relativistic and highly relativistic electrons during a substorm interval: Van allen
probes observations. Geophysical Research Letters, 41(1):20–25. 2013GL058438.
Foster, J. C., Wygant, J. R., Hudson, M. K., Boyd, A. J., Baker, D. N., Erickson, P. J., and Spence,
H. E. (2015). Shock-induced prompt relativistic electron acceleration in the inner magnetosphere.
Journal of Geophysical Research: Space Physics, 120(3):1661–1674. 2014JA020642.
Gannon, J. L., Elkington, S. R., and Onsager, T. G. (2012). Uncovering the non-adiabatic response
of geosynchronous electrons to geomagnetic dist urbance. Journal of Geophysical Research (Space
Physics), 117.
128
Gilbert, W. (1600, 1958). De magnete, translated by pf mottelay.
Gosling, J. T. (1993). The solar flare myth. Journal of Geophysical Research: Space Physics,
98(A11):18937–18949.
Gosling, J. T., Riley, P., McComas, D. J., and Pizzo, V. J. (1998). Overexpanding coronal mass
ejections at high heliographic latitudes: Observations and simulations. Journal of Geophysical
Research: Space Physics, 103(A2):1941–1954.
Green, J. C. and Kivelson, M. G. (2001). A tale of two theories: How the adiabatic response and
ULF waves affect relativi stic electrons. Journal of Geophysical Research, 106:25777.
Green, J. C. and Kivelson, M. G. (2004). Relativistic electrons in the outer radiation belt: Differ-
entiating between acceler ation mechanisms. Journal of Geophysical Research, 109(A3).
Green, J. C., Onsager, T. G., O’Brien, T. P., and Baker, D. N. (2004). Testing loss mechanisms
capable of rapidly depleting relativistic electron flux in th e Earth’s outer radiation belt. Journal
of Geophysical Research, 109(A12).
Harten, A. (1983). High resolution schemes for hyperbolic conservation laws. Journal of
Computational Physics, 49(3):357 – 393.
Harten, A. (1997). High resolution schemes for hyperbolic conservation laws. Journal of
Computational Physics, 135(2):260 – 278.
Harten, A., Engquist, B., Osher, S., and Chakravarthy, S. R. (1987). Uniformly high order accurate
essentially non-oscillatory schemes, {III}. Journal of Computational Physics, 71(2):231 – 303.
Haurwitz, M. W. and Brier, G. W. (1981). A critique of the superposed epoch analysis method:
Its application to solar–weather relations. Monthly Weather Review, 109(10):2074–2079.
Hayt, W. and Buck, J. (2012). Engineering Electromagnetics. McGraw Hill, 8th edition.
Henderson, M. G. (2010). Documentation and reference manual for lanlgeomag library. Technical
report, Los Alamos National Laboratory.
Hiorter, O. P. (1747). Om Magnet-nalen Atskillige andreingar. Kongle Swen Wetenskaps Acad.

Handlingar.
Hudson, M. K., Elkington, S. R., Lyon, J. G., Marchenko, V. A., Roth, I., Temerin, M., Blake,
J. B., Gussenhoven, M. S., and Wygant, J. R. (1997). Simulations of radiation belt formation
during storm sudden commencements. Journal of Geophysical Research, 102(A7):14087.
Hudson, M. K., Paral, J., Kress, B. T., Wiltberger, M., Baker, D. N., Foster, J. C., Turner, D. L.,
and Wygant, J. R. (2015). Modeling cme-shock-driven storms in 2012–2013: Mhd test particle
simulations. Journal of Geophysical Research: Space Physics, 120(2):1168–1181. 2014JA020833.
Hundhausen, A. J., Burkepile, J. T., and St. Cyr, O. C. (1994). Speeds of coronal mass ejections:
Smm observations from 1980 and 1984-1989. Journal of Geophysical Research: Space Physics,
99(A4):6543–6552.
Hunter, J. D. (2007). Matplotlib: A 2d graphics environment. Computing In Science & Engineering,

9(3):90–95.
129
Ida, N. (2000). Engineering Electromagnetics. Springer New York.
Jacobs, M., Chang, L., and Pulkkinen, A. (2013). Automatic segmentation and classification of
multiple coronal mass ejections from coronagraph images. In Proceedings of the International
Conference on Image Processing, Computer Vision, and Pattern Recognition (IPCV), page 1.
The Steering Committee of The World Congress in Computer Science, Computer Engineering
and Applied Computing (WorldComp).
Janhunen, P., Koskinen, K. E. J., and Pulkinen, T. I. (1996). A new global ionosphere-
magnetosphere coupling simulation utilizing locally varying time step. In Rolfe, E. J. and Kalde-
ich, B., editors, International Conference on Substorms, volume 389 of ESA Special Publication,
page 205.
Juanes, R. and Matringe, S. (2009). Unified formulation for high-order streamline tracing on two-
dimensional unstructured grids. Journal of Scientific Computing, 38(1):50–73.
Katz, V. J. (1979). The history of stokes’ theorem. Mathematics Magazine, 52(3):146–156.
Kenwright, D. and Mallinson, G. (1992). A 3-d streamline tracking algorithm using dual stream
functions. In Visualization, 1992. Visualization ’92, Proceedings., IEEE Conference on, pages
62–68.
Kincaid, D. and Cheney, E. W. (2009). Numerical analysis: mathematics of scientific computing,

volume 2 of The Sally series. American Mathematical Society, Providence, R.I., 3rd ed edition.
Kirnosov, V., Chang, L., and Pulkkinen, A. (2014). Automatic segmentation of coronal mass ejec-
tions from stereo white-light coronagraph images. In Proceedings of the International Conference
on Image Processing, Computer Vision, and Pattern Recognition (IPCV), page 1. The Steering
Committee of The World Congress in Computer Science, Computer Engineering and Applied
Computing (WorldComp).
Kistler, L. M., Baumjohann, W., Nagai, T., and Möbius, E. (1993). Superposed epoch analysis of
pressure and magnetic field configuration changes in the plasma sheet. Journal of Geophysical
Research: Space Physics, 98(A6):9249–9258.
Kivelson, M. G. and Russell, C. T. (1995). Introduction to space physics. Cambridge University

Press, Cambridge.
Knipp, D. J. (2011). Understanding Space Weather and the Physics Behind it. Space Technology
Series. McGraw Hill.
Koller, J., Reeves, G. D., and Friedel, R. H. W. (2009). Lanl* v1.0: a radiation belt drift shell model
suitable for real-time and reanalysis applications. Geoscientific Model Development, 2(2):113–
122.
Koller, J. and Zaharia, S. (2011). Lanl*v2.0: global modeling and validation. Geoscientific Model
Development, 4(3):669–675.
Laundal, K. M. and Richmond, A. D. (2017). Magnetic coordinate systems. Space Science Reviews,
206(1):27–59.
130
Lee, C., Arge, C., Odstrčil, D., Millward, G., Pizzo, V., Quinn, J., and Henney, C. (2013). Ensemble
modeling of cme propagation. Solar Physics, 285(1-2):349–368.
Lee, C. O., Arge, C. N., Odstrcil, D., Millward, G., Pizzo, V., and Lugaz, N. (2015). Ensemble
modeling of successive halo cmes: A case study. Solar Physics, 290(4):1207–1229.
Lejosne, S. (2014). An algorithm for approximating the L* invariant coordinate from the real-time
tracing of one magnetic field line between mirror points. Journal of Geophysical Research: Space
Physics, 119(8):6405–6416. 2014JA020016.
Li, W., Thorne, R. M., Bortnik, J., Baker, D. N., Reeves, G. D., Kanekal, S. G., Spence, H. E., and
Green, J. C. (2015). Solar wind conditions leading to efficient radiation belt electron acceleration:
A superposed epoch analysis. Geophysical Research Letters, 42(17):6906–6915. 2015GL065342.
Lyon, J., Fedder, J., and Mobarry, C. (2004). The lyon–fedder–mobarry (lfm) global mhd magneto-
spheric simulation code. Journal of Atmospheric and Solar-Terrestrial Physics, 66(15-16):1333–
1350.
Mays, M., Taktakishvili, A., Pulkkinen, A., MacNeice, P., Rastätter, L., Odstrcil, D., Jian, L.,
Richardson, I., LaSota, J., Zheng, Y., and Kuznetsova, M. (2015). Ensemble modeling of cmes
using the wsa–enlil+cone model. Solar Physics, 290(6):1775–1814.
McCollough, J. P., Elkington, S. R., and Baker, D. N. (2012). The role of shabansky orbits in
compression-related electromagnetic ion cyclotron wave growth. Journal of Geophysical Research:
Space Physics, 117(A1):n/a–n/a. A01208.
McCollough, J. P., Elkington, S. R., Usanova, M. E., Mann, I. R., Baker, D. N., and Kale,
Z. C. (2010). Physical mechanisms of compressional emic wave growth. Journal of Geophysical
Research: Space Physics, 115(A10):n/a–n/a. A10214.
McCollough, J. P., Gannon, J. L., Baker, D. N., and Gehmeyr, M. (2008). A statistical comparison
of commonly used external magnetic field models. Space Weather, 6(10):n/a–n/a. S10001.
McGregor, S. L., Hudson, M. K., and Hughes, W. J. (2014). Modeling magnetospheric response to
synthetic alfvénic fluctuations in the solar wind: Ulf wave fields in the magnetosphere. Journal
of Geophysical Research: Space Physics, 119(11):8801–8812.
McGregor, S. L., Hughes, W. J., Arge, C. N., and Owens, M. J. (2008). Analysis of the magnetic
field discontinuity at the potential field source surface and schatten current sheet interface in the
wang–sheeley–arge model. Journal of Geophysical Research: Space Physics, 113(A8):n/a–n/a.
McGregor, S. L., Hughes, W. J., Arge, C. N., Owens, M. J., and Odstrcil, D. (2011). The distribu-
tion of solar wind speeds during solar minimum: Calibration for numerical solar wind modeling
constraints on the source of the slow solar wind. Journal of Geophysical Research: Space Physics,
116(A3):n/a–n/a.
Millward, G., Biesecker, D., Pizzo, V., and de Koning, C. A. (2013). An operational software tool
for the analysis of coronagraph images: Determining cme parameters for input into the wsa-enlil
heliospheric model. Space Weather, 11(2):57–68.
Min, K., Bortnik, J., and Lee, J. (2013a). A novel technique for rapid L∗ calculation: Algorithm
and implementation. Journal of Geophysical Research (Space Physics), 118:1912–1921.
131
Min, K., Bortnik, J., and Lee, J. (2013b). A novel technique for rapid L∗ calculation using UBK
coordinates. Journal of Geophysical Research (Space Physics), 118:192–197.
Morley, S. K., Welling, D. T., Koller, J., Larsen, B. A., Henderson, M. G., and Niehof, J. (2010).
Spacepy—a python-based library of tools for the space sciences. In Proceedings of the 9th Python
in science conference (SciPy 2010), pages 39–45. Citeseer.
Newell, P. T., Sotirelis, T., Liou, K., Meng, C.-I., and Rich, F. J. (2007). A nearly universal solar
wind-magnetosphere coupling function inferred from 10 magnetospheric state variables. Journal
of Geophysical Research: Space Physics, 112(A1):n/a–n/a. A01206.
Northrop, T. G. (1963). The adiabatic motion of charged particles. Interscience Publishers.
Odstrcil, D. (2003). Modeling 3-d solar wind structure. Advances in Space Research, 32(4):497 –
506. Heliosphere at Solar Maximum.
Odstrcil, D. (2004a). Enlil: A numerical code for solar wind distrubances. Technical report,
University of Colorado Boulder.
Odstrcil, D. (2004b). Enlil: Organization and structure of data. Technical report, University of
Colorado Boulder.
Odstrčil, D. and Pizzo, V. J. (1999a). Three-dimensional propagation of coronal mass ejections

(cmes) in a structured solar wind flow: 1. cme launched within the streamer belt. Journal of
Geophysical Research: Space Physics, 104(A1):483–492.
Odstrčil, D. and Pizzo, V. J. (1999b). Three-dimensional propagation of coronal mass ejections

(cmes) in a structured solar wind flow: 2. cme launched adjacent to the streamer belt. Journal
of Geophysical Research: Space Physics, 104(A1):493–503.
Odstrcil, D., Riley, P., and Zhao, X. P. (2004). Numerical simulation of the 12 may 1997 interplan-
etary cme event. Journal of Geophysical Research: Space Physics, 109(A2):n/a–n/a.
Odstrčil, D., Smith, Z., and Dryer, M. (1996). Distortion of the heliospheric plasma sheet by
interplanetary shocks. Geophysical Research Letters, 23(18):2521–2524.
of Aeronautics, A. I. and Astronautics (1999). Guide to Modeling Earth’s Trapped Radiation

Environment. AIAA G: American Institute of Aeronautics and Astronautics. American Institute
of Aeronautics and Astronautics.
Olmedo, O., Zhang, J., Wechsler, H., Poland, A., and Borne, K. (2008). Automatic detection and
tracking of coronal mass ejections in coronagraph time series. Solar Physics, 248(2):485–499.
Onsager, T. G., Chan, A. A., Fei, Y., Elkington, S. R., Green, J. C., and Singer, H. J. (2004). The
radial gradient of relativistic electrons at geosynchronous orbit. Journal of Geophysical Research
(Space Physics), 109(A5).
Öztürk, M. K. and Wolf, R. A. (2007). Bifurcation of drift shells near the dayside magnetopause.
Journal of Geophysical Research: Space Physics, 112(A7):n/a–n/a. A07207.
Parker, E. N. (1958). Dynamics of the interplanetary gas and magnetic fields. The Astrophysical
Journal, 128:664.
132
Parsons, A., Biesecker, D., Odstrcil, D., Millward, G., Hill, S., and Pizzo, V. (2011). Wang-sheeley-
arge–enlil cone model transitions to operations. Space Weather, 9(3):n/a–n/a. S03004.
Perreault, P. and Akasofu, S. I. (1978). A study of geomagnetic storms. Geophysical Journal

International, 54(3):547–573.
Pizzo, V. J., de Koning, C., Cash, M., Millward, G., Biesecker, D. A., Puga, L., Codrescu, M.,
and Odstrcil, D. (2015). Theoretical basis for operational ensemble forecasting of coronal mass
ejections. Space Weather, 13(10):676–697. 2015SW001221.
Powell, K. G., Roe, P. L., Linde, T. J., Gombosi, T. I., and De Zeeuw, D. L. (1999). A solution-
adaptive upwind scheme for ideal magnetohydrodynamics. Journal of Computational Physics,
154(2):284–309.
Prevost, M., Edwards, M. G., and Blunt, M. J. (2001). Streamline tracing on curvilinear struc-
tured and unstructured grids. In SPE Reservoir Simulation Symposium. Society of Petroleum
Engineers.
Pugmire, D., Childs, H., Garth, C., Ahern, S., and Weber, G. H. (2009). Scalable computation
of streamlines on very large datasets. In Proceedings of the Conference on High Performance
Computing Networking, Storage and Analysis, SC ’09, pages 16:1–16:12, New York, NY, USA.
ACM.
Pulkkinen, A., Oates, T., and Taktakishvili, A. (2009). Automatic determination of the conic
coronal mass ejection model parameters. Solar Physics, 261(1):115–126.
Ridley, A. J. (2002). University of michigan mhd results of the geospace global circulation model
metrics challenge. Journal of Geophysical Research, 107(A10).
Roederer, J. G. (1970). Dynamics of geomagnetically trapped radiation. Springer New York, 1

edition.
Roederer, J. G. and Zhang, H. (2014). Dynamics of Magnetically Trapped Particles. Springer.
Sanna, A., Montuschi, P., and Montuschi, P. (2000). A survey on visualization of vector fields by
texture-based methods. Devel. Pattern Rec, 1:2000.
Savitzky, A. and Golay, M. J. E. (1964). Smoothing and differentiation of data by simplified least
squares procedures. Analytical Chemistry, 36(8):1627–1639.
Schatten, K., Wilcox, J., and Ness, N. (1969). A model of interplanetary and coronal magnetic
fields. Solar Physics, 6(3):442–455.
Schatten, K. H. (1971). Current sheet magnetic model for the solar corona. Cosmic
Electrodynamics, 2:232–245.
Schröder, W. (1998). Hermann fritz and the foundation of auroral research. Planetary and Space
Science, 46(4):461–463.
Schulz, M. (2013). Canonical Coordinates for Radiation-Belt Modeling, pages 153–160. American
Geophysical Union.
133
Shprits, Y. Y., Subbotin, D. A., Meredith, N. P., and Elkington, S. R. (2008). Review of modeling
of losses and sources of relativistic electrons in the outer radiation belt ii: Local acceleration and
loss. Journal of Atmospheric and Solar-Terrestrial Physics, 70(14):1694–1713.
Singh, Y. and Badruddin (2006). Statistical considerations in superposed epoch analysis and its
applications in space research. Journal of Atmospheric and Solar-Terrestrial Physics, 68(7):803
– 813.
Siscoe, G. and Odstrcil, D. (2008). Ways in which icme sheaths differ from magnetosheaths. Journal
of Geophysical Research: Space Physics, 113(A9):n/a–n/a. A00B07.
Sisto, A. and Kamath, C. (2013). Ensemble feature selection in scientific data analysis. Technical
report, Lawrence Livermore National Laboratory (LLNL).
Sparavigna, A. and Montrucchio, B. (2006). Vector field visualization with streamlines. eprint
arXiv:cs/0610088.
Steenburgh, R., Biesecker, D., and Millward, G. (2014). From predicting solar activity to forecasting
space weather: Practical examples of research-to-operations and operations-to-research. Solar
Physics, 289(2):675–690.
Stockinger, K. and Wu, K. (2004). Improved searching for spatial features in spatio-temporal data.
Technical report, Ernest Orlando Lawrence Berkeley National Laboratory.
Tao, X., Chan, A. A., and Brizard, A. J. (2007). Hamiltonian theory of adiabatic motion of
relativistic charged particles. Physics of Plasmas, 14(9).
Temerin, M. and Li, X. (2006). Dst model for 1995–2002. Journal of Geophysical Research: Space
Physics, 111(A4):n/a–n/a. A04221.
Tóth, G. and Odstrčil, D. (1996). Comparison of some flux corrected transport and total variation
diminishing numerical schemes for hydrodynamic and magnetohydrodynamic problems. Journal
of Computational Physics, 128(1):82 – 100.
Tóth, G., Sokolov, I. V., Gombosi, T. I., Chesney, D. R., Clauer, C. R., De Zeeuw, D. L., Hansen,
K. C., Kane, K. J., Manchester, W. B., Oehmke, R. C., and et al. (2005). Space weather
modeling framework: A new tool for the space science community. Journal of Geophysical
Research, 110(A12).
Tsyganenko, N. (1989). A magnetospheric magnetic field model with a warped tail current sheet.
Planetary and Space Science, 37(1):5–20.
Tsyganenko, N. (2000). Modeling the inner magnetosphere: The asymmetric ring current and region
2 birkeland currents revisited. Journal of Geophysical Research: Space Physics, 105(A12):27739–
27754.
Tsyganenko, N. A. (1995). Modeling the earth’s magnetospheric magnetic field confined within a
realistic magnetopause. Journal of Geophysical Research: Space Physics, 100(A4):5599–5612.
Tsyganenko, N. A. (1996). Effects of the solar wind conditions in the global magnetospheric
configurations as deduced from data-based field models (Invited). In Rolfe, E. J. and Kaldeich,
B., editors, International Conference on Substorms, volume 389 of ESA Special Publication, page
181.
134
Tsyganenko, N. A. (2002a). A model of the near magnetosphere with a dawn-dusk asymmetry 1.

mathematical structure. Journal of Geophysical Research: Space Physics, 107(A8):SMP 12–1–
SMP 12–15.
Tsyganenko, N. A. (2002b). A model of the near magnetosphere with a dawn-dusk asymmetry 2.

parameterization and fitting to observations. Journal of Geophysical Research: Space Physics,
107(A8):SMP 10–1–SMP 10–17.
Tsyganenko, N. A. (2013). Data-based modelling of the earth’s dynamic magnetosphere: A review.

Annales Geophysicae, 31:1745 – 1772.
Tsyganenko, N. A. (2014). Data-based modeling of the geomagnetosphere with an imf-dependent

magnetopause. Journal of Geophysical Research: Space Physics, 119(1):335–354.
Tsyganenko, N. A. and Andreeva, V. A. (2015). A forecasting model of the magnetosphere driven

by an optimal solar wind coupling function. Journal of Geophysical Research: Space Physics,
120(10):8401–8425. 2015JA021641.
Tsyganenko, N. A., Singer, H. J., and Kasper, J. C. (2003). Storm-time distortion of the inner
magnetosphere: How severe can it get? Journal of Geophysical Research: Space Physics, 108(A5).
Tsyganenko, N. A. and Sitnov, M. I. (2005). Modeling the dynamics of the inner magnetosphere
during strong geomagnetic storms. Journal of Geophysical Research, 110(A3).
Tsyganenko, N. A. and Sitnov, M. I. (2007). Magnetospheric configurations from a high-resolution

data-based magnetic field model. Journal of Geophysical Research: Space Physics, 112(A6):n/a–
n/a. A06225.
Tu, W., Elkington, S. R., Li, X., Liu, W., and Bonnell, J. (2012). Quantifying radial diffusion coef-
ficients of radiation belt electrons based on global mhd simulation and spacecraft measurements.
Journal of Geophysical Research: Space Physics, 117(A10):n/a–n/a.
Ueng, S.-K., Sikorski, C., and Ma, K.-L. (1996). Efficient streamline, streamribbon, and stream-
tube constructions on unstructured grids. IEEE Transactions on Visualization and Computer
Graphics, 2(2):100–110.
Ukhorskiy, A. Y. and Sitnov, M. I. (2012). Dynamics of radiation belt particles. Space Science
Reviews, 179(1-4):545–578.
Ukhorskiy, A. Y., Sitnov, M. I., Merkin, V. G., Gkioulidou, M., and Mitchell, D. G. (2017). Ion
acceleration at dipolarization fronts in the inner magnetosphere. Journal of Geophysical Research:
Space Physics.
Ulmer, C. D., Bayer, G. W., Choe, Y. R., and Roe, D. C. (2010). Scientific data analysis on
data-parallel platforms. Technical report, Sandia National Laboratories.
Walt, M. (1994). Introduction to Geomagnetically Trapped Radiation. Cambridge University Press.

Cambridge Books Online.
Wang, Y.-M. and Sheeley, Jr., N. R. (1992). On potential field models of the solar corona. The
Astrophysical Journal, 392:310–319.
135
Whipple, E. C. (1978). (u, b, k) coordinates: A natural system for studying magnetospheric

convection. Journal of Geophysical Research: Space Physics, 83(A9):4318–4326.
Willis, D. M. and Stephenson, F. R. (1999). Simultaneous auroral observations described in the

historical records of china, japan and korea from ancient times to ad 1700. Annales Geophysicae,
18(1):1–10.
Wiltberger, M., Qian, L., Huang, C.-L., Wang, W., Lopez, R. E., Burns, A. G., Solomon, S. C.,
Deng, Y., and Huang, Y. (2012). {CMIT} study of {CR2060} and 2068 comparing {L1} and
{MAS} solar wind drivers. Journal of Atmospheric and Solar-Terrestrial Physics, 83(0):39 –
50. Corotating Interaction Regions from Sun to Earth: Modeling their formation, evolution and
geoeffectiveness.
Wolfram (2016a). Cartesian coordinates (website).
Wolfram (2016b). Spherical coordinates (website).
Xie, H., Ofman, L., and Lawrence, G. (2004). Cone model for halo cmes: Application to space
weather forecasting. Journal of Geophysical Research: Space Physics, 109(A3):n/a–n/a.
Yu, Y., Koller, J., Zaharia, S., and Jordanova, V. (2012). L* neural networks from different magnetic
field models and their applicability. Space Weather, 10(2).
Zanetti, L. J., Mauk, B. H., Fox, N. J., Barnes, R. J., Weiss, M., Sotirelis, T. S., Raouafi, N.-E.,
Kessel, R. L., and Becker, H. N. (2014). The evolving space weather system—van allen probes
contribution. Space Weather, 12(10):577–581.
Zhao, X. and Dryer, M. (2014). Current status of cme/shock arrival time prediction. Space Weather,
12(7):448–469. 2014SW001060.
Zhao, X. P., Plunkett, S. P., and Liu, W. (2002). Determination of geometrical and kinematical
properties of halo coronal mass ejections using the cone model. Journal of Geophysical Research:
Space Physics, 107(A8):SSH 13–1–SSH 13–9.

Advanced Analysis and Visualization of Space Weather Phenomena PDF

Uploaded by

Copyright:

Available Formats

Advanced Analysis and Visualization of Space Weather Phenomena PDF

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Advanced Analysis and Visualization of Space Weather Phenomena PDF

Uploaded by

Copyright:

Available Formats

University of Colorado, Boulder

Advanced Analysis and Visualization of Space

Follow this and additional works at: https://scholar.colorado.edu/csci_gradetds

B.S., University of Maine, 2008

M.S., University of Colorado Boulder, 2013

A thesis submitted to the

Faculty of the Graduate School of the

University of Colorado in partial fulfillment

of the requirements for the degree of

Department of Computer Science

Prof. Xiao-Chuan Cai

Dr. Scot Elkington

Murphy, Joshua J. (Ph.D., Computer Science)

Advanced Analysis and Visualization of Space Weather Phenomena

Thesis directed by Prof. Xiao-Chuan Cai

1.1 A Brief History of Space Weather . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1

1.2 Connecting the Sun and the Earth . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3

1.3 Radiation Belts . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5

1.3.1 Trapped Particle Motion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5

1.3.2 The Adiabatic Invariants . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7

1.4 Analysis and Visualization of Particles in the Magnetosphere . . . . . . . . . . . . . 10

1.5 Numerical Models of the Earth’s Magnetosphere . . . . . . . . . . . . . . . . . . . . 10

1.5.1 Dipole Geomagnetic Field Model . . . . . . . . . . . . . . . . . . . . . . . . . 11

1.5.2 Empirical Geomagnetic Field Models . . . . . . . . . . . . . . . . . . . . . . . 12

1.5.3 Magnetohydrodynamic (MHD) Geomagnetic Field Models . . . . . . . . . . . 14

1.6 Thesis Format . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 16

2 Observational Uncertainties and CME Transients 17

2.1 The Model Environment . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 20

2.1.1 The WSA Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 20

2.1.2 The Cone Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 21

2.1.3 The Enlil Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 23

2.2 Mitigation of Uncertainties with Ensemble Modeling . . . . . . . . . . . . . . . . . . 24

2.3 Analyzing Effects of Uncertainty with CME Impact Domain Analysis . . . . . . . . 27

2.3.1 The 06 October 2013 CME Event . . . . . . . . . . . . . . . . . . . . . . . . 28

2.3.2 Defining an Impact . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 28

2.3.3 Finding and Populating the Impact Domain . . . . . . . . . . . . . . . . . . . 32

2.3.4 Solar Wind Impact Domains and Quartiles . . . . . . . . . . . . . . . . . . . 35

3 Adiabatic Invariant Methods and Error Analysis in a Dipole 45

3.1 Invariant Algorithms . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 45

3.2 Error Response in a Dipole Field . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 55

3.2.1 Error Response due to Field Line Footprint Inaccuracies . . . . . . . . . . . . 58

4 Working with LFM Model Gridded Output 63

4.1 The LFM Grid . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 63

4.2 LFM Grid Dipole Response . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 64

4.3 Algorithm Adjustments for LFM Data . . . . . . . . . . . . . . . . . . . . . . . . . . 71

4.3.1 The New K/B Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 73

4.3.2 Numerical Noise . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 76

4.3.3 Extremely Distorted Magnetic Fields and Search Progression . . . . . . . . . 80

4.3.4 Dealing with Multiple K Instances Per Line . . . . . . . . . . . . . . . . . . . 84

4.3.5 Final Updated Drift Trajectory Search Algorithm . . . . . . . . . . . . . . . . 85

4.3.6 Accuracy and Convergence . . . . . . . . . . . . . . . . . . . . . . . . . . . . 86

4.4 Application of Methods . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 88

4.4.1 Van Allen Probes L∗ and K Comparisons . . . . . . . . . . . . . . . . . . . . 92

4.4.2 LFM vs. LFM/RCM Full Field Comparisons . . . . . . . . . . . . . . . . . . 98

5 Parallel Operations and Timings 108

5.1 Serial Single Point Algorithm Timings . . . . . . . . . . . . . . . . . . . . . . . . . . 108

5.1.1 Single Point Computation - L∗ Overview . . . . . . . . . . . . . . . . . . . . 110

5.1.2 Single Point Computation - KB Line Search Routine . . . . . . . . . . . . . . 112

5.1.3 Single Point Computation - K/B Model . . . . . . . . . . . . . . . . . . . . . 114