The Application of Machine Learning to Source Detection and Classification of Variable Stars with TESS - Master’s Thesis

Aug 5, 2026

Machine learning is a powerful tool for modern astronomers, enabling effective processing of the large data volumes produced by advanced telescopes and surveys. This Master’s thesis applies machine learning methods to detect and classify sources observed by TESS, supporting the TESSELLATE pipeline for constructing an open variable-sky catalogue. A convolutional neural network, SourceDetect, achieves high positional recovery and identification accuracy, enabling robust recovery near the model magnitude limit. A random-forest classifier, Classifind, provides accurate variable-star characterisation, and an exoplanet-candidate pipeline extends the workflow by identifying new candidates from detections.

LanguageEnglish

CategoryResearch & Report

FormatPDF

Pages119

Reading Time5 h 0 min

AccessFree to download

Contents

1 Introduction

1.1 Overview

1.2 TESS

1.2.1 Background

1.2.2 Machine Learning with TESS

1.2.3 Ongoing work at UC

1.2.4 Motivation for this Research

1.3 Convolutional Neural Networks

1.3.1 Layers

1.3.2 Loss Functions

1.4 Random Forest Classifiers

1.5 Variable Stars

1.5.1 Instability Strip Variables

1.5.2 Other Variables

The Application of Machine Learning to Source Detection and Classification of Variable Stars with TESS Andrew Moore Supervisor Dr. Ryan Ridden-Harper A thesis submitted for the degree of Master of Science in Astronomy February 2025 School of Physical and Chemical Sciences Abstract Machine learning is a powerful tool for modern astronomers. As telescopes and astronomical surveys become more advanced, we require more powerful methods to process the large volumes of available data. In this thesis, we discuss the application of machine learning techniques to the detection and classification of objects captured by the Transiting Exoplanet Survey Satellite (TESS) . We completed this work primarily for application in the TESSELLATE pipeline which will create a complete, open catalogue of the variable sky observed by TESS. We implemented a convolutional neural network (CNN) to develop an object detection method for sources within reduced TESS images. Our method, SourceDetect, was trained on true and false sources from real TESS images and yielded an accuracy of 96 . 12% during testing. Upon application to real images injected with TESS pixel response functions, SourceDetect reliably recovered sources brighter than Tmag ≈ 17; the approximate magnitude limit of our model. SourceDetect identified 94 .24% of the recovered sources within a precision of 0 .5 pixels in both positional dimensions. These results could be improved through amendments to the training set and CNN itself but demonstrate that SourceDetect can identify large volumes of scientific targets for the TESSELLATE survey, or other purposes. Detected sources require classification to become scientifically useful. Transients and asteroids were accounted for in other work, but TESSELLATE required a variable star classifier. Therefore, we developed a variable star characterisation method, Classifind, based on a random forest classifier (RFC) . The RFC was trained and tested on characteristic lightcurve features of sources identified with SourceDetect across 6 variable classes. Formal dataset testing returned an accuracy of 96 .7% and further testing on 5000 SourceDetect lightcurves yielded an accuracy of 96 .2% . Approximately 80% of classifications had a likelihood of over 75% and we found that some lightcurves above approximately 50% likelihood were still correctly classified. Classifind would benefit from a larger training set but has proven capable of yielding accurate variable star classifications of TESS lightcurves. SourceDetect found a potential exoplanet candidate that has not been classified in previous work. The target, SIPS J0324-5945, is an M-dwarf star with periodic dips corresponding to an exoplanet of radius Rp = 0 .89 ± 0.02RJup. We appended on previous work to create an exoplanet candidate pipeline, ExoHunter, to manually identify more candidates from SourceDetect detections. ExoHunter yielded tens of potential candidates across TESS sectors 28, 29, 30 , and 31 . Some of these sources were verified as known exoplanets using ExoplanetOrbitDatabase, including 8 of the 27 known exoplanets for TESS cameras 3 and 4 in sector 29 . This verified the capability of ExoHunter to identify exoplanet candidates. We plan to gather follow-up observations for three candidates at Mt John Observatory in April 2025 and others at later dates. DECLARATION I declare that the work in this Master’s dissertation has been composed solely by myself, except where otherwise stated by reference or acknowledgement. All TESS images used during this work were obtained through the TESSELLATE pipeline ([https://github. com/rhoxu/TESSELLATE](https://github. com/rhoxu/TESSELLATE) ), which extracts calibrated and reduced photometric data for every full-frame image in the TESS archive. TESSELLATE also utilises the TESSreduce difference imaging pipeline (https://github. com/CheerfulUser/TESSreduce) . My work in this thesis contributed significantly to the source detect

FAQs

What is the goal of this thesis in relation to TESS?

The thesis develops machine learning methods to detect and classify sources captured by TESS, mainly to support the TESSELLATE pipeline for creating an open variable-sky catalogue.

How does SourceDetect work and what performance does it achieve?

SourceDetect uses a convolutional neural network to detect sources in reduced TESS images. Testing reports about 96.12% accuracy, and recovery of sources brighter than Tmag ≈ 17 with about 94.24% identification of recovered sources within 0.5 pixels.

How are variable stars classified and what results were obtained?

A variable star classifier, Classifind, is built using a random forest classifier trained on lightcurve features across six variable classes. Dataset testing reports about 96.7% accuracy, with further testing on 5000 SourceDetect lightcurves yielding about 96.2%, including high-likelihood classifications.

Aug 5, 2026

Upload by River Wang