In recent years, artificial intelligence (AI) has been applied in many
fields of research. It is particularly well suited to astronomy, in which
very large datasets from sky surveys cover a wide range of observations.
The upcoming Legacy Survey of Space and Time (LSST) presents
unprecedented big data challenges, requiring state-of-the-art methods
to produce, process and analyse information. Observations of Type Ia
supernovae help constrain cosmological parameters such as the dark
energy equation of state, and AI will be instrumental in the next
generation of cosmological measurements due to limited spectroscopic
resources. AI also has the ability to improve our astrophysical understanding
by perceiving patterns in data which may not be obvious to
humans.

In this thesis we investigate how advanced AI methods can be used
in classification tasks: to identify Type Ia supernovae for cosmology
from photometry using supervised learning; by determining a
low-dimensional representation of stellar spectra, and inferring astrophysical
concepts through unsupervised learning.

In preparation for photometric classification of transients from LSST we
run tests with different training samples. Using estimates of the depth
to which the 4-metre Multi-Object Spectroscopic Telescope (4MOST)
Time-Domain Extragalactic Survey (TiDES) can classify transients, we
simulate a magnitude-limited training sample reaching rAB = 22.5 mag.
We run our simulations with the software snmachine, a photometric
classification pipeline using machine learning. The machine-learning
algorithms struggle to classify supernovae when the training sample
is magnitude-limited as its features are not representative of the test
set. In contrast, representative training samples perform very well,
particularly when redshift information is included. Classification
performance noticeably improves when we combine the magnitude-limited
training sample with a simulated realistic sample of faint,
high-redshift supernovae observed from larger spectroscopic facilities;
the algorithms' range of average area under ROC curve (AUC) scores
over 10 runs increases from 0.547-0.628 to 0.946-0.969 and purity of
the classified sample reaches 95% in all runs for 2 of the 4 algorithms.
By creating new, artificial light curves using the augmentation software
avocado, we achieve a purity in our classified sample of 95% in all
10 runs performed for all machine-learning algorithms considered. We
also reach a highest average AUC score of 0.986 with the artificial
neural network algorithm. Having real faint supernovae to complement
our magnitude-limited sample is a crucial requirement in optimisation
of a 4MOST spectroscopic sample. However, our results are a proof
of concept that augmentation is also necessary to achieve the best
classification results.

During our investigation into an optimised training sample, we assumed
that every training object has the correct class label. Spectroscopy is a
reliable method to confirm object classification and is used to define our
training sample. However, it is not necessarily perfect and we therefore
consider the impact of potential misclassifications of training objects.
Taking the predicted error rates in spectroscopic classification from the
literature, we apply contamination to a TiDES training sample using
simulated LSST data. With the recurrent neural network from the
software SuperNNova, we determine appropriate hyperparameters
using a perfect, uncontaminated TiDES training sample and then
train a model on its contaminated counterpart to study its effects
on photometric classification. We find that a contaminated training
sample produces very little difference in classification performance, even
when increasing contamination to 5%. Contamination causes more
objects of both Type Ia and non-Ia to be classified as Ia, increasing
efficiency, but decreasing purity, with changes of less than 1% on
average. Similarly, we see a decrease of 0.1% in average accuracy, and
no clear difference in AUC score, only varying at the fourth significant
figure. These results are promising for photometric classification.
Contaminated training appears to have little impact and propagation
to cosmological measurements is expected to be minimal.
In a separate study, we apply deep learning to data in the European
Southern Observatory (ESO) archive using an autoencoder neural
network with the aim of improving similarity-based searches using the
network's own interpretation of the data. We train the network to
reconstruct stellar spectra by passing them through an information
bottleneck, creating a low-dimensional representation of the data. We
find that this representation includes several informative dimensions
and, comparing to known astrophysical labels, see clear correlations
for two key nodes; the network learns concepts of radial velocity and
effective temperature, completely unsupervised. The interpretation
of the other informative nodes appears ambiguous, leaving room for
future investigation.

The results presented in this thesis emphasise the practical capabilities
of AI in an astronomical context: Classification of astrophysical objects
can be conducted through supervised learning using known labels, as
well as unsupervised learning in a physics-agnostic process.