Stellar evolution proceeds on timescales of millions to billions of years, however, most stars in both the birthing and dying processes will punctuate their lives with short-lived (few days to a year) eruptions or explosions that lead to dramatic changes in the brightness of the star. Observational studies of these events are the foundation upon which time-domain astronomy is built. While these events are short lived, they often provide some of our most significant insights into stellar evolution.
This dissertation focuses on the use of time-domain techniques to discover and characterize these rare astrophysical gems, while also addressing some gaps in our understanding of the earliest and latest stages of stellar evolution. The observational studies presented herein can be grouped into three parts: (i) the study of stellar death (supernovae); (ii) the study of stellar birth; and (iii) the use of modern machine-learning algorithms to discover and classify variable sources.
I present observations of supernova (SN) 2006gy, the most luminous SN ever at the time of discovery, and the even-more luminous SN 2008es. Together, these two supernovae (SNe) demonstrate that core-collapse SNe can be significantly more luminous than thermonuclear type Ia SNe, and that there are multiple channels for producing these brilliant core-collapse explosions. For SN 2006gy I show that the progenitor star experienced violent, eruptive mass loss on multiple occasions during the centuries prior to explosion, a scenario that was completely unexpected within the cannon of massive-star evolution theory. I also present observations of SN 2008iy, one of the most unusual SNe ever discovered. Typical SNe take ≤3 weeks to reach peak luminosity; SN 2008iy exhibited a slow and steady rise for ~400 days before reaching maximum brightness. The best explanation for such behavior is that the progenitor of SN 2008iy experienced an episodic phase of mass loss ~100 yr prior to explosion. The three SNe detailed in this dissertation have altered our understanding of massive-star mass loss, namely, these SNe provide distinct evidence that post-main sequence mass loss, for at least some massive stars, occurs in sporatic fits, rather than being steady. They also demonstrate that core collapse is not restricted to the red supergiant and Wolf-Rayet stages of stellar evolution as theory predicted. Instead, some massive stars explode while in a luminous blue variable-like state.
I also present observations of the newly discovered FU Orionis variable, PTF~10qpf. FU Orionis stars are young stellar objects (YSOs) that exhibit long-lasting (≥10 yr), large-amplitude (≥5 mag) eruptions due to accretion instabilities in the star+disk system. These eruptions, for which there are precious few examples, play an important role in (i) determining the final mass and angular momentum of the newly born star, (ii) clearing the circumstellar envelope from which the star forms, and (iii) the formation of planets, and ultimately life, in the circumstellar disk. PTF~10qpf is only the fourth FU Orionis variable with detailed observations taken during the course of eruption. Furthermore, it is the best observed FU Orionis star prior to eruption, with both optical spectra and infrared photometry demonstrating that the star was a normal classical T Tauri star before its outburst. This discovery shows that the FU Orionis phenomenon is not reserved for only the most-massive YSOs, as had previously been suggested.
As robotic observing and automated data processing procedures render the standard transient discovery process mundane, we are quickly approaching an era where we will be overwhelmed with discoveries. The Large Synoptic Survey Telescope (LSST), with a planned start in 2020, is expected to discover ~50 million variable stars, an orders-of-magnitude leap over the currently known number of variables. As data volumes grow to these enormous sizes, it is clear that classical discovery and characterization techniques relying on the visual inspection of data are no longer tractable. In the final chapters of this dissertation, I present a scaleable machine-learning framework capable of identifying rare sources in a time-domain dataset, while also providing classifications for the most prevalent source types in the survey. Following a search of the variables identified in the All Sky Automated Survey, I discover four bright R Coronae Borealis stars, carbon-rich supergiants in a short-lived phase of late stellar evolution. The discovery of these stars, which were identified via a fully automated procedure, represent an important proof-of-concept demonstrating that advanced algorithmic procedures can unearth the rare astronomical sources that lead to leaps in our understanding of stellar evolution. I close with the presentation of a new machine-learning methodology for inferring the fundamental atmospheric properties of stars, Teff, log g, and [Fe/H], without obtaining spectra. While the development of these tools, which predict the atmospheric parameters from photometric light curves, are still in their infancy, I argue that their continued development will enable the conversion of large photometric time-domain surveys, such as LSST, into pseudo-spectrographs.
Astrophysical science should no longer be viewed as a static tableau but instead as the unraveling of violent beginnings in a dynamic cosmos. The use of machine learning to extract novel results from large astronomical datasets occupies an instrumental portion of a burgeoning 21st century revolution in the way we conduct ourselves as scientists — the poetic connection of this modern approach applied to the millennia-old practice of monitoring the heavens will uncover a universe of new mysteries.