-
Notifications
You must be signed in to change notification settings - Fork 0
/
research.html
197 lines (196 loc) · 10.5 KB
/
research.html
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
<!DOCTYPE html>
<html>
<head>
<meta charset="utf-8">
<meta name="viewport" content="width=device-width, initial-scale=1">
<title>Turibius Rozario | Research</title>
<link rel="stylesheet" href="styles.css">
</head>
<body>
<header>
<span class="header_ends"><a href="index.html">Turibius Rozario</a></span>
<nav>
<ul>
<li><a href="research.html"><b>Research</b></a></li>
<li><a href="UAVs.html">UAVs</a></li>
<li><a href="coursesActivities.html">Courses and Activities</a></li>
<li><a href="others.html">Other Projects</a></li>
</ul>
</nav>
<span class="header_ends"><a href="mailto:s175@umbc.edu">Email</a></span>
</header>
<div class="centered">
<h1>Research</h1>
<p>
I intend to pursue a PhD in Mechanical Engineering in the studies of
performance optimization,
process automation,
energy conservation, and
aeronautical systems.
Although I am yet to decide on a focus,
I am actively looking for summer research internships
(in academia and industry) in any of the above fields and
am currently in sustained lab during the academic year.
My current or previous research experiences are listed below.
</p>
<h2>Parameter Optimization | Dr. Ankit Goel</h2>
<h3>November 2021 — Present</h3>
<p class="background">
Background:
<a href="https://ankgoel.umbc.edu/">Dr. Ankit Goel</a>
is an assistant professor at UMBC.
He runs the Estimation, Control, and Learning Laboratory (ECLL)
in the Mechanical Engineering department.
</p>
<div class="fig">
<img
src="Images/Research-Goel-XOR.png"
alt="Plot of convergence of various training methods, in which
the Newton method outranks random search, which outranks gradient descent.">
Figure 1: Testing gradient free training methods
such as <code>fsolve</code> and random search method
against conventional gradient descent in approximating the XOR function
using a NN.
</div>
<p>
<!-- Under the Mechanical Engineering department at University of
Maryland, Baltimore County (UMBC), I implement various
optimization techniques, test them against existing solvers,
and aid in refining algorithms. This includes linear and
non-linear optimizers, and slicing novel optimizers into
machine learning frameworks such as Keras and Pytorch. -->
I implement and test various optimization techniques
to solve classical non-linear problems and train neural networks (NNs).
In particular, I have,
<ol>
<li>Used conventional machine learning techniques to train NNs
using both Keras and custom gradient descent,
</li>
<li>
Illustrated the ability of gradient <i>free</i> methods,
such as the Newtonian method and random search method,
to surpass conventional gradient descent in training NNs,
as shown in figure <a href="#gradientFree">1</a>,
</li>
<li>
Tested the effect on inertial measurement units
due to properties of accelerometers, gyroscopes, and magnetometers,
</li>
<li>
And implemented a novel finite time algorithm
in discreet problems such as NNs,
shown in figure <a href=#finiteTime>2</a>.
</li>
</ol>
</p>
<!-- <p>
One of my current projects is testing a novel algorithm that is
shown to have finite time convergence for all linear problems
and certain non-linear problems. In other non-linear problems,
it has demonstrated to converge at a much faster rate than
current optimizers such as stochastic gradient descent and
Adam, as shown in figure 1.
</p> -->
<p>
Additionally, I had the opportunity to write on my work
or present on my research:
<ol>
<li>
A Tutorial on Neural Networks and Gradient-free
Training (rejected, ACC 2023). <a href="https://doi.org/10.48550/arXiv.2211.17217">arXiv link</a>
</li>
<li>
URCAD — Presented poster on neural networks and gradient-free training on UMBC's undergraduate research day.
<a href="https://urcad.umbc.edu/abstract-2023/">Link to abstracts</a>
</li>
<li>
Time accelerated algorithms.
This was cancelled due to thresholding.
</li>
<li>
Modelling dynamical systems using neural networks. (In progress)
</li>
</ol>
</p>
<div class="fig">
<img
src="Images/Research-Goel-CostMNIST.png"
alt="Plot of convergence of various optimizers, in which
FTE outranks Adam, and Adam outranks SGD.">
Figure 2: Comparing training algorithm performance on MNIST dataset;
the finite time method was shown to outperform
conventional adaptive momentum and stochastic gradient descent methods
in this case.
</div>
<h2>Design of a Lab-Scale Ocean Wave-Powered Desalination System | Dr. James Van de Ven</h2>
<h3>Summer 2023 (REU)</h3>
<p class="background">
Background:
<a href="https://cse.umn.edu/me/james-van-de-ven">Dr. James Van de Ven</a>
is a Mechanical Engineering professor at University of Minnesota.
He runs the Mechanical Energy & Power Systems Laboratory (MEPS)
with a focus on fluid power.
</p>
<div class="fig">
<img
src="Images/Research-MEPS-Schematic.png"
alt="Illustration of a wave energy based desalination system.
The kinetic energy of a large flap, hinged at the sea bed, is
used to drive a hydraulic system. The hydraulic system uses
seawater to transfer power, generate electricity, and produce
freshwater.">
Figure 3: Illustration of the core processes and elements in
the desalination system.
</div>
<p>
A self-powered and decentralized wave energy converter and desalination system
was proposed in prior work by the MEPS lab. The wave energy is harvested using a large
oscillating flap hinged at the sea bed, whose kinetic energy is then transferred into
hydraulics; the pressurized seawater is used to generate electricity and freshwater.
Figure <a href="#simpleIllustration">3</a> showcases these primary traits.
</p>
<p>
My role over the summer was to help scale this full-scale system to a lab-scale
hardware-in-the-loop test system. With the help of my graduate mentor, Jeremy Simmons,
I:
<ol>
<li>Used fluid equations, efficiency computations, and dimension constraints to size
the system.</li>
<li>Identified parts required to construct the system and generated a bill of materials.</li>
<li>Designed custom parts and fittings, and a draft assembly model.</li>
</ol>
</p>
<p>
Figure <a href="#cornerCondition">4</a> compares the maximum
required pressure differential to the achievable pressure differential based on the energy losses
through a specific servo valve; the maximum required flow rate is compared against
the achievable flow of the pump. Performance evaluations such as this helped determine
whether the combination of the chosen parts would be capable of simulating the motion
of the wave energy converter.
</p>
<p>
Three deliverables were produced during this research experience:
<ul>
<li>Research poster, presented at SURE at University of Minnesota.
<a href="http://www.mrsec.umn.edu/_assets/pdf/REU-event-programs/SURE23.pdf">Link to abstracts.</a></li>
<li>Extended abstract, which can be read <a href="Documents/umn_reu_extended_abstract.pdf">here</a></li>
<li>Project hand-off memo, consisting of all research files and deliverables.</li>
</ul>
</p>
<div class="fig">
<img
src="Images/Research-MEPS-CornerCondition.png"
alt="X-axis is oil flow rate, y-axis is pressure differential across oil pump.
Plot consisting of three legends.
The first is the pressure differential available for the oil pump;
this curve starts at 4300 psi, and parabolically decreases.
The second is a cross indicating the system requirement.
This is located at the 3500 psi and 23 GPM location,
which is underneath the curve of pressure differential available.
The third is a vertical line indicating the maximum flow rate achievable (24 GPM).">
Figure 4: Maximum system requirements for pressure differential and flow rates compared
to the achievable pressure differential (4300 psi supply) and flow rate.
</div>
</div>
</body>
</html>