Transforming Technologies

In the ever-evolving landscape of financial markets, the supply and demand of successful trading strategies has become the bedrock for many traders and financial institutions. As market conditions constantly shift, many existing models struggle to maintain profitability in unseen market environments. My project addresses this challenge by enhancing an automated trading system to not only predict market movements but be adaptive to unexpected and unseen environments. By integrating advanced machine learning and causal inference methods into automated trading models, the goal is to create a model that thrives in uncertain environments, maintaining profitability even when the market takes an unexpected turn. The result of this work being a robust trading algorithm capable of handling unseen market conditions with precision of which can contribute to the market stability and efficiency of which the financial system relies upon while also offering greater insight and understanding in the behaviour of financial markets.

Project by:

• Maxwell Bruce

Pupper is a friendly robot puppy whose passion is to be an emergency responder. Oh no! The kittens are surrounded by aggressive dogs. This looks like a job for Pupper. However, we need to give Pupper the correct equipment so it can succeed.  We have given our little friend an OAK-D camera to allow computer vision so it can see its environment. For Pupper to be aware of its surroundings we have given it a LiDAR, so it maps the environment. To make our friend extra smart, we have used a technique called behavioral cloning, it involves training a neural network to imitate expert behavior. Additionally, we have developed a multi-robot communication system allowing Pupper to work in a team. Our results show that Pupper and its friends have learned to find cats while steering clear of dogs, showcasing the potential for use in real-world emergency applications.

Project by:

• Khiya Barrett 
• Cho Ting Lee 
• Muhammad Rehan 

Fault-Tolerant Task Allocation and Load Balancing in a Dynamic Distributed Network. When we're pushing the boundaries of how much work a single machine can do, our next step is to have multiple machine work together simultaneously. However, challenges can arise in such a system due to the distributed nature of its independently operating components, in the form of unexpected machine failures, sporadic loss of communication paths, and so-on. We endeavoured to design and evaluate several approaches to coordinating tasks in a dynamic distributed system, aiming to reduce the effects of failures on the system's ability to complete tasks, while measuring the impacts of these approaches on its efficacy and efficiency. The results of this project include a discrete-time distributed service simulation framework, a set of techniques that can be implemented to improve the fault-tolerance of a distributed system, and an evaluation of these techniques when run on our framework.

Project by:

• Marika Colby 

The plastic recycling processes is crucial since plastic pollution not only a threat to the environment globally but also results in the loss of critical resources. Researchers have recently looked into the possibility of combining the reforming of plastics, especially microplastics, with photocatalytic hydrogen production to simultaneously solve environmental and energy concerns. However, there are some limitations, such as the low efficiency of Photocatalyst due to the surface area, reaction condition (light intensity, temperature and pressure condition etc.). As a result, this project is to take on action as to develop some catalysts, especially, Carbon Nitride(CN) to achieve the better efficiency of Photocatalyst.

Project by:

• Ngoc Huy Nguyen 

The field of process engineering relies on safety considerations and not wasting money methods. Level control ensures high level of safety in dealing with liquids in tanks, and maintains cost effective approaches. Teaching the students the importance behind implementing control techniques is essential to their future careers as process engineers. Therefore, our project aims to deliver that knowledge in a user-friendly and interactive way of learning.

The method used in our approach to teaching third year chemical engineering students, is redesigning the current experiment. Moreover, it will be compliant and more specific to course structure. This will ensure best learning outcomes compared to the theoretical content they have been exposed to.  

The outcome is a new experiment relevant to the course content, that will emphasise the understanding of the importance of level control and approaching correct methods. In addition, the lab script is designed to help the students in an engaging way.

Project by:

• Khalifa Ahmed Khalifa Ahmed Alremeithi 
• Abdulla Mohamed Ahmed Sulaiman Almahri 

The aim of my project is to explore how continual machine learning can be used to effectively manage data drift in temporal datasets. Data drift occurs when the patterns in a dataset change over time, making traditional machine learning models less effective. The Australian Rail Track Corporation (ARTC) provided a dataset that exemplifies these challenges, with shifts in data over time impacting the ability to predict rail breaks, whilst also being a heavily imbalanced problem with few failure cases. The project experiments with different scenarios of Continual Learning by delaying the availability of the truth labels for the failure cases and restricting how much data is available to the model when training. By applying this method to the RailBreak dataset, the project aims to show that continual learning is an effective tool for dealing with data drift in real-world scenarios.

Find out more - project video

Project by:

• Neil Mazumdar 

Slow internet could be a thing in the past with terahertz technology. By developing advanced antennas to enhance high-speed wireless communication using terahertz waves. They can transmit large amounts of data rapidly and are useful for applications like high-resolution imaging and non-contact security screening. To meet the demands of this emerging technology, we are designing two innovative antennas: a "phase-correcting lens antenna" and a "dual-reflector Cassegrain antenna".

Using specialised electromagnetic simulation software, we model and optimise these antennas to achieve high performance. The phase-correcting lens antenna provides compact integration and beam-steering capabilities, while the dual-reflector Cassegrain antenna is designed for high gain, crucial for long-distance communications.

After finalising the designs, we fabricate and test them to measure key performance parameters such as gain, radiation pattern, and reflection coefficient. The outcome will be antennas optimised for next-generation terahertz communication systems, supporting faster internet and improved precision in sensing and imaging applications.

Project by:

• Tin Chu Ng 
• Junwei Liang 
• Chuntian Zhang 

Our project investigates how data from mobile phones, particularly location data, can be used to uncover critical insights about how facilities are used. By analysing data from shopping centres, we aim to understand how patterns of movement and behavior can reveal important details about operations, such as where people gather, how often certain areas are visited, and the roles of individuals within these spaces. We use statistical analysis and machine learning to detect these patterns and evaluate the potential risks they pose. The outcome of this project will help us identify vulnerabilities and develop strategies to improve security and efficiency in such facilities.

Project by:

• Matthew Parsons 

Would you trust a diagnosis from an AI doctor? How comfortable would you feel in a self-driving car? The emerging use of artificial intelligence has had a profound impact on our lives, but many people are hesitant to put their complete trust in this new technology. Despite state-of-the-art performance, AI models are still prone to making overconfident predictions, which can have disastrous outcomes in high risk settings. Without properly quantifying the uncertainty in a prediction, it is impossible to have trustworthy and reliable AI. One way to overcome this is through calibration, which ensures that predicted probabilities are equal to observed frequencies. For example, a weather forecaster is well calibrated if for all the times it predicts 70% chance of rain, it actually rains 70% of the time. My research focuses on measuring calibration using nonparametric kernel methods, providing a mathematically rigorous answer to AI's trust question.

Project by:

• Peter Moskvichev 

Australian agriculture faces growing challenges like labour shortages, safety concerns, and inefficiency of manual tasks. While autonomous machines have been adapted to the industry, they often struggle in dynamic environments. They are required to be  agile yet efficient in handling various tasks.  Our project aims to tackle these challenges by developing an autonomous robotic system capable of navigating unknown environments and performing tasks such as object retrieval. We integrate a robotic arm on an autonomous rover that can move on uneven surfaces and pick and place objects. We implemented the latest robotic framework and algorithms for path planning, obstacle avoidance, and arm control. The robot uses sensors like LiDAR and cameras to create real-time maps of its surroundings, allowing it to plan efficient routes while avoiding both static and dynamic obstacles. We have thoroughly tested our robotic system in computer simulations before moving to real-world robots.

Project by:

• Jia Jun Ho
• Wenqi Lyu
• Hanyuan Ma 

Superconducting quantum interference devices (SQUIDs) are expected to generate electrical voltages in response to changes in the electromagnetic environment, and hence can be used to measure magnetic fields very precisely. The aim of this project is to experimentally verify this theoretical behaviour, at low temperatures. We verified the superconducting behaviour of SQUIDs by measuring a resistance-less current (i.e. zero voltage even in the presence of a current). Additionally, we verified the response of SQUIDs to magnetic fields, by observing a sinusoidal voltage response. SQUIDs provided by QuANTeG were cooled to -268  °C using liquid helium. Experimental data matched expectations, confirming the quantum behaviour of SQUIDs.

Project by:

• Declan O'Callaghan 

From plankton to plants and people, the evolution of multicellular organisms has played an undeniably important role in shaping life on Earth. Despite its importance, much about what conditions are required such that early multicellular organisms are capable of participating in evolution is still a mystery. The goal of this project is to analyse evolution within models of early multicellular populations in order to uncover parts of this mystery. Simulations have been performed to determine the behaviour of populations with certain traits in a specific environment and for quicker analysis, deterministic models have been employed to estimate the expected behaviour. Evolution is then able to occur through allowing a mutation of these traits to occur in an individual. Analysing this individual's behaviour reveals if the new trait will spread through the population or die out.

Project by:

• Daniel Fraterman 

As the world of underwater exploration develops; diverse capabilities are needed. The domain of application-specific Autonomous Underwater Vehicles (AUVs) is growing with potential applications in defence, science, and mining. This project designed and built an AUV that mimics cuttlefish undulating motion and can operate autonomously. Undulating motion has benefits over traditional propeller propulsion including; less tangling risk, lower sonar signature and less impact on the marine environment. Stability, controllability and waterproofing were determined as the most important aspects for AUV design and thus underpinned this project. A new 3D-printed hull was designed, alongside modularised code to manage the craft's communication, automation and movement, allowing this iteration of the AUV to be easily replicated and extensible. Improvements over previous iterations were also made in employing waterproof servo motors outside the main hull to reduce water ingress. Metasepia provides a strong foundation for the future development of bio-inspired AUVs.

Project by:

• Callum Menadue 
• Giuseppe Grasso 
• Matthew Beahan 
• William Craddock 

How can we uphold safety when maintenance work is performed in an industrial process?

Controls and isolations are required to eliminate and manage hazards and their risks. Permit to work systems are used for the planning, preparation and isolation of site activities to maintain safety. Process safety is essential for preventing major accident events and mitigating consequences. Major accident events have been caused by technical or human failures, with many arising due to safe work failures. Failure causes may result from poor practices due to complacency without incident which can escalate into safety critical equipment failures resulting in a major accident event. Permit to work systems are a critical process safety management element that when implemented correctly can reduce the probability of such an escalation occurring. Formulating metrics to measure the ongoing effectiveness of the permit to work systems will contribute to proactively maintaining an effective permit to work system.

Find out more - project video

Project by:

• Lucy Thiselton 

The Adelaide University Motorsport Team (AUMT) race against universities from around Australia with a formula-style vehicle and are striving for the next breakthrough in vehicle performance. A carbon fibre composite chassis has the potential to significantly save weight amongst other performance improvements, enabling AUMT to challenge the top teams at the annual competition. The aim of the project is to design and build a composite chassis for AUMT, enabling AUMT to continue to build composite chassis in the future. Existing composite materials and chassis have been researched, theoretical analysis and structural simulation has been conducted for initial panel selection, before iterative experimental testing of panels was conducted to refine the design and enable the chassis to be manufactured. We present a fully manufactured carbon fibre composite chassis that AUMT can use for their 2025 vehicle, which has the potential to provide increased vehicle performance and improve AUMT's competition placing.

Project by:

• Hannah Dinning
• James Mattiske 
• Harvey Michels 
• Saatvik Rudrapatna 

When an autonomous robot enters an unfamiliar scene, it needs to understand the surrounding area to complete its task. This cannot always be preprogrammed, so the robot needs a way to comprehend scenes in real time using visual context. Our goal is for our robots to understand scenes in real time, using machine learning to learn object relationships for scene comprehension. To support this, we designed an autonomous navigation system and implemented an object detection pipeline. By processing detected objects, the robot develops a database of objects from which it can comprehend scenes. Our demonstration shows robots exploring and comprehending a scene to complete a search and rescue mission.

Project by:

• Daniel Webber 
• Adam Rice 

Does having more engines benefit an aircraft, particularly in lift generation? Distributed propulsion systems (DPS) involves dispersing an aircraft's propulsive forces about its airframe, typically by using additional engines. The aim of our project was to investigate the extent in which distributed propulsion can enhance the lifting ability of an aircraft wing. We performed an investigation into relevant literature to develop a single-wing design utilising DPS. The concept was manufactured with 3D printed wing segments and internal aluminium supports. Several preliminary experiments were conducted to ensure the structural integrity of the wing and determine the thrust production of the propulsion system. Our model was placed into a wind tunnel to simulate real life application. All forces produced by the wing were recorded using a load cell and analysed to compare against current technology to determine the effectiveness of our design.

Project by:

• Austin Gogel 
• James Prior 
• Kynan Skein 
• Sabine Molloy 
• Sebastian Caruso 

What if we could slash the travel time between Adelaide and Sydney from 2 hours to 40 minutes. This crewed hypersonic aircraft is designed to achieve just that!  

Using Computer-Aided Design software, the team has postulated a conceptual design for an aircraft that can travel up to five times the speed of sound. Using state-of-the-art fluid simulation software, a combined turbojet and ramjet engine was calculated and designed. This engine configuration allows the aircraft to operate from take-off to hypersonic speeds without rockets. Flying at fast speeds induces high heat and pressure on the aircraft's surface. The team has simulated and defined areas of the plane that experience large pressure loads and extreme levels of heat. This research and testing will inform aircraft manufacturers to choose appropriate materials for hypersonic aircraft.

So why not come and have a look at the new crewed hypersonic aircraft? It will be out of sight before you know it!

Project by:

• Neel Puri 
• Louis Nitschke 
• Thomas Lightfoot 

Our project is developing an adaptable, autonomous olive harvesting system tailored for the Australian Super High Density (SHD) farming environment, inspired by successful models in Chile and the USA. We've made significant strides in designing and modelling key components, including the frame, beater, and conveyor subsystems, with initial testing showing promising efficiency and adaptability. This cost-effective, tow-bar attached design aims to address the lack of suitable solutions in Australia, enhancing productivity and mitigating Labor constraints. Future work will focus on system refinement and field testing, in collaboration with the University of Adelaide and agricultural stakeholders, to ensure the solution meets the needs of olive farmers and boosts the sustainability and competitiveness of the Australian olive industry.

Project by:

• Jemma Francis 
• Bayden Stanley 
• Edward Tilbrook 
• Sam Button 

Traditional fixed-wing aircraft are unsuitable for pollinating Australian crops, such as greenhouse tomatoes, due to their large size, poor manoeuvrability, and high energy consumption. In contrast, flapping-wing micro aerial vehicles (FWMAVs), inspired by insect flight, can manoeuvre through greenhouses with minimal disturbance to the surrounding environment. Our project focuses on optimising a two-winged FWMAV for agricultural purposes.

Despite their potential, developing FWMAVs as unmanned aerial vehicles is challenging due to the complex and poorly understood nature of flapping wing flight. Our research presents that a two-winged FWMAV for plant pollination is feasible, offering reduced mechanical complexity and lower environmental impact than alternative FWMAV designs. By exploring the relationship between insect flight mechanics and wing optimisation, we aim to enhance the energy efficiency, manoeuvrability, and load-carrying capacity of FWMAVs. Ultimately, our goal is to build an FWMAV capable of stable hover, pushing the boundaries of bio-inspired aerial innovation and sustainability.  

Project by:

• Melarn Murphy 
• Max Buttignol 
• Kosta Dimitropoulos 
• Jack Kelton 

The human brain is understood to be the most complex organ, capable of performing intricate functions such as computing, memorising and learning. At the most basic mechanistic level, these complex processes rely on the unique ability of brain's neurons 'biological microprocessor units' to communicate and transfer information to each other through electric potential signals generated by precise regulation of ions across membrane nanochannels in their synapses' junctions between adjacent neurons. Despite the apparent simplicity of this electrical energy generation mechanism, to mimic and implement this ability of neurons' ionic nanochannels in functional synthetic membranes remains challenging. As such, research attention has been paid towards the engineering of biomimicking artificial devices to sustain the growing requirement of computing power on a hardware level, harnessing fluidic ion transport mechanisms to imitate synaptic signals in the human brain. Successfully doing so will unlock faster and stronger computing power by combining physical memory and processing components into one unit.

Project by:

• Charles Lord 

Cherry production efficiency has greatly improved with the introduction of agricultural automation. Mobile Elevated Work Platforms (MEWPs) automate the pruning process by mechanically lifting orchard workers to the treetops, reducing the need for labour-intensive and hazardous ladder use. Unfortunately, Adelaide Hills cherry orchards often have up to 57% terrain gradients, which exceeds the safe and efficient operating capabilities of current commercial MEWP options. The goals for this project were to design an affordable MEWP capable of efficiently operating on severely sloped terrain, with self-levelling features for enhanced operational safety. Throughout the project, a solution was developed that addressed both the structural design and control approach. Significant analysis and simulation were performed to support design process and ultimately reduce risk for a future functional prototype. While further verification and validation is required, a MEWP has been designed with the potential to be a viable option for Adelaide Hills cherry producers.

Project by:

• Adam Slimming 
• Ahmed Shehata 
• Benjamin Cornish 
• Jackson Tilley 
• Zachary Martin-Peddey 

The field of terahertz engineering aims to exploit the so called  'terahertz gap' - the underutilised frequency band between microwave and optical frequencies in the electromagnetic spectrum. This project aims to adapt the method of transformation optics to the terahertz regime. A platform for terahertz integrated circuit design and well established methods for designing components on the platform would help to establish the field of terahertz as a more viable solution to modern problems in communications, medical technologies, and imaging, transitioning the technology from ad hoc laboratory setups into mass fabricated products. This project proposes that the adaptation of transformation optics to the terahertz regime will add a great degree of design flexibility to wave manipulation. This project has made use of transformation optics design techniques to design and simulate the performance of an integrated terahertz device. Use of similar design flows may be useful for future terahertz component designs.

Project by:

• Taine Murray 

This project explores the application of advanced artificial intelligence (AI) techniques, particularly machine learning (ML), to enhance the processing and analysis of radio frequency (RF) signals in complex electromagnetic environments. The primary aim is to evaluate the efficacy of these novel AI approaches in deinterleaving RF signals and detecting anomalous pulses, comparing their performance to current signal processing standards.

The methodology involves developing ML models capable of clustering similar pulses and tracking their evolution over time. We preprocess Pulse Descriptor Words (PDWs), extract relevant features, and train unsupervised learning algorithms to group pulses based on their similarities. These models are then tested using real-world datasets that incorporate noise and other environmental effects such as multipath.

Project by:

• Lennox Avdiu 

There is an increasing need for the testing of materials after their manufacture to ensure they are safe and without defects. With good penetration through many dielectric materials and sub millimetre resolution, terahertz waves can be utilised to image many internal defects. However, conventional two-dimensional imaging is limiting, thus we turned to the power of holograms. The aim of the project is to determine if the 3D imaging technique of digital holography in the terahertz spectrum is a viable test option to view internal defects of fibreglass samples. An optical system, utilising 3D printing has been designed to capture holograms, as well as code created to raster scan and create 3 dimensional reconstructions of the objects under test. We experienced several technical challenges, such as suboptimal camera performance at our operational frequency, and unexpected results when capturing images. Despite this, we have made significant progress towards achieving the aim.

Project by:

• Aaron Hamilton 
• Sebastian Cox 

The ball bearing motor has been a mystery since it was first discovered. To this day, there is still no consensus on how the ball bearing motor is able to generate torque. This phenomenon has been dubbed the Huber effect. There are three existing theories for how the ball bearing motor is able to spin though none of them have been fully proved through experimentation.

The goal of this project is to investigate various properties of the ball bearing motor and ascertain whether any of the existing theories are viable explanations for the Huber effect. This investigation involves modelling the ball bearing motor or a simplified version of it in COMSOL to simulate the electromagnetic interactions within a ball bearing motor. Additionally, physical models of the ball bearing motor will be made and experimented on.

Project by:

• Rui Yan
• Deyan Sarac
• Atef Zadeh 

Imagine you had a picture of someone's face or some other style of picture, how would you see their whole body or everything around them? This is where my project comes into play! This project is a tool to take a picture of that face and visualise what the rest of the person looks like and what kind of background they might be in. You can also upload your face to customise a version of yourself with a different body. This application studies many images to learn what people and their surroundings usually look like. Then, when it sees just one face, it guesses and draws the rest of the picture. You'll get a complete picture, including the body and background, even if you only have one face at first.

Project by:

• Yukun Chen 

Our project aims to improve the accuracy and reliability of testing methods used to analyse how brittle materials behave under pressure, particularly during an unwanted failure known as "snapback." Snapback, a reversal of displacement due to unstable energy release, presents challenges in structural engineering. Our goal is to develop better testing methods that accurately predict and control it, providing engineers with a more reliable understanding of materials to design safer structures.

To tackle this issue, we conducted experiments using three tests: the three-point bending test, the Brazilian disc test, and the uniaxial compression test. These tests simulate conditions under which cracks form and propagate in brittle materials. We used tools like digital image correlation systems and acoustic emission sensors to capture detailed data on the stress-strain behavior during these tests.

Project by:

• Siu Chun Yeung 
• Ming Gong 
• Man Lee Wong 
• Yuen Lung Wong 

Autonomous robot swarms have the potential to transform many industries through advancements in AI and robotics. This project addresses key challenges, such as enabling drones to locate each other within a swarm and maintain stable flight paths despite disturbances. The work establishes a foundation for systems made up of multiple robots. Initially, a method was developed in MathWorks, allowing for the creation of control algorithms that manage attractive and repulsive forces based on the distances between drones. The simulation was then advanced to a Linux platform using ROS, a software framework that supports the future integration of these algorithms on real-world rovers and drones. This research, done alongside industry-sponsored PhD projects, helps study how groups of robots work together in different situations, like underwater tasks and communication between robots. These efforts bring us closer to a future where groups of autonomous robots play an important role in technology and development.

Project by:

• Eric Wei

There exists much academic debate over a question that seems to be a simple physics-based puzzle: if a frictionless rail car filled with water develops a perfect vertical leak in an off-centre position, will the car move forward? Or backward? This unknown effect is difficult to observe amongst other real-world phenomena, hence making it challenging to develop effective experimental techniques to settle this debate. However, through investigating the theory behind this phenomena we are offered a deeper understanding of all forces and concepts involved. From there, experiments testing different design parameters isolated in simulation have been performed to validate the theoretical findings. In doing this we have been able to test our simulated results in person to finally determine the solution to the leaky tank mystery.

Project by:

• Eric Tsoukatos 
• Michael Stefani 

Most people had heard of what drones are, a small flying machine. Various type of those machines were used in farming, rescuing or surveying an area with build in camera, and more. Traditional drones use blades as a form of propulsion, but it is often loud, dangerous with the blades, and unable to do sharp turns. This is where flapping wings drone, also know as ornithopters, came in. It is quite, lightweight, safe and mimicking bird allowing it to make sharper turns. However, with an unorthodox method of propulsion, came a challenge in controlling it. The project focuses on developing a control system that allows the drone to stably move semi-autonomously. With this system we will be able to open up a relatively uncharted area of areal vehicle, allowing a smoother, safer and less disrupting drone to assist in many different field.

Project by:

• Low Teck Han 
• Minshan Zha 

Our project is about understanding how sandwich panels bend or change shape when pressure is put on them. Sandwich panels are used in things like airplanes, cars, and buildings because they are lightweight and strong. But it's important to know exactly how they bend and move so that engineers can make better and safer designs.

To do this, we use computers to create simulations, which are like super-detailed digital models. These simulations show us how the panels behave under different pressures and conditions, helping us predict how they'll act in the real world. Instead of guessing, we use precise math and coding to get accurate results.

In the end, our project helps engineers design better, more reliable panels. This can make planes lighter, cars safer, and buildings stronger, all by understanding how sandwich panels bend and flex.

Project by:

• Chun Yip Tam
• Nicholas Bigg 
• Aman Surani 

Autism Spectrum Disorder (ASD) affects 1 in 25 children in Australia, with many experiencing significant challenges in communication and social skills. To address these issues, our project developed an AI-powered platform designed to enhance social skills and mental health support for children with ASD. Leveraging Machine Learning (ML) and Natural Language Processing (NLP), the platform creates personalised AI personas that engage children in meaningful, tailored conversations. These personas act as interactive friends, helping children navigate social interactions, express emotions, and build communication skills. The integration of dynamic AI characters, inspired by popular figures, makes interactions both educational and therapeutic. Gamification features, including a rewards system, are implemented to boost user engagement and motivation, enhancing the effectiveness of interventions. The outcome is an innovative tool that empowers children with autism to connect, learn, and thrive in a supportive digital environment, offering a new approach to tackling real-world challenges in autism care.

Project by:

• Vidit Patel
• Vaibhav Aggarwal
• Shaan Sadanand

Air pollution significantly impacts public health and the environment, but understanding the causal relationships between pollutants and weather conditions remains challenging. This project investigates the causal link between pollutants such as carbon monoxide levels (CO) and temperature (T) in air quality time series data. Using advanced machine learning techniques, such as Variational Autoencoders (VAE), Invariant Variational Autoencoders (iVAE), and Independent Component Analysis (ICA), we uncover hidden structures and dependencies between these variables. By applying these models, we analyse the data to detect causality and improve air quality predictions. The results show potential insights into how temperature fluctuations influence CO levels, with implications for environmental policies and predictive models in urban settings. This project demonstrates the power of modern AI tools in solving real-world environmental problems.

Project by:

• Xuanming Feng 

Advancing ecological monitoring, "Snail Trail" leverages the power of YOLOv9 and the forthcoming YOLOv10 models to enhance the precision of small snail detection in natural settings. This project tackles the unique challenges of tracking small, dynamic creatures, where traditional models falter due to size and environmental camouflage. Our approach enriches dataset diversity and trains the models to discern subtle movements and features, crucial for accurate ecological assessments. We've refined our algorithms to reduce misdetections, crucial in complex backgrounds where snails exhibit erratic movements. Our findings aim not only to push the boundaries of object tracking technology but also to provide essential tools for biodiversity conservation. By optimising these advanced detection models, we enhance their applicability in real-world ecological monitoring, offering insights into subtle ecological dynamics previously obscured by technological limitations.

Project by:

•  Ziyi Liu 

Snails play an important role in ecosystems, yet their behavior and lifestyle remain largely a mystery. Our project uses cameras and smart tracking technology to observe the movements and habits of snails in their natural environments. By recording their activities, we can better understand their way of life and how they impact the plants and animals around them. This project helps scientists protect snail habitats more effectively, maintain biodiversity, and promote ecological balance, while raising awareness of the important role these "slow movers" play in nature.

Project by:

• Yuanyuan Fu 

Counting plants and animals on large farms is difficult and time-consuming. Farmers need to know how much crop they are growing, or how many sheep and cattle they have on their fields. Manual calculations can lead to errors and waste valuable time. Our project aims to help farmers by developing an intelligent system that can automatically count and track plants and animals.

We created a special program to view pictures and videos of the farm using a camera and computer. This program can identify and count objects such as plants in a field or animals in a pasture. It works much like how our eyes see, but it can be done faster and without fatigue.

The result of our project is a practical tool that tells farmers the exact number of plants or animals they have. This saves them time and energy, allowing them to focus on other important tasks. It makes farming easier and helps ensure food production is more efficient and feeds everyone.

Project by:

• Yilei Mei 

In everyday life, we use pictures to understand the world around us. But how can a computer figure out what's important in an image? My project is all about exploring those techniques that teach computers to focus on the right parts of a picture. Imagine you're looking at a picture of a dog in a park. Instead of letting the trees or sky distract it, I aid the computer in "seeing" the dog.

I use something called 'Class Activation Maps' (CAMs), which act like a spotlight to show the important parts of an image. This can be used in many ways, like helping doctors understand medical images or even helping computers write captions for photos. By improving how well these 'spotlights' work, my project is all about exploring different CAM methods for better understanding.

Project by:

• Malik Muhammad Arslan 

Atrial fibrillation (AF) is a major global health issue, affecting millions, including over 500,000 Australians, and increasing risks for stroke, dementia, and death. Current AF treatments vary significantly in effectiveness due to the complexity of the disease, highlighting the need for personalised interventions. The concept of a "cardiac digital twin" offers a promising solution. This advanced computational model simulates an individual's heart, allowing precise risk assessment, detailed understanding of disease progression, and accurate prediction of treatment outcomes. However, developing these digital twins faces calibration challenges, where model parameters need fine-tuning to match the patient's actual heart function. This project proposes using "AF dynamics scores," derived from routine clinical data, to enhance the calibration process, improving the accuracy and efficiency of digital twins. By leveraging these scores, the initiative aims to revolutionise AF treatment personalisation, paving the way for advancements in precision cardiology.

Project by:

• Huu Hung Quang Nguyen

Have you ever read a story written by a computer? While computers can write stories, they sometimes aren't very good at making them interesting. Large Language Models (LLMs) hold great potential in automating creative tasks like storytelling, but they often struggle with key elements such as plot coherence, character development, and emotional depth. This project explores how LLMs can be used to generate high-quality stories by addressing these limitations. By using a combination of generative models, we aim to improve the richness and depth of AI-generated stories, making them more comparable to those written by humans. The findings can contribute to enhancing the quality of AI-generated stories, with applications in entertainment, education, and other creative fields.

Project by:

• Xiangxiao Meng 

My project aims to enhance the way AI agents, such as drones, communicate and collaborate in a framework called Multi-Agent Reinforcement Learning. In this setup, multiple drones work together to complete tasks like navigating through forests and putting out fires. The real-world challenge is that these drones have limited information and visibility, so they must share critical details, such as the location of fires or obstacles, to make better decisions as a team.

The methodology involves designing an adaptive communication protocol where drones learn when and what information to share. Using reinforcement learning, they adjust their communication strategies based on the environment and task demands. This system continuously improves through feedback, enabling more effective coordination.

The outcome of this project is a smarter, adaptive communication system that boosts the efficiency and teamwork of drones in complex, dynamic environments, demonstrating how enhanced communication can significantly improve the performance of MARL systems.

Project by:

• Abdul Kareem 

Cryptocurrency is a type of virtual money which can be traded online and used to pay for things. Just like the stocks, you can buy cryptocurrencies online, but they not belong to any companies and banks, and the value of cryptocurrencies can change a lot due to market trends and news. People always buy cryptocurrencies just for earning money with chance. Our projects investigate some algorithms that can make cryptocurrencies traded with program, and get the highest profit at same time. We search different kinds of algorithms, and test them on historical data to evaluate the profit. A good algorithm should make users earn much money. Our goal is finding and writing such kinds of algorithms, it will trade automatically and try to get highest return. If the algorithms are good enough, we may be able to earn money when we sleep.

Project by:

• Yuhan Xia  
• Ruize Wang 

Deep learning has achieved significant progress in tasks like image classification, which is critical for applications such as facial recognition and medical diagnostics. However, this progress depends on large, high-quality datasets, which are expensive and time-consuming to collect. When data is limited, deep learning models often struggle to perform well.

My project explores whether AI-generated images, using models like Stable Diffusion, can replace real images for data augmentation. By generating new images with AI, this project aims to increase the diversity of training data and improve the accuracy of classification models, and explore if AI-generated images can make deep learning models more effective, reduce the need for costly data collection, and improve performance when real-world data is scarce.

Project by:

• Mengyao Lin 

Software simulations are a useful way to test real-life ideas without expending significant resources. Big simulations need to share tasks, with many programs working together to reach the same goal. However, talking between programs can become difficult, especially when adding new features. For example, adding a new animal to a zoo simulation might require large changes to logic. Our project examines a new way of handling communication called High-Level Architecture (HLA). It makes it easier to add new things to our simulation by having all communication go through one main program, instead of each entity talking to each other. We used this to improve an existing tank battle royale game, called Tank Royale, where players compete to code the best tank and be the last one standing. Our project improves the original game by making it easier to extend and serves as a proof of concept for future HLA projects.

Project by:

• Peter Yeoh 
• Heath Rampazis 
• William Moore 

The current insurance claim process is complex, convoluted and confusing. Currently, property managers lodge claims on behalf of their clients, but oftentimes the process varies wildly for each combination of insurers, building inspectors, property managers and other involved parties. As such, it's often difficult to understand what information is required and what property owners actually need to do, leading to delayed claim approval. Our project is to create a system that will revolutionise how home insurance claims are handled, simplifying the process for everyone involved. By pulling readily available data from open API endpoints, we can automatically fill in a large amount of the data required when filling out claims, reducing the stress and work required for all.

Project by:

• Nathan Do 
• Nico Vidal 

As the global population grows and concerns about food safety and animal welfare increase, the livestock industry faces the dual challenge of improving production efficiency and animal welfare. We aim to help farmers manage their livestock more effectively using smart technology. Traditional livestock management is inefficient and cannot provide round-the-clock monitoring. We solve this problem through video analysis technology. We use artificial intelligence and large language models to analyse the behavior and health of animals in videos. This acts as a smart monitoring tool for farmers, allowing them to quickly identify abnormal behaviour, such as illness or animals that need special care. Early results show that the system can identify animal behaviour and speed up the response to animal conditions.

Project by:

•  Zhenhao Li 

Imagine having a tiny robot buddy that can zip around your home, listen to you, and even see things like a super detective. This special friend uses its 'superpowers'—listening, talking, and seeing—to help you out. For example, if you ask, "Is there an apple in the kitchen?" it can scoot over, take a look, and tell you all about it.

We built this robot to be a great helper around the house. It uses smart tools to understand what you say and to see clearly, so it can find things and answer your questions. We teach it the best paths to move around without any bumps or crashes.

Now, our robot can chat with you, find things by looking at them, and even guide itself around your home safely. It’s like having a tiny, talking guide dog that’s also a whiz at spotting things!

Project by:

• Yuchen Wang 

Supplying content to users over a network in an efficient and reliable manner is challenge faced by Content Distribution Networks (CDNs). However in their current state CDNs have serious issues. The underlying connection methods suffer high overhead and the networks have high complexity. Additionally they cannot provide an assurances of the origin of content, how it has traversed the network, and that the content was actually delivered to a user. Given these issues a new type of Content Delivery Network was designed and simulated. The design is based on Named Data Networking, where data is delivered intuitively based on its name, instead of predetermined network locations. Cryptography is used to prove the data’s origin, who can access it, and track it through the network from request to delivery. The design shows that this type of network is feasible whilst also providing greater simplicity and efficiency, which is reflected in the simulations.

Project by:

• Karl Asenstorfer 

The primary objective of every brewery is to brew great tasting beer. As such, brewers need to implement strategies that prevent or limit potential contaminants from affecting their beer. Hops are a key ingredient in brewing and are highly sensitive to being damaged by oxygen. Therefore, oxygen is considered a contaminant during the late stages of brewing, as contact with oxygen can produce off-putting flavours and aromas. Hence, minimising the oxygen exposure during late-stage brewing stages such as dry hopping and kegging will prevent the production of off flavours/aromas and result in a better tasting beer.

Project by:

• Maxine Tsoukatos 
• Julia Damato 
• Georgio Tsipanitis 

Large language models (LLMs) have the ability to automate the generation of fuzz inputs to improve software vulnerability discovery. LLMs have shown impressive capabilities in natural language processing tasks, assisting in the analysis of complex languages for tasks like information retrieval and summarising papers. LLMs can potentially read and analyse complex software descriptions written by humans to generate useful inputs for automated testing. 

This project aims to understand the shortcomings of state-of-the-art LLMs in analysing complex descriptions of software. This is done by creating benchmarks for evaluating ChatGPT 4o’s AI model on the RFC9000 document. By automating this process, any user, no matter the expertise can use AI to guide them through complex documents. Upon completion, this project can expand to other complex documents following successful evaluations.

Project by:

• Wei Long Wan 

Rheumatoid arthritis is a global health complication, causing inflammation and damage to the joints and bones. The significance of early detection and diagnosis is crucial for improving patient outcomes in which Magnetic resonance imaging (MRI) is utilised to great effect by medical professionals. However manually interpreting hundreds of images is time consuming. Enter Artificial Intelligence (AI), which has demonstrated an incredible ability to detect critical features in MRI scans with up to 96% accuracy, potentially surpassing human capabilities. 

My project’s focus is to enhance the traditional hip arthritis diagnosis via deep learning algorithms, segmenting critical regions and identify disease progression based on known disease traits. My model aims to streamline and optimise rheumatoid arthritis detection in clinical settings, by employing a semi-supervised learning approach. This innovative approach ensures accurate training and generalisation, overcoming the challenge of limited medical datasets.

Project by:

• Eugene Xue 

In collaborative settings, people generally perform better alongside other people. However, hiring multiple humans for one single task can be expensive. To address this issue, there are behaviour-cloning AIs that mimic human actions, enabling them to collaborate effectively with humans. However, training these AIs typically demands a substantial amount of human data, which can be costly to gather. This project aims to overcome this obstacle by recombining pieces of existing human data to generate additional training data for the AIs. By stitching together these data segments, we can exponentially enhance the volume of usable human data, thereby lowering the costs associated with behaviour cloning. This method makes the training process more efficient and economical with minimal performance drawbacks.

Find out more - project video

Project by:

• Ka Hei Samuel Kwok

Fatigue analysis of all aircraft components as well as crack propagation and monitoring is critical to ensure the ongoing safety and reliability of in-service aircraft. As a non-destructive examination technique, Thermal Stress Analysis (TSA) provides engineers with accurate full field stress data across components. Due to its non-destructive nature, improvements in TSA procedures have become a large field of research with the potential to greatly expedite aircraft structural testing protocols. Whilst research is currently being undertaken to develop and analyse TSA data to obtain full-field 3-dimensional stress maps, the application of this technology to non-planar crack tracking remains largely unexplored. This project aims to implement 3-dimensional TSA methodologies for the tracking of out of plane cracks. By combining cutting-edge imaging with computational models, we aim to optimise current crack tracking capabilities for improved fatigue analysis of aircraft structures.

Project by:

• Niamh Michael-Roubos

The goal of our project is to make future wireless networks (6G) faster and more stable. Imagine being able to download a movie in seconds, or talk to friends and family in 3D anytime, anywhere. What's even cooler is that driverless cars can also be realized through this network, avoiding traffic accidents and making travel more convenient and safe. Terahertz communication is such a very fast wireless data transmission method, but because of technical limitations, it is still often wrong. We are using a method called "pilot" and "LDPC" to reduce these errors, make the data transmission more accurate and stable, and improve the terahertz technology. With the help of this technology, the future of the network will be more powerful, making life more convenient and fun.

Project by:

• Yuexi Ma 
• Marlon Kha 

Ever heard loud banging noises when you turn the tap off too quickly? That's a water hammer (otherwise known as a transient), and it can be quite damaging to any hydraulic system. Now imagine this event occurring on a larger scale - a long pipeline within an industrial water distribution system buried deep underground. The hydraulic transient analysis of water pipe with consideration for pipe-soil interaction is complex; there are limited existing numerical methods that can provide an accurate analysis. This project intends to contribute by investigating and attempting to create a simulation that can successfully predict the behaviour of pressure within a pipe with consideration of external pressure from the surrounding soil during a transient event. With this understanding, protection from hazards due to transients and vibration in the hydraulic system can be achieved by timely analysis.

Project by:

• Qayyimah Zamri 

An antenna system is not much more than a fancy stick or dish hooked up to an electric system that sends waves of different sizes that are invisible to the eye. But they can be seen by a device that is set up to only see them. Normally these sticks are made of metals like copper, silver and or gold. These materials used traditionally in sending out these invisible waves. In the past few years it has been found by some clever thinkers, that these waves could be received by liquid water, introducing a new way of sending radio-waves without using metals.

Find out more - project video

Project by:

• Ryan Dan 
• Minh Anh Nguyen 

Large Language Model(LLMs) like ChatGPT have garnered significant attention due to their remarkable ability to comprehend human language and generate text outputs accordingly. Based on their large-scale parameter sizes and a broad range of training datasets, LLMs have been actively applied to code generation tasks in recent academic and industry studies. Using NL2Code (natural-language-to-code) configuration, the code generation model takes natural language instruction as input and outputs code snippets. However, due to the size of the model and the restricted context window that limits the number of tokens LLM can proceed, instructing the model to perform accurately is considered a challenge.

One can always fine-tune a model using some datasets, but the computational overhead serves as a roadblock especially when dealing with LLMs.

Prompt engineering, which focuses on improving the quality of prompt structure, template, and composition, has emerged as a promising solution. The fundamental premise of prompt engineering is to retrieve the learned knowledge embedding within the language model through the optimization of prompt structure, template, and token selection.

In this setting, the language model is kept 'frozen' and is often referred to as a pre-trained language model(PTM) to avoid the huge computational overhead caused by fine-tuning the modern LLMs. 

However, prompt engineering is still a time-consuming process. Significant manual labor efforts are needed to explore the optimal structure of the prompts, and manually searching for the most downstream task-related keywords that provide conditioning to the model is a highly tedious process.

In this work, we propose as a novel solution to address the prompt optimization problems. Our solution is based on Diffusion, inspired by its ability to reshape noise into the desired high-quality sample, instead of training and storing task-specific prompts separately, we are aiming to train diffusion as a prompt projector that project Gaussian noise sample into the optimal prompt embedding for the code generation task. 

Find out more - project video

Project by:

• Jinyang Li

Microservice architectures have been widely adopted to improve various aspects of software systems. In the microservice design, complex applications are broken down into a set of small services which run independently but they can communicate and work together. Each microservice is allocated an amount of resource to run, however, when the system experiences an unexpected high workload, the allocated resource might not be sufficient to maintain the system’s desired performance. Smart HPA is an autoscaler which resolves this challenge by transferring surplus resources from idle services to more busy services. However, the original implementation of Smart HPA assumes an error-free environment. Our project focuses on analysing and implement fault-tolerant techniques for various failure scenarios during Smart HPA operations. This ensures microservices can be scaled by Smart HPA regardless of partial failures. Therefore, the system can efficiently handle fluctuating workloads, resulting in higher system’s performance, availability, resource utilization and user satisfaction.

Project by:

• Nguyen Thanh Toan Le