Understanding of open stellar clusters and peculiar spectra in sky surveys using machine learning

Čotar, Klemen (author), Zwitter, Tomaž (mentor)

The increasing number of fully automated and highly multiplexed telescopes produces an ever-increasing number of observations whose reduction and analysis tend to be as automated as possible. As a limited number of researchers is unable to process and check all collected information manually, this degree of automation will steadily increase with time. This does not mean that used processes are flawless, and science will be in the near future handed to machines. Quite the opposite, the complexity of machinery and investigated scientific topics still requires a person to tweak the processes, identify previously unrecognised problems and think about the next scientific problem. Machine learning tools are not magical solutions as they only try to solve the problem introduced by the operator. Therefore if a question is poorly defined or data improperly prepared and filtered, results will also be burdened by those decisions. This presents an important need of not only knowing the data at hand but also how were they prepared and possibly quality flagged during production. This thesis presents the results of our exploration of extensive GALactic Archaeology with HERMES (GALAH) data set with various machine learning tools and emphasizes the need for control and understanding of input data. Presented results deal with the exploration of 16 open stellar clusters and their chemical signature, where the main problems are homogeneity and correctness of the determined chemical compositions. By understanding the limitations of chemically tagging open cluster stars with homogeneous composition, the findings were applied to the chemical exploration of their surrounding. The second large part of this thesis deals with finding smaller subsets of peculiar spectra in the GALAH data set. By various approaches of comparison between observed and modelled normal spectra, we found 918 spectra with pronounced molecular bands of C2 molecule, 10,364 spectra with visible emission lines and analysed 329 possible spectroscopically unresolved multiple systems.

• Language: English
• Keywords: astronomy, open clusters, peculiar stars, binaries, classification, machine learning
• Work type: Doctoral dissertation
• Typology: 2.08 - Doctoral Dissertation
• Organization: FMF - Faculty of Mathematics and Physics
• Year: 2020
• PID: 20.500.12556/RUL-121266
• COBISS.SI-ID: 31512067
• Publication date in RUL: 02.10.2020

Secondary language

• Language: Slovenian
• Title: Razumevanje razsutih zvezdnih kopic in posebnih tipov spektrov v pregledih neba z uporabo strojnega učenja

Abstract

Naraščajoče število popolnoma avtomatiziranih in visoko multipleksiranih teleskopov nam posreduje vedno večje število opazovanj, katerih začetna obdelava in analiza sta vedno bolj in bolj avtomatizirani. Ker omejeno število raziskovalcev ne more ročno obdelati in preveriti vseh zbranih informacij, se bo stopnja avtomatizacije s časom le še povečevala. To ne pomeni, da so uporabljeni procesi brezhibni in bo znanost v bližnji prihodnosti predana računalnikom. Ravno nasprotno, zapletenost teleskop in opazovanj ter znanstvenih vprašanj še vedno zahteva, da operator prilagodi procese, ugotovi prej neopažene težave in razmišlja o naslednjem znanstvenem problemu. Orodja za strojno učenje niso čarobna orodja, saj le skušajo rešiti težavo, ki si jo je zamislil njihov uporabnik. Če je vprašanje slabo opredeljeno ali so podatki nepravilno pred-pripravljeni in filtrirani, bodo tudi rezultati obremenjeni z vhodnimi odločitvami. To predstavlja pomembno težnjo po poznavanju ne samo razpoložljivih podatkov, ampak tudi kako so bili pripravljeni in označeni z raznimi statusi njihove kvalitete. Disertacija predstavlja rezultate našega raziskovanja obsežne podatkovne baze GALactic Archaeology with HERMES (GALAH) z različnimi orodji strojnega učenja in poudarja potrebo po nadzoru in razumevanju vhodnih podatkov. Predstavljeni rezultati obravnavajo raziskovanje 16 razsutih zvezdnih kopic in njihov kemični podpis. Pri tem sta glavna problema homogenost in pravilnost izračunanih kemičnih zastopanosti. Pri kemičnem raziskovanju okolice razsutih kopic smo uporabili razumevanje omejitev uporabe metode kemičnih podpisov razsutih zvezdnih kopic, ki naj bi imele homogeno sestavo. Drugi večji sklop disertacije se ukvarja z iskanjem manjših podskupin posebnih tipov spektrov v podatkovni bazi GALAH. Z različnimi pristopi primerjave med opazovanimi in modeliranimi normalnimi spektri smo našli 918 spektrov z izrazitimi molekularnimi pasovi molekule C2, 10.364 spektrov z emisijskimi črtami ter analizirali 329 kandidatov za spektroskopsko nerazpoznavne večkratne sisteme.

• Keywords: astronomija, razsute kopice, zvezde posebnega tipa, dvojne zvezde, klasifikacija, strojno učenje