Search the portal

Please enter a term

Full record

« Back to home page
TitleDecoding astronomical spectra using machine learning
Authorde Mijolla, Damien
AbstractSpectroscopy is one of the cornerstones of modern astronomy. Using spectra, the light from far-away objects measured on Earth can be related back to the physical and chemical conditions of the astronomical matter from which it is emitted. This makes spectroscopy an essential tool for constraining the physical and chemical conditions of the matter in stars, gas, galaxies and all other types of astronomical objects. However, whilst spectra carry a wealth of astronomical information, their analysis is often complicated by difficulties such as degeneracies between input parameters and gaps in our theoretical knowledge. In this thesis, we look towards the rapidly growing field of machine learning as a means of better extracting the information content of astronomical spectra. Chapters 2 and 3 of the thesis are dedicated to the study of spectra originating from the interstellar medium. Chapter 2 of this thesis presents a machine learning emulator for the UCLCHEM astrochemical code which when combined with a Bayesian treatment of the radiative-transfer inverse problem enables a rigorous handling of the degeneracies affecting molecular lines (all within short enough computational timescales to be tractable). Chapter 3 extends upon the work of Chapter 2 on modelling molecular lines and investigates the appropriateness of Non-negative Matrix Factorization, a blind source separation algorithm, for the task of unmixing the gas phases which may exist within molecular line-intensity maps. Chapter 4 and 5 are concerned with the analysis of stellar spectra. In these chapters, we introduce machine learning approaches for extracting the chemical content from stellar spectra which do not rely on manual spectral modelling. This removes the burden of building faithful forward-models of stellar spectroscopy in order to precisely extract the chemistry of stars. The two approaches are also complimentary. Chapter 4 presents a deep-learning approach for distilling the information content within stellar spectra into a representation where undesirable factors of variation are excluded. Such a representation can then be used to directly find chemically identical stars or for differential abundance analysis. However, the approach requires measurements of the to-be-excluded undesirable factors of variation. The second approach which is presented in Chapter 5 addresses this shortcoming by learning which factors of variation should be excluded using spectra of open clusters. However, because of the low number of known open clusters, whilst the method constructed in Chapter 4 is non-linear and parametrized by a feedforward neural network, the approach presented in Chapter 5 was made linear.
TypeThesis; Doctoral
PublisherUCL (University College London)
Source Doctoral thesis, UCL (University College London).