The Open University
Unveiling the Hidden Cosmos: Machine Learning for Astronomical Deconvolution and Cirrus Noise Removal
Abstract
dc:description.abstractThis thesis explores the application of machine learning, specifically convolutional autoencoders, to enhance the resolution of far-infrared astronomical images affected by interstellar cirrus emission.Cirrus, a filamentary foreground structure composed of cold dust, is a significant source of confusion noise that limits source detection and photometric precision. By combining deep learning with realistic synthetic data, this work aims to improve the quality of deconvolved images and enable the detection of faint astrophysical sources. The study builds on previous research using Herschel and James Clerk Maxwell Telescope data, extending the approach to more challenging observational scenarios. Synthetic cirrus was generated using fractional Brownian motion (fBm) simulations to approximate the statistical properties of interstellar dust. These simulations were used to train autoencoder models capable of distinguishing between compact sources and structured foreground emission. Models trained on fBm-augmented datasets demonstrated superior performance compared to those trained on clean data, achieving better source recovery, positional accuracy, and photometric fidelity. This work highlights the potential of combining physical simulations with machine learning to overcome the limitations imposed by confusion noise in far-infrared astronomy. The methods developed here provide a foundation for integrating machine learning-based deconvolution into standard astronomical data processing pipelines, with the potential of supporting ongoing and future analysis with instruments such PRIMAger, but also making use of archived data. Overall, the research presented demonstrates that deep learning frameworks, when trained with synthetic data, can significantly improve the quality of far-infrared astronomical images, possibly leading to new discoveries in galaxy evolution and star formation studies. Parts of this work have been published in Donnellan et al., 2024b.
Degree
thesis:*- Level dc:type.qualificationlevel
- mphil
- Grantor dc:publisher.institution
- The Open University
- Year dc:date.issued
- 2026
Author and committee
dc:creator, dc:contributor.*- Author dc:creator
-
- Sorrell, Christopher
Rights
- Language dc:language
- en