Ryan Roberts

I am an Astrophysics PhD student at the LJMU Astrophysics Research Institute in Liverpool. Alongside this, I am also currently working as a data scientist at Aspia Space as my industry placement for the LIV.INNO Centre for Doctoral Training. My astrophysics research focuses on the application of machine learning models to cosmological hydrodynamical simulations, as we seek to investigate the complex relationships between galaxies and their dark matter haloes. At present, we are probing the physical origin of scatter in the circumgalactic gas mass fraction (f_CGM) of galaxies at fixed halo mass (M_200). We are utilising an XGBoost (eXtreme Gradient Boosting) model in tandem with SHAP (SHapley Additive exPlanations), a powerful explainable AI method for retroactively calculating feature importances in machine learning models by utilising probabilistic game theory. These combined methods are enabling us to go further than with less sophisticated techniques such as correlation coefficients, by not only establishing correlation strengths but the relative importance of various properties in establishing the target property.

I am also a member of the Aspia Space team, an Earth-observation and data science company based in Cornwall, UK. As part of their team, I assist in developing machine learning models that are used to extract useful and meaningful information from big data sets, such as Sentinel-2 satellite data and outputs from their flagship ClearSky model. One of the main projects I have worked on is the release of AstroPT, a set of GPT-like "Large Observation Models (LOMs)" trained on 8.6 million galaxy cutouts from the DESI DR8 survey, in which we trained the largest model of the group at 2.1 billion parameters on LJMU's Prospero supercomputer. One of the aims of the project is to continue developing AstroPT on big data sets (such as the vast quantity of publicly-available astronomy data) and translate the techniques used to EarthPT, a model analogous to AstroPT but instead focusing on Earth-observation data. AstroPT is an open-source collaborative project, and by continually developing it we hope to improve Aspia Space's in-house EarthPT model.

The AstroPT release paper can be found here: https://arxiv.org/abs/2405.14930

My GitHub page can be found here: https://github.com/RJ-Roberts

Degrees

2022, University of Liverpool, United Kingdom, MPhys Astrophysics with 1st Class Honours