Turibius Rozario Email

Research

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.

Parameter Optimization | Dr. Ankit Goel

November 2021 — Present

Background: Dr. Ankit Goel is an assistant professor at UMBC. He runs the Estimation, Control, and Learning Laboratory (ECLL) in the Mechanical Engineering department.

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 fsolve and random search method against conventional gradient descent in approximating the XOR function using a NN.

I implement and test various optimization techniques to solve classical non-linear problems and train neural networks (NNs). In particular, I have,

  1. Used conventional machine learning techniques to train NNs using both Keras and custom gradient descent,
  2. Illustrated the ability of gradient free methods, such as the Newtonian method and random search method, to surpass conventional gradient descent in training NNs, as shown in figure 1,
  3. Tested the effect on inertial measurement units due to properties of accelerometers, gyroscopes, and magnetometers,
  4. And implemented a novel finite time algorithm in discreet problems such as NNs, shown in figure 2.

Additionally, I had the opportunity to write on my work or present on my research:

  1. A Tutorial on Neural Networks and Gradient-free Training (rejected, ACC 2023). arXiv link
  2. URCAD — Presented poster on neural networks and gradient-free training on UMBC's undergraduate research day. Link to abstracts
  3. Time accelerated algorithms. This was cancelled due to thresholding.
  4. Modelling dynamical systems using neural networks. (In progress)

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.

Design of a Lab-Scale Ocean Wave-Powered Desalination System | Dr. James Van de Ven

Summer 2023 (REU)

Background: Dr. James Van de Ven is a Mechanical Engineering professor at University of Minnesota. He runs the Mechanical Energy & Power Systems Laboratory (MEPS) with a focus on fluid power.

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.

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 3 showcases these primary traits.

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:

  1. Used fluid equations, efficiency computations, and dimension constraints to size the system.
  2. Identified parts required to construct the system and generated a bill of materials.
  3. Designed custom parts and fittings, and a draft assembly model.

Figure 4 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.

Three deliverables were produced during this research experience:

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.