STFC Centre for Doctoral Training in Data Intensive Science
The Astrophysics Research Institute at LJMU has a number of PhD studentships starting in 2022, which include a 4-year stipend to support you during your studies.
Prospective students should complete this application form for projects at Liverpool John Moores University (LJMU). Please indicate the specific project(s) you would like to apply for; this helps identify projects that are best suited for your research interests. However, general applications will also be considered.
Each project will give you access to LIV.INNO’s comprehensive training in data science, consisting of lectures, seminars, international workshops and schools. You will also get opportunities to contribute to the centre’s outreach activities. Each LIV.INNO student is required to undertake a 6-months placement in industry where you will work on a research challenge outside of your core PhD project. This exciting opportunity will help boost your wider skills and employability.
We welcome and encourage applications from the UK, EU and other parts of the world. We would like to encourage in particular applications from women and other STEM minority groups. LIV.INNO actively helps overcome barriers to access; qualifying students can receive additional funding for research-related travel costs. It is also possible to realize many of our PhD projects part-time, over a longer total period.
For further details about the LIV.INNO CDT, please see the LIV.INNO website.
Projects being offered for starters in Oct. 2022
1) Reconstructing the assembly history of our Galaxy using neural networks
Supervisors: Andreea Font (ARI, LJMU), Sandra Ortega Martorell (CSM, LJMU), Ivan Olier-Caparroso (CSM, LJMU)
Internship: Liverpool Heart and Chest Hospital (TBC)
In the standard cosmological model, galaxies form hierarchically by accreting and tidally disrupting smaller ('dwarf') galaxies over billions of years. The debris left from the destruction of satellite dwarf galaxies can be found today in the form of tidal streams in the stellar haloes of host galaxies. From the number and shapes of these stellar streams one can reconstruct the accretion history of galaxies and constrain the nature of dark matter.
The identification of tidal streams in our Galaxy is difficult because these features are extremely faint, and they also lose coherence quickly after a few orbits around the Galaxy. Machine learning (ML) techniques are critical in finding these features and in categorizing their shapes. In this project, we will use a combination of convolutional and recurrent neural networks (CNN and RNN) to extract the contextual information related to the tidal streams and to model their time evolution. We will train these networks on Artemis, a suite of state-of-the cosmological simulations that follows the formation of 45 galaxies like the Milky Way in a realistic cosmological context. This suite contains a large dataset of simulated tidal streams that will allow us to test and reduce the bias of current ML stream identification techniques. Additional improvements over the existing techniques will be achieved by including other information captured by these simulations, namely the kinematics and chemical abundance of the stars in the streams. By training the NNs on a multi-dimensional parameter space we will improve the identification of tidal streams, not only in the simulations but also in observational surveys (e.g. Gaia, LSST, Euclid).
The methods developed in this project will be also applied in the detection of degradation of artificial heart valve using cardiac ultrasound ("echocardiography") cine images. This collaboration with the Liverpool Heart and Chest Hospital will provide the largest, most detailed and longest running dataset of its kind in the world.
2) Using neural networks to learn how dark matter and dark energy affect structure formation in the Universe
Supervisors: Ian G. McCarthy (ARI, LJMU), Ian H. Jarman (CSM, LJMU)
The field of large-scale structure cosmology is presently on the cusp of a revolution, with a large number of imminent wide-field surveys (particularly LSST and Euclid) poised to make unprecedentedly precise measurements of the distribution of matter in the Universe in order to constrain the nature of dark matter and dark energy. To achieve these aims, these surveys will rely heavily on cosmological simulations to predict what the Universe should look like for different dark sector scenarios. However, cosmological simulations are very computationally expensive, particularly when non-gravitational (“baryonic”) interactions are incorporated. This prevents theoretical astrophysicists from systematically exploring the full parameter space associated with both baryonic and cosmological processes. In order to address these issues, we will use neural networks to: i) develop models for mapping the relations between less expensive gravity-only simulations and full cosmological hydrodynamical simulations including baryons; and ii) link the outcome of simulations (e.g., weak lensing mass maps) and the input parameters specifying the initial conditions and the nature of dark matter and dark energy. The neural networks will be trained and tested on a grid of simulations carried out on the LJMU HPC facility, Prospero. The derived models will be used to produce accurate predictions for LSST and Euclid.
3) Using neural networks and clustering algorithms to understand the mass flows and energy cycles at the heart of our Galaxy
Supervisors: Steve Longmore (ARI, LJMU), Qizhou Zhang (CfA, Harvard University)
The inner few thousand light years of the Milky Way – the Central Molecular Zone (CMZ) – hosts the nearest supermassive black hole, largest reservoir of dense gas, most massive/dense stellar clusters, and highest volume density of supernovae in the Galaxy. As the nearest environment for which it is possible to simultaneously observe many of the extreme physical processes shaping the Universe, it is one of the most well-studied regions in astrophysics. However, the potential of the CMZ as a laboratory of extreme physics is fundamentally limited by the lack of a unified framework to understand how global processes determine the location, intensity and timescales for star formation and feedback. This joint PhD project with Harvard will be based on analysis of large datasets that we’ve been awarded on the world’s foremost mm-wave telescope, ALMA, intended to overcome this limitation. This project will produce the largest (250Tb), highest velocity resolution map of the most physically, chemically and kinematically complex region in the sky. The enormity and complexity of these data make the analysis particularly challenging and very well suited to a data intensive PhD project.
4) Constraining the complex relationship between galaxies and their dark matter haloes with machine learning
Supervisors: Rob Crain (ARI, LJMU), Ivan Baldry (ARI, LJMU), Paul Bell (CSM, LJMU)
Detailed numerical simulations of the formation of cosmic large-scale dark matter structure show that the spatial clustering of dark matter haloes (the sites where galaxies form) depends primarily on their mass, but also significantly on secondary properties such as their assembly history or local environment. New state-of-the-art gas-dynamical numerical simulations indicate that these secondary properties also have an important influence on the properties of the galaxies forming within. However, owing to the diversity and complexity of the physics underpinning galaxy evolution, a consensus view of precisely how the assembly history and environment of dark matter haloes influences the evolution of galaxies has yet to emerge.
Machine learning (ML) presents an exciting opportunity to disentangle the complex relationships between the diverse present-day populations of galaxies and the dark matter haloes. Whilst ML has seen widespread adoption in astrophysics as a means to predict observable galaxy properties based on halo catalogues drawn from detailed, but physically “simple” dark matter-only simulations, it can also be used in the reverse sense: to predict halo properties based on galaxy properties. Here, we aim to train ML models using state-of-the-art gas-dynamical simulations (e.g. EAGLE-XL, IllustrisTNG-300) to learn the complex relationships between galaxies and haloes that each simulation predicts. The models will then be used to create mock galaxy catalogues that will be confronted with the rich, multidimensional data yielded by the ongoing and forthcoming observational “megasurveys” (spectroscopic, e.g. WAVES, DESI, and deep/detailed imaging, e.g. LSST, Euclid), which are set to revolutionise extragalactic astrophysics over the coming decade, enabling the discrimination between leading models and placing unprecedentedly stringent constraints on galaxy formation theory.