Newcastle University

Application deadline: Monday 13 April

Find out about our internships, who can apply, and the stipend

Prospective projects

Below is a list of prospective projects you can apply for, complete with a short explanation and the lead supervisor's name and department. Please contact them before making your application; you will be asked if you have done this on the application form.

If you are interested to learn more about a specific project before applying, download the full project proposal at the bottom of the page.

From Simulation to AI: Physics-Informed Neural Networks for Chemical Kinetics in Turbulent Hydrogen Swirling Flames. Umair Ahmed, School of Engineering

This project investigates whether physics-informed neural networks (PINNs) can predict chemical kinetics in turbulent hydrogen swirling flames. Hydrogen combustion is promising for net-zero energy because it produces no carbon emissions, but hydrogen–air flames behave differently from hydrocarbon fuels and can be unstable under swirling conditions. Using direct numerical simulation (DNS) datasets from an ongoing EPSRC project and pretrained neural networks, the student will benchmark AI surrogate models, assessing prediction accuracy and physical consistency.

VERIFAI: A Scalable Verification Framework for Transparent AI Models. Alex Chan, School of Engineering.

This project investigates how high-performance computing (HPC) supports verification and explainability of machine learning (ML) models in complex decision-making systems. The student will explore methods on analysing how these ML models arrive at specific outcomes, and develop prototype tools to extract explainable representations of these decision processes. Here, HPC enables efficient state exploration involving combined large inputs and model behaviour, which the project will evaluate to improve transparency and trust in AI systems.

Explainable Health Data Analytics with Tsetlin Machines. Ekin Can Erkus, Electrical and Electronic Engineering

This project builds a reproducible, HPC-enabled pipeline for explainable health data analytics using routinely collected hospital records, with MIMIC-IV as a public testbed. The intern will train Tsetlin Machines to learn transparent rule-based predictors and will benchmark them against standard baselines under matched pre-processing. The evaluation includes fairness across patient subgroups, and explanation stability under missingness and resampling. Outputs include an auditable codebase, automated explainable model that supports deployable, trustworthy prediction from healthcare data.

Scaling up an interactive web app of insights into the data of a digital charity. Clement Lee, Lecturer in Statistics

Freegle is a digital exchange platform where users give away household items to other users. It is operationally beneficial to know how far an average user travels for items, which has been visualised by a prototype web application. Implementing this measure in Freegle’s systems requires results of a higher spatial resolution, which are naturally computationally intensive. It is also desirable to incorporate the idea of isochrone, which further adds to the computational cost. This project will equip the student with the skills of using R to simultaneously create an interactive application and tackling the computational requirements.

Deployed End-to-End Machine Learning on an Autonomous Racing Car, Chris Holder, Computing

While enormous progress has been made towards autonomous road vehicles, translating this to the domain of motorsport, where vehicles operate at much greater speed at the edge of their dynamic limits, poses significant challenges. End-to-end machine learning, where an AI model learns to translate sensor data directly to vehicle controls offers a potential solution to these complex challenges. This project seeks to deploy an end-to-end machine learning model onboard an F1Tenth small scale autonomous racing car, with the goal of developing a system capable of successfully completing laps on previously unseen track layouts fully autonomously.

Just-in-time compute for Neural Architecture Search Benchmarks in Teaching and Research. Stephen McGough, School of Computing

Neural Architecture Search (NAS) is a Deep Learning process for identifying the ‘best’ architecture for a particular problem and dataset. Due to the ‘No free lunch’ proposition there is no single architecture which will be ‘best’ in all cases. One of the research directions in NAS is to produce benchmarks where you train thousands of architectures on a dataset recording the performance. This benchmark can then be used to try out different NAS approaches. We need more of these for teaching. But rather than pre-compute all architectures we propose a just-in-time approach where architectures are only trained when needed.

Data Pre-Processing and Generation for Real-World Neural Architecture Search Applications. Stephen McGough, School of Computing

A big disparity between real-world and academic data science projects is the availability and reliability of the data. CIFAR-10, a common benchmark dataset, contains images all the same size and modality, it is perfectly class balanced, and contains no missing values. Working with our industry partners we are working to create high-quality training datasets and accurate models. One of the aims is to explore and enhance their data, creating clear labels, and generating data samples to make up for the missing use cases and biased training sets. We will investigate how style transfer can be used to expand these datasets.