Astrophysical Light Curve Analysis in the Era of Big Data

With next-generation observatories such as the Vera C. Rubin Observatory, ESA[’]{dir=”rtl”}s Euclid mission, the Nancy Grace Roman Space Telescope, and the Square Kilometre Array (SKA), astrophysics is witnessing a revolution in data availability. These powerful instruments will generate petabytes of data annually, requiring innovative analytical techniques to efficiently process and interpret light curves — time-series measurements of the brightness of celestial objects. Analysing these light curves is crucial for understanding transient astrophysical events like supernovae, variable stars, and exoplanet systems. Three difficult challenges will arise from these new facilities: 1) in some situations data rates may be so large that only “high-level” data products will be stored, therefore requiring the analysis of “intermediate-level” data products (e.g. to identify transients) to be both efficient and reliable before the information is lost, 2) large data volumes will render the analysis/modelling of individual sources (such as periodic binaries) impossible/ineffective through “conventional” methods such as Bayesian analysis, 3) fusion of heterogeneous datasets (from different facilities) will be desired to maximise scientific return. The “Astrophysical Light Curve Analysis in the Era of Big Data” PhD project will attempt to make in-roads to tackle these challenges. This research will focus on taking advantage of state-of-theart techniques such as simulation-based inference (SBI) and normalizing flows with a goal to develop physics-informed neural networks (PINNs) to improve astrophysical time series analysis, with applications in areas such as fast radio bursts, periodic variables, and (exoplanet) transits. The use of SBI and PINNs is aimed at creating the paradigm shift necessary to tackle the three challenges. For instance, we want to investigate if the implementation of PINNS into fast radio bursts detection, for instance through the behaviour related to the wavelength-dependent dispersion and scattering laws could provide more a robust automated selection of events (challenge 1). Likewise, the search for exoplanet transits is an area which should gain from implementing PINNS, as their signals should obey patterns that could at least be written a set of empirical differential equations (challenge 2). In the area of periodic variables, much attention has been devoted to source classification. The characterisation of these sources, however, is a very expensive task which tends to rely on Bayesian parameter inference using MCMC methods. By moving towards an SBI approach, it would become possible to implement characterisation of variables on large scales thus providing opportunities to perform population studies, but also to identify targets of interest for further follow-up based on “physical” parameters rather than empirical features (challenge 2). We could even envision moving towards an SBI-based classification which would embed characterisation outright as a one-step solution. Normalizing flows Quantifying uncertainties in parameter inference obtained with this approach would also constitute a long-term goal as to enable statistically meaningful statements to be made. Finally, enabling the fusion of heterogeneous datasets within these would be an ambitious objective to realise as it would fully leverage the inference capabilities (challenge 3). A possible first approach could rely on incorporating prior information through feature engineering.

Desirable Student Background:

Background in Astrophysics/Physics or Mathematics, with good programming experience in Python (and possibly machine learning frameworks), some knowledge of at least one of the following areas: linear algebra, Bayesian inference, deep learning, optimisation methods.

How to apply:

Students should apply through the below link under the PhD in Artificial Intelligence CDT program:

https://pgapplication.manchester.ac.uk/psc/apply/EMPLOYEE/SA/s/WEBLIB_ONL_ADM.CIBAA_LOGIN_BT.FieldFormula.IScript_Direct_Login?Key=UMANC1251000021489F