Skip to content
IJZIsaac Johan Ziebarth

UQ coursework · 2025 – 2026

Selected coursework

Nine pieces of coursework from my degree at UQ, newest first. Each says whether it was mine alone or a team effort, and what my part was.

Courses
AERO4450, AERO4470, METR6203, METR4202, METR4201, METR3100, ELEC4630, MATE7015
When
2025 – 2026
Tools
Python, MATLAB and Simulink, ROS 2, fastai, Fusion 360

The full write-up

Scramjet cycle with nitrous augmentation

AERO4450 Aerospace Propulsion · 2026 · Part 1 on my own, final report in a team of 2

A 1-D cycle analysis of a scramjet that burns part of its hydrogen with on-board nitrous oxide to keep thrust up at higher Mach numbers. The combustion part was mine alone: in Python I worked out the heat release of hydrogen burning with nitrous oxide, 164.6 MJ/kg. For the final report, where I did most of the work, we built the cycle model and matched the lecture's Mach 10 example to within 7.2% on thrust. Without nitrous the critical Mach number was 10.65; at Mach 13, burning about 11% of the hydrogen with nitrous oxide restored the brief's 0.5 thrust margin. It is a simulation of a fictional scenario.

Two plots against flight Mach number from 8 to 15: thrust margin and specific impulse, each with and without nitrous augmentation
Thrust margin and specific impulse against flight Mach number, with and without augmentation.

Reflected shock tunnel analysis

AERO4470 Hypersonics · 2026 · my own report on shots taken for our class group

An analysis of shots taken for our class group in UQ's Drummond reflected shock tunnel with a Mach 7 nozzle, at fills of 15 and 85 kPa. In Python I calibrated the wall pressure transducers in place, detected shock arrival automatically and measured shock Mach numbers of 3.676 ± 0.088 and 2.671 ± 0.047, carrying the uncertainty through every step. The measured plate pressures, 6.39 ± 1.15 and 12.39 ± 1.36 kPa, sat within one standard deviation of both the equilibrium and the calorically perfect predictions.

Two traces of plate pressure against time, one for each fill pressure, with the useful test window shaded on each
Pressure on the test plate at both fills, with the test windows marked.

Cold gas thruster against nozzle theory

AERO4450 Aerospace Propulsion · 2026 · lab in a group of 8

Firing a small cold gas thruster into back pressures from 1.35 to 81.35 kPa and comparing its thrust with isentropic nozzle theory, using the Summerfield criterion for flow separation. I did most of the analysis and wrote the notebook behind every figure and table, including the thrust-stand calibration (R² 0.995). Measured thrust ran 44 to 86% below theory, and repeat firings scattered by up to 0.8 N, more than the instrument uncertainty.

Measured thrust points with error bars well below the falling line of isentropic theory, against back pressure from 0 to 100 kPa
Measured thrust against isentropic nozzle theory across the back-pressure range.

Titanium crank arm for metal 3D printing

MATE7015 Additive Manufacturing · Sep – Oct 2026 · on my own · Fusion 360, static FEA, FDM printing in ASA

A left bicycle crank arm in Ti-6Al-4V that I designed in Fusion 360 for laser powder bed fusion. At the 1800 N test load my static study peaks at 408 MPa, a safety factor of 2.11, and the arm weighs 251.4 g, 21% lighter than my solid starting design. It is a design only: the titanium arm has not been built.

Crank arms are usually forged, so every length needs its own dies. Printed in titanium, crank length could be a single model parameter, so a rider with a leg-length difference could get a custom arm, if the design suits the process.

  • Crank length is one parameter of the model.
  • The arm is hollow and needs no supports inside: a closed internal cavity with 2 mm skins and ends 50° above horizontal, so the arm builds upright from the bottom bracket boss with no supports inside it. I chose the process partly for its fine resolution, which the thin skins need.
  • Two 3.5 mm holes let the unfused powder out. My first holes, in the top skin, raised the peak stress to 591 MPa, so I moved them to the side wall.
  • In my static study, at the 1800 N fatigue test load from BS EN ISO 4210-8, the peak von Mises stress is 408 MPa, a safety factor of 2.11 on the 860 MPa heat-treated minimum yield.
  • I printed 3 full-size ASA prototypes of earlier versions on my own printer. A 98.7% shrinkage setting brought the last one's bores within 0.3% of the CAD.
  • Over my weight target: at 251.4 g the arm is 76 g heavier than my stretch target, the 175 g Shimano arm.
  • Not built or tested: the titanium arm exists only as a design and a static study; no fatigue test has been run.
Section along the crank arm showing the closed internal cavity, its thin skins and its angled ends
Section along the arm: the closed internal cavity, its 2 mm skins and the ends angled 50° above horizontal, so it prints without internal supports.
Render of the final crank arm with its two powder-escape holes in the side wall
The final design rendered in Fusion 360 (not built), with the 2 powder-escape holes in the side wall.

Control design for three rigs

METR6203 Control Engineering 2 · 2025 · two team assignments and one on my own

On a gyro-actuated inverted pendulum I was responsible for the MATLAB and Simulink model, the LQR capture controller and integration, with Stateflow switching between an energy-based swing-up and balance. In simulation it swung up in about 1.1 s and was caught, but we ran out of time to get the swing-up working on the rig. On a Quanser 3-DOF helicopter I built the linearised hover model and the Simulink model that joins the observer, controller and plant, and the controller flew under joystick control. On my own, I designed a state feedback controller and observer for a linearised Whipple bicycle; in simulation its step response overshot by 3.4%.

Stateflow chart with three states, swing up, capture and swing down, and the angle conditions that move between them
The Stateflow logic I wrote for the gyro pendulum: swing up, capture once it is near upright, and swing down to retry.

METR4202 Robotics and Automation · 2025 · team of 4

A TurtleBot3 had to explore a map it had never seen and find the ArUco markers placed around it, with nobody driving. It maps with SLAM Toolbox and drives with Nav2. As the team's integration lead I wrote the exploration and recovery side: a frontier detector that scores each frontier by how much unknown space a ray cast says it would reveal, goal ranking and blacklisting, stuck detection with recovery, and a coverage monitor. I made 81 of the team's 101 commits.

Node diagram: SLAM Toolbox feeds the frontier detector and exploration monitor, which drive a Nav2 controller, alongside ArUco marker detection
How the nodes fit together: mapping, my frontier detector and exploration monitor, marker detection and Nav2.

Furuta pendulum balance controller

METR4201 Control Engineering 1 · 2025 · on my own, in simulation

Balancing an inverted pendulum on a motor-driven rotating arm. I derived the motor and gearbox model by hand, checked it in MATLAB, and placed a PID controller's zeros on the root locus. In simulation it met the brief's limits of 0.2 s peak time and 20% overshoot, with 0.040 s and 16.3%. An outer PD loop with a low-pass filter then held the arm at its set angle with no drift. It is a simulated design, on the course's nonlinear Simulink model rather than the rig.

Step response of the pendulum angle rising to its set point inside the shaded region allowed by the peak-time and overshoot limits
Simulated step response against the peak-time and overshoot limits.

Functional safety case for a LiDAR check

METR3100 Control System Implementation · 2025 · team of 5; I did the analysis and wrote the report

A fictional mine wanted a LiDAR check to catch partly dumped ore cars before they derail. Event trees over 249,660 cars a year showed the human supervisor alone still let 8 derailments a year through, and that the LiDAR needed 97.5% detection with a 0.018% false-flag rate, which is SIL 1. A 1oo2 voting design cut derailments to 0.2 a year but used all 365 unnecessary re-dumps the brief allowed. An STPA found 7 unsafe control actions and 16 loss scenarios, and I recommended a 300 ms command timing limit.

Event tree from 249,660 ore cars a year through the supervisor and the LiDAR, ending in yearly counts of re-dumps, unnecessary re-dumps and derailments
Event tree for the 1oo2 design over a year of ore cars. Steve is the human supervisor in the course's fictional case.

Road signs and rice leaf disease

ELEC4630 Computer Vision and Deep Learning · 2025 · on my own

For road signs I built a MATLAB bank of 3,268 templates across 7 sign shapes, found candidates with colour masks and edges, and scored each match on template overlap and chamfer distance; on the course's 10 or so test photos it found the yellow diamond signs more than 90% of the time. On a Kaggle rice leaf disease task in fastai, tuning took a ResNet26d from 88.1% to 97.0%, and my best ConvNeXt Large scored 98.387%, just short of the brief's 98.846% benchmark.

Loss against learning rate on a log scale, falling to a minimum near 0.1, with two suggested starting points marked
Learning-rate finder for the ConvNeXt classifier, with its suggested starting points.

Credits

Team entries say what my part was; the rest was my teammates'. Course briefs, rigs, the shock tunnel and the course models belong to UQ.