Transforming Technologies

Have you ever noticed that when you ask an AI a question, the answer can change depending on how you ask it? That is the challenge my project looks at. Right now people often guess how to phrase their instructions, which are called prompts, and sometimes the AI gives a great answer but other times it misses the point.

My project tries to remove the guesswork from this process. I built a system that creates many different prompts, makes small changes such as swapping words, changing the order of sentences or mixing different styles and then tests which ones give the best results. It works like practice, where the AI keeps trying until it gets better.

The outcome is a smarter way to talk to AI. It helps the AI write summaries of articles, conversations or news stories that are clearer, more accurate and easier to understand.

Project by: Aditya Karthully 

Quantum technologies rely on materials that can carry electricity with no resistance, called superconductors. One promising application is in superconducting quantum interference devices (SQUIDs), which can measure extremely small magnetic fields with high precision. However, producing superconducting materials with the right properties remains a technical challenge.

This project focuses on growing very thin layers of the metal niobium on silicon substrates using a method called electron beam evaporation. By carefully adjusting growth conditions—such as temperature, pressure, and beam strength—different film samples will be produced. These samples will then undergo electrical and structural testing at room temperature to identify which growth conditions yield the most promising properties. If successful, further testing will be conducted at low temperatures to confirm superconductivity.

The outcome of this project will be a clearer understanding of how growth conditions influence superconducting films, providing valuable insights for advancing SQUID technology and future quantum devices.

Project by: 

• Ishaan Oak 
• Sebastian Alfred

Within environments of excess danger or difficulty for human exploration such as subterranean caves and disaster relief, robots and drones play the critical role to find objects of interest. But how do we enable these robots to navigate and explore without the Global Positioning System? This project tackles the implementation of a distributed simultaneous localisation and mapping algorithm on physical robots, exploring and documenting the process in building robot teams that will be able to perform tasks too risky for humans. With feedback from real world experiments, a robust solution is developed that can also be adapted easily to various robot platforms. This system demonstrates the enhanced capabilities of a distributed robot team, forming the foundations necessary for further research in the field.

Project by: Tom Xiaoguang Zhu

In the electronics industry, antennas are used to send and receive signals. An example is the TV antenna you may have on your roof that allows you to watch the news. These TV antennas are typically a few rulers in length and have various metal poles sticking out that allow them to receive signals. A special type of antenna called “superconducting quantum antennas” or SQAs receive signals just like TV antennas do but are only half the length of your finger! The problem is SQAs are so good at receiving signals that they also pick up unwanted noise making the original signal hard to understand. It’s exactly like when you can’t hear what the person you’re speaking to is saying because the background noise in the room is loud. Thus, this project focuses on measuring how much unwanted noise SQAs pick up to help make the signals they receive understandable.

Project by: Joshua Graham

The hardest bug in software teams is poor communication, and we test whether AVW-Space can fix it. Communication skills are crucial for effective software development teams, but those skills are difficult to teach. The goal of this study is to evaluate the effectiveness of teaching face-to-face soft skills (presentation and empathy) using AVW-Space, a platform for video-based learning that provides personalized nudges to support students’ engagement during video watching. Within the space created by their teacher, students can watch the videos, write comments, and rate their peers’ comments. We will recruit software engineering students, administer brief surveys before they begin AVW-Space and after they complete the activities, and compare the results. This pre/post design allows us to see clearly whether AVW-Space helps software engineering students build the soft skills that effective development teams rely on.

Project by: Ali Zeeshan

Imagine if you had a very smart robot friend who could write computer code just by listening to instructions. This robot is powered by something called “large language models” (LLMs), which are tools trained to understand and create human like text. My project looks at how good these robot friends really are at writing code.

The aim is to find out which robot does the best job at solving problems, following instructions, and writing code that actually works in real life. To do this, I give them the same set of challenges and then test their answers using simple checks, like whether the code runs correctly or not.

The outcome of this project shows which robot is the most reliable, accurate, and useful for programmers. This can help businesses, schools, and everyday people choose the best tool to make coding faster, easier, and more creative.

Project by: Durga Sai Eswar Chodavarapu

Large language models (LLMs), like ChatGPT, have become a part of daily life for many, helping with writing, coding, brainstorming and more. However, generating text with such models can be extremely slow, as for each token generated, it requires loading the model's parameters, which may be hundreds of billions, onto memory. To combat this, speculative decoding has emerged, where a small draft model predicts the original model’s output, then the original model verifies the drafter’s predictions, allowing for multiple tokens to be generated for each loading of the large model, improving efficiency. EAGLE-3 is one such speculative decoding method, and like the others, it can be sped up. Our project, OptimEAGLE, aims to explore the capabilities of, and optimise, EAGLE-3, by experimenting with context compression, LLM-instructed optimisation of inference parameters and the technique's KV cache usage. Faster LLMs makes their impressive capabilities accessible to more people.

Project by: Ebubekir Pulat

The upcoming Adelaide University merger will cause a significant increase in students. Currently, the university has a few flow visualisation rigs that are used for practical sessions. However, they have several limitations such as footprint and ease of use. The lack of any documentation is also a significant issue preventing the creation of further rigs needed for the increase in students. Therefore, this project aims to design, build and test a flow visualisation apparatus to be used for teaching purposes in the new Adelaide University. The new design is an optimised upgrade which also makes use of 3D printing technology to deliver a custom solution that is easily re-creatable. Accompanying detailed documentation also ensures that the design can be utilised well into the university's future. Overall, our work will allow student engagement through practical, intuitive learning, and ease of interpretation of results.

Project by:

• Ahmad El-Kaissi
• Jonathan Rentsch

My project is about helping teachers make better maths lessons using smart computer helpers called AI. In schools, teachers often spend a lot of time creating questions and answers for their students. This can be tiring and take them away from teaching. I wanted to find a way to make this easier, safer, and more accurate, while still matching Australia’s school standards. Several small tests with fast AI have been done to see which one could best create maths materials. These models learned from examples of school work and information about the students, like their year level and skill level. In the end, this model can quickly make maths questions and answers, for example, addition, subtraction, multiplication, or division, that match what teachers need for different classes. This means teachers can spend less time on preparation and more time helping their students learn.

Project by:

• Franky Wong 

Micro-structured optical fibres are thin threads of glass with complicated internal geometries, critical to applications such as the Internet, medical imaging, and industrial laser systems. Fabricating optical fibres involves a drawing process in which a preform is heated and stretched to obtain a final fibre with desired diameter and geometry. This process is influenced by factors such as draw and feed speeds, surface tension, applied pressure, heating, and cooling. Sometimes, an instability called 'draw resonance' occurs during this process, compromising fibre uniformity and thus the optical properties of the final fibre. This project investigates the occurrence of this instability and its effect on fibre geometry. Using mathematical modelling, we show that surface tension and pressurisation are destabilising while cooling is stabilising, and that the combined effect of these three factors leads to more complicated behaviour. This improved understanding of the stability will help to minimise the production of defective fibres.

Project by: Caitriona Lightbody 

Can showing an AI one solved example help it solve the next? Large Language Models (LLMs) like ChatGPT are powerful at answering questions, but they often fail at tasks requiring multi-step reasoning, such as math word problems. Our project investigates in-context learning a technique where models are given solved examples as prompts to improve their problem-solving accuracy.

We build baselines by testing models on hundreds of competition-style math questions without any context, then compare these results to when semantically similar solved examples are provided. we explore whether AI systems can “reason” more effectively without retraining or fine-tuning.

Our findings will help researchers design smarter, more reliable AI with applications ranging from education and tutoring to safer automated decision-making systems.

Project by: Jay Chavan 

Imagine a limitless supply of power that can be plucked straight from the air. Our project explores wireless power transfer using highly directive arrays to deliver energy over long distances. The aim is to make mobile devices such as drones less dependent on batteries, reducing size, increasing performance and increasing run times. To achieve this, we simulated and tested multiple different transmitting arrays ranging from patch antennas to Fabry Perot structures to determine the best performance for our system. Further, we designed and tested the receiving antenna and rectifier circuit to convert the transmitted RF energy into the desired DC power for the remote devices. Through our simulations, we discovered the shortcomings of the different antennas and the impact on the system. The limiting criteria were size, directivity or design complexity which limits scalability. The project highlighted the challenges and potential for a high-efficiency wireless power transfer system.  

Project by:

• Benjamin Butler 
• Caleb Tonkin 
• Benjamin Whiting 

High-energy laser systems for fusion and high-energy-density physics require precise pulse compression to reach peak powers in the picosecond or femtosecond regime. To improve the reproducibility of research results and measurements, controllable motorised stages need to be adapted for laser researchers. Manual alignment of the multiple optical components involved is slow, error-prone, and hard to reproduce. This project aims to develop a cost-effective multiplexing system that allows a single motor controller to operate multiple stages, reducing hardware complexity. Hardware prototyping and custom software enables automated alignment in the pulse-compression stage of a laser system, optimising pulse compression and improving experimental reproducibility.

Project by: Genis Mulaj 

Music concerts and stage productions are renowned for their dynamic lighting effects, which amplify the atmosphere and immerse the audience. Scaling these effects to large outdoor events is challenging due to the distance between the stage and the crowd’s fringes.

Wearable lighting devices have been used in large productions, most famously by Coldplay and Taylor Swift, but current systems have significant limitations. They either rely on infrared signals to “paint” patterns across the audience, which requires constant line-of-sight, or use Bluetooth connections through each user’s phone, requiring user intervention.

To address these challenges, this project develops a system of bespoke wearable devices that utilise ad-hoc flood networking. This approach removes the need for external infrastructure, improves resilience to physical obstructions, allows the network to scale dynamically with crowd size, and reduces user setup requirements. The result is a more robust and flexible platform for audience lighting effects at any scale.

Project by:

• Bradley Langfield 
• Timothy Szabo 
• Luke Waldeck 

Porous liquids (PLs) combine permanent porosity with liquid-like mobility, enhancing gas solubility and mass transfer compared to conventional solvents. Type III PLs, created by dispersing metal-organic frameworks (MOFs) in sterically hindered solvents, are especially promising for catalysis and selective gas or liquid separations but are limited by particle aggregation and reduced pore accessibility. This project investigates the use of amphiphilic polymers to stabilise MOF colloids by forming surface coatings that prevent aggregation while maintaining access to internal pores. By systematically varying the ratios of hydrophobic, hydrophilic, and charged monomers, polymer architectures that balance colloidal stability with porosity retention can be identified. Dispersion behaviour and catalytic performance will be evaluated to establish structure-function relationships, providing design principles for stable, high-performance PL systems with applications in catalysis, gas capture, and chemical processing.

Project by: Daniel Piacquadio

To build an audrino system of Low-cost sensing system for condition moinitoring of structures, as in the strain time or displacement time curves of structures.

Project by:

• Mrinaal Nahata 
• Dyaan Abrar Biswas 
• Callis Fernandes 

Logging in geology means recording what is found in the ground after drilling or digging. We are building a digital system that makes this logging process faster and easier. Instead of the old system consisting of paper notes or manually updated spreadsheets, we propose a better solution. Teams can enter information directly into online sheets, and the system will automatically keep those sheets synchronised. This means that when one person makes a change, everyone else can see it instantly; reducing mistakes, saving valuable time, and ensuring that the whole team is always working with the most recent information.

Project by:

• Tristan Hyams 
• Anthony Tran 
• Wisam Aburas 
• Blake Jones 
• Sam White 

Imagine needing to read thousands of news reports every single day just to find out which countries are being discussed. This is a real challenge for risk intelligence companies that work to keep businesses safe from global threats. Using the most powerful AI to do this is often too slow and expensive, which means important information can be missed.

To solve this, my project uses a method like a "teacher" and a "student". A very large, smart AI acts as the teacher, and we use it to train a smaller, faster AI student that learns to do the same job at a fraction of the cost.

The result is a new, efficient system that can tag huge amounts of information, closing a critical "intelligence gap". This allows human experts to quickly find the information they need, helping them better protect organizations from real-world risks.

Project by: Rajan Paneru

Spin crossover (SCO) complexes act as molecular switches that reversibly alter their magnetic, optical, electrical, and structural properties in response to external stimuli. The physical characteristics of SCO complexes can be further engineered by incorporating a shape-shifting molecule, bullvalene, into the design of the complexes, offering dynamic tunability to an otherwise binary system (high or low spin). The aim of this project is to synthesise one such shape-shifting SCO complex and to investigate its potentials as a molecular switch. An existing model of SCO complexes was modified and prepared via alternative synthetic routes. Spectroscopic studies have been performed to confirm the structure of the complex and determine the external stimuli required to trigger the switch. We herein report a novel shape-shifting SCO complex that operates as a tuneable multistate switch for potential applications such as data storage, photonic devices and electronic circuits.

Project by: Jiapei Liu 

Ever clicked an email that looked legit but wasn't? Phishing remains the web’s oldest magic trick, fooling people and security tools alike. Our project asks a simple question: which modern AI “co-pilot” — ChatGPT, Gemini, LLaMA or Grok — is the most reliable spam-spotter for real-world inboxes? We built a no-cost, reproducible testbench using public email datasets and a transparent prompt-and-evaluate pipeline. In plain terms: we gave each model the same emails, asked “is this phishing and why?”, and measured three things people care about: accuracy (gets it right), specificity (avoids false alarms), and efficiency (speed/cost). So far, results show clear trade-offs: some models are keen detectives but over-zealous, while others are calmer and miss fewer legitimate messages. Our open benchmarks and prompts help teams choose the best fit for their risk tolerance and budget, and point to practical ways to combine AI with existing filters to keep inboxes safer.

Project by: Mohammad Samman Hossain 

Quantum field theory (QFT) is currently the best model of the fundamental constituents of reality. Classical theories describe forces on particles due to fields, but QFT implies that particles themselves are waves propagating through quantum fields. What we generally recognise as forces are in fact interactions between fields.

When multiple quantum systems interact, they can become entangled, meaning the states of the systems are coupled to each other. Quantifying these relationships naturally leads to a description of the information encoded in the composite system. Hence, quantum information theory (QIT) is an area of interest in particle physics.

In this project we learn the basic principles of QFT; particularly, Feynman's rules for computing scattering amplitudes of strongly interacting particles such as quarks. We then investigate whether the same physical outcomes can emerge with QIT as a foundation, potentially offering a fundamentally different way of explaining particle interactions.

Project by: Luca Bastians

For communities near airports, aircraft noise affects both residents’ quality of life and health. This has led many airports to implement curfews at night. This project aims to support Blue Spirit Aero’s development of its hydrogen-electric aircraft, Dragonfly, by investigating ways to reduce its noise emissions and allow it to operate in curfew hours. This therefore aims to minimise its impact on people living near airports.

To do this, we use computer simulations to model the noise from multiple propellers. In conjunction, we performed experiments using 1:3 scale propellers to validate the numerical results. We then use these results to propose ways that the propeller acoustics can be reduced. The outcomes of this project are to provide data for Blue Spirit Aero on the optimal configuration of propellers for reducing noise.

Project by:

• Angus Braendler 
• Corey Wright 
• William Bai 

Terahertz radiation is an emerging area of interest, and viable components are necessary to manipulate it. Current methods are costly, difficult to implement and design, and ineffective, thus a better solution is required for terahertz radiation to be used commercially. Our project attempted to solve this by focusing on contactless dispersion tuning. By building a resonator featuring electronically controllable moving metal plates, effective tuning of terahertz radiation in silicon waveguides was achieved, demonstrating the viability of adoption of this method due to its modularity and low cost.

This is a critical step towards the widespread adoption of technologies using terahertz radiations. Potential applications of this project exists in countless different fields, ranging from safer medical imaging of tissue, improved and non-destructive quality control of products, to better security through non-invasion scanning.

Project by: Jordan Nedosyko 

Hospitals can get very crowded, making it hard for patients to move through quickly and safely.

Our project aims to make patient visits smoother by finding ways to reduce waiting times. We studied how patients move between hospital wards using a computer program that looks at real hospital data, like a map showing where people get stuck.

We also created a pretend hospital on the computer to test ideas, like sharing doctors between wards or combining waiting lines. Our tests showed that sharing doctors and combining lines helped patients get through the hospital faster, especially when it’s not too busy.

This means hospitals could use these ideas to help patients wait less and get better care. 

Project by: Chia-Yuan Kuo 

The project will explore the possibility of optimising the mechanical performance of multilayer ceramic coatings within a FEA (Finite Element Analysis) informed design framework with the aim of mimicking a shell-like animal to improve hardness and decrease rate of fracture of multilayer ceramic coating. The short-term goals are to reproduce a benchmarked multilayer model in the literature, calibrate the behaviour of the model against published experimental data, and then iteratively optimise the structure by varying layer thicknesses, material combinations and gradient parameters. The ultimate outcome is to suggest a coating structure that will optimise the force of impact that a coating can handle without compromising the high hardness levels, in effect breaking the traditional performance trade-off.

Project by:

Moath Al-Rumhi 
Weiven Tan 
Scott Tiller 

The rapid transition to electric transportation has created strong demand for electric motors with higher torque density, efficiency, and lower cost. However, most high-performance designs rely on rare-earth permanent magnets, which are expensive and environmentally challenging. This project aims to design and analyze an axial-flux electric machine that reduces rare-earth dependency and compare their cost and performance with conventional PM motors.

Design and analyze an axial flux motor using analytical models, equivalent circuit methods, and finite element simulation (ANSYS Maxwell). Key performance parameters are evaluated using a combination of computational and experimental methods, while also learning the difference between 2D and 3D studies.

The results highlight the advantages of axial-flux topologies in torque density and compactness, while also comparing PM and non-PM options in terms of efficiency, cost, and suitability for electric transportation.   

Project by:

• Jiazhen Wang 
• Liquanze Zhao 

Australia is surrounded by oceans, home to vibrant coral reefs and thousands of kilometres of underwater pipes and cables carrying energy, internet, and other vital resources. Many reefs are stressed by climate change, while ageing subsea infrastructure is often difficult or too dangerous for humans to inspect. The Cuttlefish Autonomous Underwater Vehicle (CAUV) project aims to design a small, agile submarine to monitor both marine life and underwater infrastructure safely and efficiently. Using undulating fins, the CAUV moves quietly through the water, avoiding disturbance to wildlife. Its compact, highly manoeuvrable design allows it to explore confined or complex spaces that traditional submarines cannot. Upgrades this year include a variable ballast system for precise vertical control, letting the CAUV dive deeper than ever before. Additionally, complex systems engineering and rigorous testing ensure it is more reliable, easier to maintain, and more effective at accomplishing its mission than ever.

Project by:

• Adam Longmire 
• Ethan Roberts 
• Jin Hong Lim 
• Lucas Bartle-Browne 

A current industrial challenge in the brewing industry is the optimization of the brewing process to achieve the desired quality. This is particularly evident when developing new recipes and their upscaling for large scale production. Therefore, the creation of digital twins to predict the fermentation process is useful. A digital twin is an advanced modelling tool that can integrate real time data into the model to provide process predictions and improve efficiency.

Our project focuses on The University of Adelaide Waite Microbrewery and the optimisation of the fermentation portion of the brewing process, specifically for the ale style of beer. In this project, data was collected and a model was created by conducting small-scale experiments, where small batches (20L), were brewed while varying the mash temperature and yeast used. The model was then validated with a large-scale batch (200L).

Project by:

• Nicole Richards 
• Olivia Pugsley 

Imagine if you had a robot friend who could read the news for you and tell the story in just a few simple sentences. That’s what my project is about! I use AI to turn long, complicated articles into short and clear summaries. But here’s the trick: the way we ask the AI makes all the difference.

I test five different styles of instructions, from giving examples, to asking for “kid-friendly” explanations, to guiding the AI step by step. I also play with mutations—tiny changes in words or style—to see how they change the results.

By checking readability and accuracy, my project shows which prompts help AI give the clearest, most useful summaries. The outcome is a step closer to making information easier and faster for everyone to understand—whether you’re a student, a busy worker, or even an alien visitor!

Project by: Jiyin Shao 

Cryptocurrency is typically thought of as a safe haven during time of fear and low trust in financial markets. It has become apparent that this is not always the case. Blockchain assets have been subject to high volatility during bear markets in the crypto industry, with the potential for some to evaporate entirely. It is important then to ask how one can properly model the various blockchain ecosystems in a way that allows for quick and accurate prediction of price changes in the near future. This paper proposes that network statistics such as transaction fees and validation rewards could be used as financial signals to indicate the overall health of the network and hence predict changes in price for blockchain asset.

Project by: Matthew Theiley 

Imagine you want to teach a robot how to write a short summary of a long story. The tricky part is figuring out the perfect instructions to give the robot so it does a good job.

Our research explores a way to make the robot figure out the best instructions on its own.

We start by giving the robot a bunch of different instructions. We then test each one and give it a score based on how good its summary is. The instructions that get high scores are "winners." We then take these winning instructions, mix them together, and make small changes to create a new set of instructions.

By repeating this process, the instructions "evolve" and get better over time. The final result is a set of highly effective instructions that the robot discovered itself, helping it become an excellent story summarizer.

Project by: Gaurav Singh 

My project focuses on making computer systems learn faster and with less effort. Training deep learning systems usually takes huge amounts of time, energy, and computer power since every model starts from scratch. I am solving this problem by letting them “share what they know,” so knowledge from one model can be reused by others.

To do this, I used CoDeepNEAT, an evolutionary method that designs and improves deep neural networks over many generations. I integrated weight sharing into CoDeepNEAT, so that when modules (small building blocks of the networks) are reused across different models, they can also reuse the same learned weights instead of starting fresh. This is like classmates not only sharing notes but also sharing the actual answers they worked out.

The result is a system that builds powerful models faster, using less computing power, making advanced technology more sustainable and efficient.

Project by: Sachin Parimanumkuzhi 

When a group of people decides to build something together (a bridge, pool, or playground), everyone wants different things and has different financial capabilities. This gives us the problem: how do you split the benefits fairly so that nobody lies about what they can afford? We created super-fair judges (represented by tiny computer programs) to help situations like this. These computer referees can figure out how much reward to distribute among group members based on their contribution, in order to make sure that everyone benefits fairly from the project, and therefore have good reasons to tell the truth at the beginning. We used a special training method where a big "teacher" computer keeps making better and better small computer judges. Our method successfully improved the worst-case scenario for groups of up to 10 people over the old approaches. This helps communities make better decisions together.

Project by: Huy Trung Truong 

Crypto prices jump and fall with online chatter. Our project asks a simple question: can reading the crowd’s mood make an automated trading bot smarter and steadier? We collect posts from Twitter and Reddit, score how positive or negative they feel using a language model, and blend that mood with basic chart signals. We then “paper trade” coins like Bitcoin, Ethereum, and Dogecoin to see if sentiment helps the bot avoid bad trades and ride better ones. Early testing suggests that filtering by mood can reduce false alarms in choppy markets and smooth results. At Ingenuity, we’ll show a live demo and walk through how a tweet becomes a trading decision—step by step, in plain English.

Project by: Shashank Gupta 

Generating high resolution 3D images of objects has a variety of applications, including for security screening, non-destructive evaluation of materials, and medical imaging. With terahertz Synthetic Aperture Radar (SAR) imaging, instead of taking images with visible light like a camera, terahertz waves are used. Depending on the geometry and materials, the wave reflections will differ, allowing for a detailed 3D image to be reconstructed. This image not only shows the surface and shape of the object, but also internal details due to terahertz wave’s ability to penetrate into many materials. This project aims to demonstrate that two existing techniques: SAR and Optical Coherence Tomography (OCT), can be combined to produce high resolution 3D images at terahertz frequencies. This would allow these technologies to be applied to imaging applications, potentially increasing resolution of images obtained, or reducing the cost or time taken to complete a scan.

Project by: 

• Ryan Rutherford 
• Sam Heathershaw 
• Sean Priestley 

Trading in cryptocurrency markets, like Binance, is fast and unpredictable. Many people try to design algorithms, called trading strategies, that can buy and sell at the right time. The challenge is that these strategies are usually trained on past data, and if the market changes, they often fail. My project aims to make trading strategies more reliable by giving them better “practice” data.

I use a type of artificial intelligence model called a Generative Adversarial Network (GAN). This model can generate new, realistic versions of past market behaviour, extending the amount and variety of data available for testing. By combining synthetic price movements with real historical data, I aim to expose trading strategies to a wider set of scenarios. This is similar to stress-testing, where I'm pushing the system under many different conditions to see how well it adapts when the market shifts unexpectedly.

The outcome aims to show that strategies trained with this mix can adapt better, helping traders reduce losses and make smarter decisions in unpredictable markets.

Project by: Karen Veigas 

Setting Up the Frameworks to Upcycle Plastics

2,000 garbage trucks of plastic are dumped into oceans, rivers, and lakes daily. Plastic pollution has shaped up to be one of the biggest problems humans have faced, with many methods investigated for plastic recycling and degradation.

This project aims to use a Zr-based Metal-organic Framework (MOF) as a platform for Ruthenium nanoparticles for the hydrogenolysis of polyolefins, a type of plastic found in pipes, plastic bottles, and bottle caps.

We synthesised three variants of the MOF MIP-206 to investigate the effect of a changing pore environment on catalyst efficiency. This research could help in the global effort to reduce plastic waste, and further understanding of MOFs as platforms for catalysis.

Project by: Isaiah Korcz 

Electric motors come in all shapes and sizes. They are what drive electric vehicles, and they are the fans that cool our computers. With the increasing usage of electric motors in our lives, they make their presence known through the sounds that they make well before they are ever seen. The basis of this project is to document the process of predicting the acoustic noise generated by an electric motor before they even leave the design floor. This project was done through recording the noise and vibration of a physical motor, and simulating said motor with software. This allows for comparisons to be made between the physical model and simulated model, and to further refine our models.

Project by:

• Aidan Sern Arn Ong 
• Hao Hung Nguyen 
• Blayne Fishlock 

Designing Siamese Neural Networks in image retrieval tasks are really time consuming and critical if it is done manually and there is the question of having mediocre performance through manual design. That is why Neural Architecture Search is gaining popularity in Siamese Network design where research has shown that Automated projection heads perform better than manually designed ones. This project proposes a guided evolutionary algorithm based Neural Architecture Search which will search for optimal multilayer perceptron models using zero-cost proxy estimation and genetic mutation of the model. Finally This project aims improve the performance of Siamese Network by automating the projection head through guided evolutionary algorithm from a vast search space.

Project by: Nusher Jamil Kazi

Emergency situations are unpredictable, complex, rapidly changing, and dangerous. Due to advances in technology, robotic platforms working in partnership with human responders could reduce risk to personnel, improve response times, and enhance mission effectiveness. This project aims to build a team of autonomous robots that can collaborate to map the unknown and hazardous environments of emergency situations, enabling the rescue of humans and animals. The robots use advanced exploration and detection algorithms to understand the environment and coordinate actions as they explore the area of interest. Specifically, they can detect humans, dogs, cats, and other objects of interest, while assessing environmental conditions and providing situational awareness. Presented is a multi-robot system, which is capable of simultaneously cooperative mapping, object recognition, and real-time decision-making.

Project by:

• Arnav Vaid 
• Izak Stentiford 
• Thuong Vo 

The top quark (heaviest and shortest-lived elementary particle) pairs up with its corresponding antimatter to produce toponium. This discovery is new for unstable quark states. The challenge becomes finding the underlying cause, with a possibility of a new particle decaying into toponium. The aim of this project is to examine samples of this pair production to produce independent results to validate the findings. Different software and clusters will be used to gather the samples and analysis techniques, similar to the discovery techniques, will be used to produce independent results. We will present a poster that displays the analysis techniques and findings of this project. The discovery of toponium and run 3 of the Large Hadron Collider (LHC) helps high energy physicists understand the interaction between quarks and gluons, along with any other new particle or adaptions to the Standard Model.

Project by: Philip Kovacs

Gravitational waves are extremely small distortions in space, far smaller than the width of a proton. In order to detect these waves, scientists use a Michelson interferometer. This tool works by splitting a laser beam into two paths, bouncing them off mirrors, and then bringing them back together. But here’s the problem, if the mirrors move even a little bit, the interferometer becomes unstable and is unable to take accurate measurements.

The aim of this project was to stabilise a benchtop Michelson interferometer using a feedback system called a PID (Proportional-Integral-Derivative) controller. This system continuously adjusted the position of one of the mirrors based on the light signal received. Initially, the controller parameters were poorly tuned, causing instability. Through systematic adjustment, values that produced a steady, reliable signals were identified.

Project by: Lucy Scholten

Birds and insects can fly with amazing grace, using their wings to glide, hover, and dart through the air. Our project takes inspiration from nature to explore how small flying robots could one day do the same. Unlike conventional drones, these robots use flapping wings to move through the air, much like a bird or a butterfly.

The goal of our project is to design a robust and reliable control system that enables the “robotic insects” to fly smoothly and maintain balance. The system contains several sensors and control algorithms that manage movement along six axes of motion. We tested the design both in simulation and on a physical prototype to demonstrate its stability and ability to maintain controlled flight.

Our project demonstrates the potential of flapping-wing robots for tasks such as examining narrow environments, surveillance, or wildlife studies, and can help us understand how to mimic natural flight. 

Project by:

• Sanjeev Garfield 
• Chi Hao Ng ​​​
• Eurie Jean Miro ​

This project explores adversarial attack methods against deep neural network models. In particular we consider algorithms that seek to challenge and better understand certified defenses developed for deep neutral networks. In this regard, we consider current attack vectors and potential future methods to better understand failure models of certification methods in practice.

Project by: Evam Kaushik

This project looks at how large language models (LLMs) can help detect signs of depression in text. I used several datasets, including Reddit, Twitter, and the Extended Distress Analysis Interview Corpus (E-DAIC). I tested different methods such as zero-shot prompting, one-shot prompting, and few-shot prompting, where the model is guided by examples. I also explored using prompts based on DSM-5 criteria and chain-of-thought reasoning to make the answers clearer. To measure results, I used standard metrics like accuracy, precision, recall, and F1 score. Visualisation tools such as confusion matrices, bar charts, and word clouds helped me understand where the models perform well and where they make mistakes. The main goal is to see how effective LLMs can be for mental health text classification and to find safe, practical ways to use them.

Project by: Yinzhi Tian

Data collected from and processed by humans has played a vital role in the development of AI tools in the field of social science. But how can we trust AI when the data itself reflects human bias and subjectivety? This project explores methods to learn not only from crowdsourced data, but also about the crowd behind them. Neural networks, a popular and commonly used type of AI models, have been considered a black-box algorithm that performs well but lacks interpretability. Our approach uses context-aware network structures to make the AI model more interpretable, uncovering morality and perspective hidden in texts, and measuring how annotators agree and disagree. By experimenting with different architectures of neural networks, we propose two frameworks that measure human bias and confusion in their annotations respectively, while maintaining high accuracy.

Project by: Yi Ren 

A digital twin of the fermentation process provides a predictive model of fermentation activity. This can be used to determine the outcome of recipe changes before the beer is brewed. By collecting data from a range of brews, a mathematical model can be adjusted to fit the yeast behaviour. This model will be validated with by comparing small and big scale brews. The outcome was a practical model that could be used to predict lager yeast behaviour for any Helles lager recipe.  By inputting the mash and fermentation temperature, a specific gravity prediction could be made.

Project by:

• Thomas Riches 
• Edward Ward 
• Hoang Phong Vo 

The knowledge of underlying topographic features such as in an open channel is often limited. Accurate knowledge of the bottom topography is desirable to safely design and construct maritime infrastructure, with potential applications including designing the hull of a ship to minimise surface waves. Our project focuses on identifying the bottom topography given an observed free-surface flow using the forced Korteweg–de Vries equation (fKdV). We consider the practical situation in which the free-surface observations are corrupted by noise. The findings highlight a systematic method which is capable of predicting the bottom topography to a required experimental accuracy. We note that this method can be applied when measurements of various types of free-surface flows are known.

Project by: Caitlin Anchor

The Standard Model of physics successfully explains most particle interactions, but it is known to be incomplete, leaving open questions about possible new physics. This project investigates whether new physics could be revealed through the scattering of up (u) and down (d) quarks. Using Mathematica, we will calculate scattering amplitudes for the process ud→ud under both the Standard Model and beyond-Standard Model scenarios. By studying how these different amplitudes affect the quantum information carried by the process, we explore whether quantum information techniques can provide new possibilities to physics beyond the Standard Model.

Project by: Tom Haley

Large language models (LLMs) are powerful AI tools that can read, write and respond in natural human language. They are already used in apps like ChatGPT, but how safe are they when applied to healthcare? This project compares five LLMs and tests how well they answer medical questions and summarise long doctor notes. Using sets of medical questions and doctor notes, each model is scored for accuracy and clarity. The results show which models are more reliable and where they might make mistakes. The aim of this project is to help doctors and patients in the future by making sure these AI tools can be trusted. This project shows the benefits and risks of using AI in medicine, so that it can be used safely and responsibly to support healthcare.

Project by: Subin Pulliyil Santhosh

While graphite is commonly associated with pencil lead, its cousin white graphite, or hexagonal boron nitride (hBN), possesses amazing nonlinear optical properties that can produce light with the ability to change the world!

In the quantum world, single-photon sources are essential, powering applications from quantum computing and secure communication to precision measurement.

By shining a high-powered laser onto an atomically thin layer of hBN, I looked for signs of second-harmonic generation, which causes incoming light to double its frequency. Detecting this would confirm that hBN has the right properties to produce single photons and be used as tool in many quantum devices.

While no clear signal was observed, the project highlighted the challenges associated with imaging atomically thin materials under immensely low levels of light. Quantum applications of nonlinear optics help pave the way toward building stable, scalable, and room-temperature quantum light sources, the building blocks of tomorrow’s quantum technologies.

Project by: Bailey Hoare

Tracking manoeuvring targets is necessary for safe traffic. Passive radar is a unique type of radar that utilizes existing signals in the environment, such as TV, radio and WiFi, instead of sending out its own signal. This make it cost-effective and harder to detect, but difficulty also increases when there are more moving targets.

Traditionally, multiple targets tracking is done using the Kalman filter, which predicts target movement based on mathematical models. While effective, this method often struggles in complex or unpredictable scenarios. This project aims to have a comparison between this approach with machine learning, a technique that allows computers to identify patterns by learning from data.

Accurate data from flights is used to train the selected machine learning model. After training, comparison results can be obtained using recorded radar data. The results will be used to further demonstrate the practicality of tracking in passive radar using machine learning.

Project by: Kar Fai Lock 

Imagine waking up to a world without music! When COVID-19 hit the world in 2020, online platforms like Spotify, Apple Music were booming with users, who resorted to digital music on the internet to alleviate the suffering of social isolation (Zhao et al. 2025, p. 354). Indeed, research has shown that music has several cognitive benefits and is a powerful tool of self-expression (Magadum et al. 2024, p. 77470). Music Recommendation systems leverage explicit user feedback in terms of ratings or implicit feedback regarding user actions to recommend similar music to the user (Aggarwal 2016; Kotu and Deshpande 2019; Rebala, Ravi and Churiwala 2019). To date, many music recommenders have been proposed. While some recommenders struggle with adding more diversity to the songs they recommend, like Spotify, others struggle with leveraging related song features for a user cohort (Vashistha et al. 2024; Wu & Sun 2024). In this project, a content based recommendation system is proposed, that is shown to be effective at recommending relevant songs to a sensitive and sparse user cohort.

Project by: Syeda Ramisa Fariha 

Birds soar through the sky with wings that can change shape, allowing them to glide, dive, and turn with ease. Airplanes, however, have fixed wings, which can make them less efficient in different flying conditions. This creates a challenge: how can we design aircraft that are more dynamic, efficient, and stable in the air?

Our project explores this question using a model of Boeing’s UCAV 1303 aircraft. The focus is on morphing wings: wings that can twist and adapt. We have taken a particular focus on supporting the future development of this project by addressing the buildup of previous issues. The aircraft model was redesigned with improved aerodynamics and detachable sections, making it easier to add data collection devices inside. A new test stand was also built so the model can be accurately tested in a wind tunnel to advance the understanding of morphing wing benefits for aircraft performance.

Project by:

• Marcus Burman 
• Harry Lilburne 
• Harrison Pearce 
• Isaac Brown 
• Karthik Robin 

In the last few years, schools have started to use artificial intelligence in the form of tools like ChatGPT, adaptive learning platforms, and automated grading systems. These new technologies can be great for helping students learn, but they also make us think about fairness, privacy, and the changing role of teachers. Many teachers are stuck in the middle: they want to know what AI can do, but they don’t know how to use it well or in a way that is right. Recent research shows that teachers need to be ready right away and points out a clear gap in professional development programs that are meant to help teachers learn about AI. Some groups, like UNESCO (2021) and ACARA in Australia, have started to implement AI-related curriculum ideas, but most of these efforts are focused on students, which doesn’t give teachers the help they need. Schools could unintentionally make existing inequalities worse or take away teachers’ power and freedom if they don’t get targeted professional development. This project aims to address this significant gap by creating professional development that helps teachers not only understand AI but also use it responsibly and confidently in their classrooms.

Project by: Anamika Cherukat 

What instruments bring a martial arts duel to life? Classic Chinese films use rich soundtracks blending traditional and Western instruments, but identifying them has long required hours of manual annotation.

Our project develops a machine learning system that automatically recognises which instruments are playing and when. Using digital signal processing to reveal spectrograms, which are visual sound “fingerprints” sound, a convolutional neural network is trained on thousands of unique sound samples to detect instruments such as the Guzheng, Suona and gong even in noisy, overlapping film audio.

We apply this system to case-study films, including Come Drink with Me and Dragon Inn, and a Touch of Zen, generating time-stamped maps of instrument use. The results not only support cultural preservation and film scholarship but also demonstrate that machine learning is effective in  analysing soundtracks.  

Project by: 

• Hugh Signoriello 
• Jianzu Lin 
• Junran Hu 
• Yusheng Zhang 

Chatting with an AI language model can feel like typing instructions to a smart robot. But sometimes, the robot doesn’t fully understand what we mean. The words we use in our instructions, called “prompts,” can make a big difference in whether the robot gives the answer we want. My project explores how to make these prompts better.

Instead of humans guessing randomly, I use an approach inspired by nature: evolution. Just like animals evolve to survive better in their environments, I let the computer “evolve” different versions of prompts. The computer tries out many prompts, keeps the ones that work well, and improves them step by step.

Project by: Zijie Luo

Years of intensive industrial farming and climate change have led to poor soil health that has impacted crop yields. Adding to this challenge, the world's population is projected to reach 9.8 billion by 2050, highlighting the demand for farmers to increase crop production to meet global food needs. Our project supports farmers by leveraging smart farming technologies, autonomy, and data fusion to gather real-time, in-situ soil health information quickly and accurately. Our system gathers data, including soil organic carbon and soil compaction, without the need to dig soil up, helping farmers make informed decisions to improve their crop yields, reduce costs, and protect the land for the future.

Project by:

• Hamish Allan 
• Nisandi Fernando 
• Stacey Heng 
• Matthew Pinyon 
• Kathy Tran 

Many industries such as aerospace and electronics need materials that are both strong and able to carry heat effectively. Copper is excellent at transferring heat, while steel is durable and tough. The problem is that no single material can achieve both qualities on its own.The aim of this project is to create copper–steel “bi-metal” parts using advanced 3D printing and to study how printing conditions affect their surface quality and strength. The method used was Laser Powder Bed Fusion, where a high-powered laser melts layers of fine metal powder to build solid parts. Samples of copper, steel, and copper–steel combinations were printed under different laser power settings. Their surfaces were then measured with microscopes to test smoothness and bonding quality. The outcome showed that changing the printing settings has a clear impact on how well copper and steel can be joined into reliable, high-performance parts.

Project by:

• Blake Harder 
• Ambrose Biggins Baker 

If you have ever played with LEGO, you will understand that small, simple building blocks can assemble to create beautiful and complicated structures. In mathematics, we are also interested in doing this. It is often helpful to break a larger space into its smaller constituent parts, which are easier to understand and analyse. Sometimes this can give us useful information about mathematical objects, like how spaces can be looped or twisted in weird ways, or if they have holes or disconnected pieces.

Relatively recently, examples from geometry and theoretical physics motivated this theory to be 'upgraded' to a 'higher' level of abstraction. For example, when mathematical objects have connections between them, it is often insightful to investigate connections between these connections. This project aims to focus on these 'higher' ideas, and use them to classify and describe certain types of spaces.

Project by: Alex Marciano

Within the Electric Vehicle class of Formula Student, the competitors are free to design their powertrain alongside the given ruleset by the event organisers. The Adelaide University Motorsport Team (AUMT) have been running a conventional single motor design since it switched to the EV class in 2019. The Team is looking to explore a change to running individual motors for each wheel to allow for better packaging and more system control. Due to the size restrictions of not only the motors themselves but also the car itself, the suitable motors require a gearbox to reduce the motor rpm in favour of torque. These gearboxes are not available off the shelf and need to be custom designed and manufactured to withstand the load and abuse, but also maintain a low weight to be effective in the motorsport application. 

Project by:

• Alexander Zbroja 
• Zane Walker 
• Linsen Du 

The nature of Dark Matter is one of the largest open questions in physics, and experimental searches for it are at the cutting edge. The DAMA/LIBRA experiment has operated for more than 20 years in Italy, and claims to have a strong detection of Dark Matter; however, no other experiment has been able to replicate its findings.

The SABRE Experiment is planned to begin operating within the next year at Stawell Underground Physics Laboratory with the aim of confirming or refuting DAMA's results. SABRE will require an incredibly low noise background, and as such efforts are underway to characterise all sources of noise. Photomultiplier Tubes (PMTs) are one such source that need to be understood. Through building a light-tight environment and collecting large amounts of experimental data this project outlined a method for characterising one of the models of second-hand PMTs that will be used by SABRE. 

Project by: Isaac Covington-Groth 

Imagine moving products or packages without wheels, chains, or belts. Our project achieves this through creating a Contactless Air Conveyor System that uses air instead of physical rollers to transport items smoothly. The idea comes from the problem that regular conveyor belts wear out quickly, need lots of maintenance, and can be damaged or contaminated through frictional contact.

We designed a flat platform with tiny holes that blow out air. This air lifts the object slightly above the surface, just like how a hovercraft floats. By angling the direction of the air, we can push and guide the objects wherever they need to go. Sensors and a small computer help track each object to make sure it moves safely and accurately.

In the end, our system offers a clean and low-maintenance way for factories to transport goods, making production low-cost and easier without the usual breakdowns.

Project by:

• Benjamin Brierley 
• Peiyan Sun 
• Kai Priest 

It is another challenge to create computer "brains," called neural networks, because we have millions different ways to build them. Some are good, while most are inefficient or confusing. Our primary question is this one how can we discover good designs without wasting time and energy?

Our project applies evolution that same driving force that gets living organisms to adapt to improve these networks automatically over generations. The best designs are preserved, and improved. Where most methods are focused on coming up with networks that are very good in terms of accuracy, ours are as good, but also interpretable. That is, people can very much understand why a network is coming to the decision that it is, rather than the decision that it is coming to.

The end product is a way to build artificial-intelligence systems that are both powerful and transparent. Taking us closer to AI that is trustworthy.

Project by: Aryan Arora 

Neural networks are computer models inspired by the human brain, used in fields like healthcare, robotics, and self-driving cars. Building an effective network is challenging because both its design (structure) and training method affect performance.

This project improves an existing framework called EvoCNN, which uses the idea of evolution—similar to survival of the fittest—to automatically design networks. EvoCNN currently evolves only the structure of networks. My work extends this by also evolving the training method, known as the optimizer. Since the same design can perform very differently depending on how it is trained, evolving both together leads to stronger results.

The outcome is a smarter and more efficient way to create neural networks with less manual trial-and-error.

Project by: Husain Gadiwala

Photonic Integrated Circuits (PIC) promise great potential in innovating laser science for quantum sensing and information applications. Using PICs, we aim to reduce the Size, Weight, and Power (SWaP) of complex optical components to the size of a penny, providing opportunities for optical atomic clocks made at the fraction of the cost of currently available systems.

At the Institute of Photonics and Advanced Sensing (IPAS), current research shows more work needs to be done to reduce noise of the on-chip laser to maintain stringent precision requirements. Noise characterisation techniques, including the self-heterodyne detection test and the Relative Intensity Noise measurement, are key to develop discernment into further improving laser performance.

Optimising the current source to the on-chip laser, significant reductions in the experiment's SWaP have been achieved, while maintaining acceptable precision. Key insights into noise source origins have also been identified, which are crucial for future optical atomic clock operations.

Project by: Paul Pounendis

The project is situated in the mainstream of current research in elliptic complex geometry.  Approximation and interpolation are key themes across all of mathematical analysis.  They have been studied extensively in complex analysis and specifically, in recent years, in geometric settings where homotopy-theoretic obstructions may arise. The main goal of the project is to give a complete detailed proof of a Runge approximation theorem with interpolation and properness for holomorphic immersions of open Riemann surfaces into the complex affine plane C^2.

Project by: Chun Hei Lee 

This project develops a universal and intelligent DC switch for future smart houses operating on DC nanogrids. With renewable energy and battery storage becoming central to modern homes, DC power enables appliances to connect directly without unnecessary AC/DC conversions, improving efficiency and reliability. The switch provides a safe way to link appliances, energy sources, or storage units within the nanogrid, supported by built-in protection and intelligent control. It also offers remote monitoring and scheduling, allowing households to better manage their energy use in off-grid conditions. By combining flexibility, safety, and smart control, the project contributes to the vision of fully DC-powered homes that make renewable energy more practical and accessible.

Project by: 

• Khanh Nguyen 
• Bashir Bilal 
• Yuchen Zhu 

My project focuses on improving how computers design models for medical image classification. Normally, creating these models takes a lot of time because each possible design must be trained before knowing if it works well. This is especially slow for 3D medical scans, which need huge amounts of memory and computation.

To address this, I combined two fast evaluation methods, called ZiCo and JacCov, into a “hybrid” scoring system. These methods don’t require full training — instead, they quickly estimate how promising a model design is by looking at its mathematical properties. The hybrid system balances stability (ZiCo) and expressiveness (JacCov), making the evaluation more reliable.

Using this hybrid method inside an evolutionary search algorithm, I was able to identify stronger architectures more efficiently. The outcome was improved ranking accuracy and faster search times, meaning high-quality medical AI models can be developed with fewer resources.

Project by: Vigneshwar Ramachandran Rajasekaran 

We all have experience with shapes and symmetries in our everyday life—from simple things like rotating wheels to much more intricate designs, like snowflakes or patterns in art and architecture. In mathematics, the field of study that seeks to learn more about the world through symmetries is called group theory. There are many phenomena in physics which can be modelled using this theory, and it can be of great use in combinatorial calculations.

Understanding an object’s symmetries often tells us in great detail how that object is intrinsically connected, shaped, or even constructed. Studying symmetries can help us understand complicated structures, make complex problems simpler, and gain insights into how patterns repeat or interact with one another. This project aims to give an introduction to this theory, and explain how we can construct and model spaces using it.

Project by: Samuel Groocock

Have you ever wondered why some people wake up feeling tired, even after a long sleep? One reason could be hidden patterns in the brain called cyclic alternating patterns (CAP). These are small bursts of brain activity that show how stable or restless our sleep really is. Doctors usually find these patterns by looking at brain wave recordings (EEG), but this process is slow and often misses details.

Our project tries to help computers spot these patterns more accurately. First, we build “fake” brainwave signals using a tool called a generative model—like teaching a computer to draw new but realistic brain squiggles. Then we train another smart program that learns to recognise these patterns from both real and fake data.

The result? A system that can understand sleep better than before. In the future, this may help doctors quickly detect sleep problems and improve treatments.

Project by: Fandi Sun

Imagine talking to a robot helper that can work with your computer files, but the robot often forgets what you just said, or repeats steps in a confusing way. My project looks at how to make this helper 'smarter' when it talks about GitHub, a popular place where people build and share software. I use a new rulebook called the Model Context Protocol (MCP). MCP acts like a universal translator, letting big computer brains (AI models) connect directly to tools. The aim of my project is to give these AIs a better 'memory' so they do not lose track of the conversation, and to make their answers more accurate. By testing and comparing different ways of managing context, I showed that an AI can use MCP not only to understand tasks better, but also to complete them more reliably.

Project by: Yao Fu 

Farmers face many challenges in managing large fields, such as keeping track of plant growth, spotting diseases early, and protecting crops from animals or pests.  My project uses artificial intelligence (AI) to give farmers “smart eyes” that can automatically recognize different objects in fields, such as plants, fruits, and even animals.

I trained a computer system using many images so that it can learn to tell these objects apart, much like how people learn by looking at examples. The system can then process new photos or videos from cameras or drones flying over the fields, quickly detecting what is present.

By doing this, farmers can save time, reduce manual labor, and make better decisions about crop care.  The ultimate aim is to support more efficient and sustainable farming, leading to healthier harvests and improved food production.

Project by: Pengtian Yang

Computers learn to recognize things like cats, faces, or even medical images by following patterns. But often, when the world changes a little—like a blurrier photo, different lighting, or even a tiny trick added by a hacker—the computer gets confused and makes mistakes. This is a big problem because we want computers to work safely in the real world, not just in perfect test conditions.

Our project aims to design smarter “blueprints” for these computer models so that they can stay steady and reliable even when the world is messy. Instead of just picking the best-performing blueprint on easy tests, we also check how “stable” it is—whether small changes make it wobble or not. We use a mix of comparison learning (teaching the computer to tell differences between things) and special math tools that measure stability. The outcome? More reliable computer systems that are harder to trick and better prepared for real-life challenges.

Project by: Jianchun Zhao

With the widespread adoption of Large Language Models in recent years, AI is increasingly woven into educational practice. Students now have many AI options for learning, while teachers’ needs and uses for AI vary greatly. This project aims to categorize current teacher needs for AI applications and explore the causal relationships between these needs. We collected public posts from teachers, generated tags using LLMs, and manually cross-verified them. We then explored causal relationships within structured data. The results can inform current AI application scenarios for teachers, helping to prioritize investment and training in educational AI. It can also indicate the specific application scenarios that should be prioritized for the industry.

Project by: Shuhao Liu

The transformer has become the dominant architecture in machine learning tasks. Tokenisation (how the input is split before its fed into the model) is one of the most critical steps in designing a transformer based system. Existing tokenisation methods are effective for language tasks, but fall short in representing other forms of structured data.

This research aims to advance machine 3D compositional reasoning by exploring tokenisation methods tailored for LEGO LDraw CAD files. LDraw is the open-source standard enabling LEGO enthusiasts to create and share virtual LEGO structures.

To achieve this a variety of LDraw preprocessing and tokenisation methods are systematically evaluated to assess their impact on downstream model performance. Models are assessed on how effectively they can reconstruct a partially given LEGO structure. Results thus far show custom LDraw tokenisers outperform text tokenisers, and preprocessing steps that simplify brick coordinates and orientations are critical.

Project by: William Saliba 

This project aims to make AI responses fair when two companies want to speak at once, while still being helpful to the user. We built a simple, three-part system: two brand-specific writers suggest wording; a scheduler shares the talking time so no one takes over, giving more turns to a brand that has waited longer and linking overall share to what it pays; and an independent reviewer checks each sentence for sense, safety, and usefulness. To keep messages current, a retrieval step slips an up-to-date slogan into the notes before the reply is written. In early trials, this design gave more balanced exposure than a basic “highest-payer wins” rule. However, mixing the two writers word-by-word sometimes sounded awkward, while the slogan approach could feel too promotional. Next, we will let the neutral writer do most of the speaking, give each brand only a small share, and work to reduce delay.

Project by: Leon Xie 

Video games often become boring if they are too easy, or frustrating if they are too hard. At the same time, most computer characters (called agents) aim to learn only one "best" strategy, which makes playing against them repetitive and predictable. This project aims to make games more engaging by creating agents that can both adapt their difficulty to match the player while also not sticking to one constant strategy. We use large language models in conjunction with reinforcement learning to turn semantics into reward functions that guide how agents learn, resulting in agents exhibiting diverse learned behaviour. This method has been implemented in playable strategy games and the outcome shows how this can be a flexible and automated system that creates self-balancing and adaptive AI for video games.

Project by: Xiaoyu Tan 

Our project is about teaching computers how to design their own “brains” for solving problems. Normally, people have to build these computer brains, called neural networks, by hand, which takes a lot of time and trial-and-error. Our aim was to make this process smarter and faster.

We used an approach inspired by nature, similar to how animals evolve. Instead of picking the “strongest” animal, our system tries out many different network designs, checks how well they work, and then keeps the ones that are both accurate (make the right predictions) and fast (take less time to run). This is like choosing the fastest cheetah and the most careful owl, not just one or the other.

The outcome of our project was a set of network designs that balance speed and accuracy. This means computers can learn quickly and still give reliable answers.

Project by: Ish Shailesh Shah 

Vehicles today are similar to moving computers as they are filled with tiny, intelligent, electronic units that communicate using a “messaging app” called the CAN bus to make the vehicles run smoothly. Unfortunately, since anyone can communicate through this “app”, an impersonator could easily trick the car by sending bad commands, which would be dangerous for the passengers. My project aims to make intelligent vehicles safer by building a smart intrusion detection system (IDS) that listens to these messages and identifies intrusions. Using advanced machine learning models, the IDS is trained to recognise patterns. The key challenge I have addressed is how to make the IDS compact enough to fit in automotive environments without slowing down the system while maintaining high accuracy. Testing and evaluations on real and simulated vehicle data show promising results in terms of safeguarding modern vehicles from cyber-attacks and thereby improving passenger safety. 

Project by: Shafia Husna

Artificial Intelligence (AI) is now part of our everyday lives. It helps recommend movies, filter emails, and even assist doctors. But there’s a problem: many AI systems are like “black boxes.” They make predictions, but people can’t always understand why. This can make it hard to trust them, especially when decisions affect jobs, health, or safety.

My project aims to make AI both accurate and understandable. I built a system that designs and tests different AI models automatically, like a digital “evolution” where the best models survive. Each model is judged on two things: how often it makes correct predictions (accuracy) and how easy it is to explain its decisions (transparency).

The outcome shows that we can create AI that is both smart and clear. This helps people trust AI more, making it safer and more useful in the real world.

Project by: Tahmina Ahmed

Smartphones and wearable gadgets, such as iPhones, Apple Watches, and Fitbits, are widely used for tracking steps, movements, and health data. These devices are powerful, but it remains unclear how reliable their measurements are. This project aims to test that accuracy in a controlled setting. Instead of relying on humans, a custom-built mechanical rig will be designed to replicate repetitive human motions such as walking, while also investigating ascending stairs, and heart rates. Devices will be put on the rig, and their data will be compared to the machine's calibrated measurements. This technique will enable the team to determine when devices recorded movements correctly and when errors occur. Understanding these strengths and weaknesses is critical for applications such as forensic investigations, health monitoring and fitness tracking where accurate data can have substantial real-world effects.

Project by: 

• Sebastian Zillante 
• Kelvin Nguyen

Say I flip a coin five times and it lands heads every time. How much would you bet on the outcome of the next flip? What if I flipped five more heads in a row? Making an informed decision would require taking into consideration the uncertainty in the probability of flipping heads. Similarly, the capacity to quantify the uncertainty is useful in many areas of machine learning: the real-world is riddled with sources of uncertainty.

However, quantifying uncertainty is challenging for prominent models like neural networks. Traditional methods focusing on parameters yield unreliable/overconfident results due to the their complex nature and size.

To address such limitations, we explore an approach that allows direct uncertainty quantification in the function space (space where points correspond to rules that could be learned), meaning uncertainty is captured not by some model parameters but by its actual predictions and behaviour across inputs. 

Project by: Rahul Rajendra Kumar

Have you ever noticed how sometimes you sleep peacefully, and other times you toss and turn? Sleep is not just about resting—it is when our brain and body recover. Scientists know that certain hidden patterns in sleep, called Cyclic Alternating Patterns (CAP), can tell us if someone’s sleep is healthy or disturbed. The challenge is that these patterns are very hard for humans to spot because they happen quickly and in complicated ways.

Our project uses artificial intelligence (AI), a type of computer “brain,” to learn how to recognise these patterns automatically. We trained the AI using real sleep recordings and taught it to find CAPs more accurately than humans can.

The outcome of this project is a system that can help doctors better understand sleep problems, which could lead to improved diagnosis of sleep disorders and healthier lives for many people.

Project by: Kevan Ghorecha