Student papers

I have been fortunate to work with excellent postgraduate students over the years, both as main supervisor and as collaborator of international PhD students. In all cases, I take my role as mentor seriously, doing all I can to train the next generation of bright scientists. Several students, both MSc and PhD, have published papers based on my ideas and with my guidance. I detail these projects here.

Astronomaly at Scale: Searching for Anomalies Amongst 4 Million Galaxies

I developed Astronomaly to identify anomalous sources in massive datasets, though initial tests were limited to tens of thousands of sources—datasets still manageable for manual inspection. To push the framework’s capabilities further, I assigned my PhD student, Verlon Etsebeth, to test it on a much larger scale. I conceptualised the project, guided him throughout, and edited the resulting paper (Etsebeth et al., MNRAS, 2024). Verlon applied Astronomaly to 4 million optical galaxies from the DECALS survey, employing a novel feature extraction method and stress-testing the machine learning algorithms. The results were impressive: Astronomaly successfully identified strong lens candidates, unusual merging systems, and several unclassified sources. This work paves the way for deploying Astronomaly on the massive datasets anticipated from the SKA and LSST. Verlon is now advancing this research with a second paper, applying Astronomaly to detect diffuse sources in MeerKAT data. His progress underscores his potential for an outstanding scientific career.

Enabling Unsupervised Discovery in Astronomical Images through Self-Supervised Representations

Feature extraction, the process of converting complex data like images into simpler numerical representations for machine learning algorithms, is both the most critical and most challenging step in anomaly detection. Recognising this, I explored cutting-edge machine learning techniques, specifically self-supervised deep learning, to enhance feature extraction beyond the methods used in our original Astronomaly paper (Lochner & Bassett, 2020). I tasked my MSc student, Koketso Mohale, with testing state-of-the-art self-supervised methods on the public Galaxy Zoo dataset of optical galaxies. Among these, the technique Bootstrap-Your-Own-Latent (BYOL) proved the most effective. Koketso applied BYOL to galaxy images, enabling a clustering algorithm to automatically group galaxies of the same class without human intervention. Building on this, I applied Astronomaly to these refined features, achieving excellent anomaly detection performance, which demonstrated the transformative potential of self-supervised learning for astronomical data. Together, we co-authored a paper (Mohale & Lochner, MNRAS, 2024), with significant contributions from both of us in writing, analysis, and visualisations. Koketso has since continued with me as a PhD student, focusing on developing a novel method to evaluate the effectiveness of feature extraction for unsupervised machine learning tasks.

Classification of Multiwavelength Transients with Machine Learning

Transient classification is a critical challenge in modern time-domain astronomy, as the vast volumes of data generated by current and future surveys demand robust and efficient automated methods. I supervised my MSc student, Kimeel Sooknunan, in developing a machine learning framework to classify transients by integrating data from radio and optical telescopes, building on the foundation of Lochner et al. (ApJS, 2016). This approach is particularly relevant to telescopes like MeerLICHT, which operates alongside MeerKAT to capture simultaneous optical and radio observations. Kimeel implemented the algorithms, fine-tuned the models, and conducted a comprehensive analysis of the results, while I conceptualised the project, provided ongoing guidance, and refined the manuscript. Our work, published in Sooknunan et al., (MNRAS, 2021), demonstrated the power of machine learning to distinguish between radio transient types, underscored the importance of multiwavelength data for classification, and laid the groundwork for real-time classification in the era of LSST and the SKA. Following this success, Kimeel earned a prestigious President’s Scholarship and completed his PhD at Imperial College London.

Unsupervised Machine Learning for Transient Discovery in Deeper, Wider, Faster Light Curves

In 2019 I was on the Scientific Organising Committee for the 6-week Kavli Summer Program for Astrophysics (KSPA) in Santa Cruz, USA, where the focus that year was on machine learning. KSPA aims to bring together top PhD students from around the world to work with senior researchers on cutting edge research projects in a short timeframe. During the programme, I worked with Sara Webb, an exceptional PhD student from Swinburne University, Australia. Together we crafted an ambitious project to apply unsupervised machine learning to the enormous dataset of light curves from the Deeper-Wider-Faster programme. The aim was to try to detect anomalous objects in the data and perhaps discover something new. Under my mentorship, Sara analysed over 60 000 light curves, finding several new variable stars and an unusual ultra-fast flare. She published an excellent first-author paper (Webb et al., MNRAS, 2020), successfully graduated her PhD and now has a permanent position at Swinburne University.

Real-time detection of anomalies in large-scale transient surveys

At the KPSA I also worked with the highly talented Daniel Muthukrishna, a PhD student from Cambridge University, UK. I suggested Daniel focus on the highly challenging problem of detecting anomalous astronomical sources in real-time. I mentored and worked with Daniel for several years resulting in a fascinating paper (Muthukrishna et al., MNRAS, 2022) and a proceedings at a highly prestigious machine learning conference (Muthukrishna et al., NeurIPS, 2021). As Daniel worked as a postdoc at MIT, he began to supervise an extraordinary young high school student from California, Rithwik Gupta. It was a delight to see the student become a mentor in his own right and together we published a follow-up paper (Gupta, Muthukrishna & Lochner, RASTI, 2024).

TEGLIE: Transformer encoders as strong gravitational lens finders in KiDS

I met Margherita Grespan, a PhD student at the National Centre for Nuclear Research, Poland, when I taught at the Tonale Winter School on Cosmology in 2022. Margherita reached out to me after the school asking if she could visit my group using some funding that she had secured. I was delighted to accept and she stayed with my group in South Africa for several months, working with me on a project to use machine learning-based techniques to look for strong lenses in KiDS data. This is an extremely challenging project as lenses are incredibly rare, less than 1 in 10 000 sources, and she was working with fairly raw, real data not simulations. I’d actually consider this to be one of the most challenging “needle-in-a-haystack” problems I have ever worked on. She published her first paper (Grespan et al., A&A, 2024) focusing on using supervised machine learning to locate strong lenses and I acted as editor to help train her in writing her first (excellent) paper. She subsequently returned to using Astronomaly to search for strong lenses as an alternative technique and is currently completing the draft. In the meantime, Margherita graduated with her PhD and will soon start a postdoc at Oxford, UK.

Page updated

Google Sites

Report abuse