Extracting precise information from large observational data is playing a vital role in modern astronomy. In particular, machine learning as a promising approach for the efficiency in gaining precise estimations on large astronomical data has become a growing field. In this thesis I have applied different machine learning algorithms to estimate redshift and quantities of dust from large photometric data sets of different astrophysical objects. Hereby I also investigated, if such machine learning methods provide improved estimates of the quantities and properties, compared to other, commonly used methods.
First, I estimate photometric redshifts by applying a method that contains two
probabilistic algorithms on a data set from the Sloan Digital Sky Survey cross matched with the Wide-field Infrared Survey Explorer. I employ an infinite Gaussian mixture model (IGMM) to classify the observed objects in the data set as three different classes (i.e. stars, galaxies and quasars). Then, I estimate probability density functions (PDFs) of photometric redshifts for the objects in the data set, by feeding the outcome of the probablistic classification into a mixture density network (MDN). Secondly, to infer quantities of dust in supernovae, I use a simulated data set of supernova spectral energy distributions calculated by an advanced fully three-dimensional radiative transfer code, MOCASSIN, to design three scenarios for estimating quantities and properties of dust in and around supernovae with synthetic magnitude adjusted to James Webb Space Telescope (JWST) telescope bandpass filters. I developed a neural network to estimate
the dust properties along with the corresponding predicted uncertainties that reflect the sufficiency of the given photometric information for representing the dust quantities.
By exploring a rich parameters space that is covered by the simulated sample, this trained algorithm became a promising approach to estimate dust properties with JWST photometric observations. Moreover, by implementing a feature selection framework on the neural network predictions I have found a minimum set of JWST filters for inferring dust in supernovae that could benefit future observational strategies. Moreover, as an ongoing project for improving estimating photo-zs, I apply two different methods as alternatives to two different parts of the photo-z estimator in the first project. First, instead of classifying objects in the feature space, I classify the objects in a latent space
by implementing a variational auto-encoder instead of an IGMM. Secondly, for reflecting the measurement errors in the estimated PDF photo-zs, I employ an importance sampling method instead of including the measurement errors as additional input features to the MDN, which may lead to more precise photo-z estimations.