Automated methods to find high-redshift quasars
File(s)
Author(s)
Lenz, Lena
Type
Thesis
Abstract
Quasars are among the most luminous objects in the Universe and as such can be found to have
existed within the first 700 million years after the Big Bang. This makes them useful for analyzing
the epoch of reionisation and the black hole mass evolution. They can be identified by a distinct
break in their spectral energy distribution (SED) around Lyα, which shifts into the near-infrared at
z ≳ 6.5. However, due to their rarity and large contaminating populations, only a handful of quasars
at z > 7 have been found in total so far. Current methods contain visual inspection of images from
photometric surveys as part of the classification process. Manually examining a significant part of the
candidates becomes infeasible for upcoming surveys such as the Large Synoptic Survey Telescope
(LSST) and Euclid, which are expected to produce ∼ 1011 − 1012 bytes of data each night. In this
work, we reduce the amount of required visual inspection by testing candidates against a model of
a stationary point source that follows a quasar SED at 5.2 ≲ z ≲ 12. Bayesian model comparison
based on colour information is used to retrieve a quasar probability and the best fit SED. To automate
this process, we implement a data processing pipeline that extracts relevant information from survey
databases and photometric images. This information is then used to apply heuristic cuts in order to
obtain high-redshift quasar candidates. To optimise the classification, the collected information is
passed to decision tree and random forest machine learning classifiers. We also feed the raw image
data to a convolutional neural network to compare performance. Our pipeline is implemented for
the Sloan Digital Sky Survey (SDSS) and the United Kingdom Infrared Telescope Infrared Deep Sky
Survey (UKIDSS), and can easily be extended to future surveys.
existed within the first 700 million years after the Big Bang. This makes them useful for analyzing
the epoch of reionisation and the black hole mass evolution. They can be identified by a distinct
break in their spectral energy distribution (SED) around Lyα, which shifts into the near-infrared at
z ≳ 6.5. However, due to their rarity and large contaminating populations, only a handful of quasars
at z > 7 have been found in total so far. Current methods contain visual inspection of images from
photometric surveys as part of the classification process. Manually examining a significant part of the
candidates becomes infeasible for upcoming surveys such as the Large Synoptic Survey Telescope
(LSST) and Euclid, which are expected to produce ∼ 1011 − 1012 bytes of data each night. In this
work, we reduce the amount of required visual inspection by testing candidates against a model of
a stationary point source that follows a quasar SED at 5.2 ≲ z ≲ 12. Bayesian model comparison
based on colour information is used to retrieve a quasar probability and the best fit SED. To automate
this process, we implement a data processing pipeline that extracts relevant information from survey
databases and photometric images. This information is then used to apply heuristic cuts in order to
obtain high-redshift quasar candidates. To optimise the classification, the collected information is
passed to decision tree and random forest machine learning classifiers. We also feed the raw image
data to a convolutional neural network to compare performance. Our pipeline is implemented for
the Sloan Digital Sky Survey (SDSS) and the United Kingdom Infrared Telescope Infrared Deep Sky
Survey (UKIDSS), and can easily be extended to future surveys.
Version
Open Access
Date Issued
2023-04
Date Awarded
2024-02
Copyright Statement
Creative Commons Attribution NonCommercial Licence
License URL
Advisor
Mortlock, Daniel
Leistedt, Boris
Publisher Department
Physics
Publisher Institution
Imperial College London
Qualification Level
Doctoral
Qualification Name
Doctor of Philosophy (PhD)