Research theme
Machine-learning methods for scientific discovery in the physical sciences — expressive generative models, simulation, inference and tools that let AI act as a collaborator in physics analyses.
Machine learning is reshaping how we do science. I'm interested in the practical edge of that change — building methods that don't just process data, but help us decide what to measure, design experiments, and reason about uncertainty. As a Turing Fellow at the Alan Turing Institute I lead a programme developing ML tools for scientific discovery in the physical sciences.
Our group has developed expressive generative models that can both simulate physical processes and infer their parameters at once, showing how the same model can serve as simulator and as analysis engine. We have applied similar ideas to liquid-argon detector simulation, using generative networks to accelerate the slow Monte-Carlo simulations that bottleneck rare-event searches.
Beyond modelling, I have worked on machine learning as a scientific advisor — algorithmic systems that propose new measurements or new configurations of an experiment, rather than just classifying events. This ranges from active-learning approaches in particle physics to AI-augmented experimental design.
Current threads include agentic data-engineering systems for open-data re-analysis, generative simulation for next-generation dark-matter detectors, and a wider programme of training and skills development — through the 4IR Centre for Doctoral Training in Data Intensive Science — that aims to make this kind of work routine across physics, astronomy and beyond.
An auto-regressive Gaussian-mixture density model designed for problems where the observables have hard physical boundaries and depend conditionally on external (theory) parameters — two features that defeat most off-the-shelf generative models. Projecting onto a configurable latent space gives the network enough expressive power to (a) constrain physical parameters from data, (b) sample as a Monte-Carlo surrogate, and (c) estimate likelihood ratios for event classification, all from one trained model. Demonstrated on anomalous electroweak Z+jj production at the LHC.
Builds high-capacity density models for high-energy physics by designing a custom latent space tailored to the structure of physical observables, improving on generic deep generative architectures for simulation and inference.
Current
Principal Investigator
Diverse interests across particle physics, software development and data analysis at colliders and in direct dark matter detection, high-frequency gravitational waves, neutrino physics, phenomenology, instrumentation, and the deployment of AI techniques.
PhD student (2023–)
Advancing direct detection frontiers for dark matter, new neutrino physics and high-frequency gravitational waves through detector modelling, optimisation and novel sensor technologies.
Alumni
Postdoctoral Research Associate
Vector boson fusion studies at the ATLAS experiment.
PhD student
Fast inference techniques in the physical sciences.
MPhys students · 2020–21
Applications of data science to HSBC UK Asset Finance.
MPhys students · 2020–21
Applications of data science to HSBC UK Asset Finance.
MPhys students · 2019–20
Generative adversarial network approaches to enhance simulation of liquid-argon time projection chambers and dark matter detection sensitivity.
MPhys student · 2017–18
Future release of the CERN ATLAS 13 TeV proton–proton Open Data.