Large astronomical surveys obtain high-resolution spectra of millions of stars that can be used to understand the formation and evolution of our own Milky Way Galaxy. But measuring the abundance of different chemical elements from these spectra in an automated manner is challenging. In this project, the SUDS scholar will adapt our successful astroNN methodology for determining elemental abundances using a deep-learning technique to spectra from the new SDSS-V survey and use them to make a chemical map of our Milky Way. Specifically, the SUDS scholar will adapt the existing astroNN implementation (in tensorflow and keras) so it can be applied to spectra from SDSS-V, run tests of the adapted implementation and check the accuracy of results, work with other people in the group to incorporate the adapted technique into the SDSS-V pipeline, and explore the chemistry of the Milky Way with the resulting abundances.

Because we can only observed the Milky Way at the present time, obtaining ages for large numbers of stars is crucial to unraveling our Galaxy’s history. However, ages are notoriously difficult to obtain using traditional astronomical techniques. The most robust method for determining ages uses time series of red giant stars; a star’s age is directly reflected in the random oscillations that such stars undergo and that we can observe using detailed time series observations. However, these observations are expensive and difficult to model.

Obtaining high-resolution spectra using a diffraction grating is much easier and such samples now consist of about a million stars. But while we believe these spectra contain age information, we have no robust theory to extract it. This is where machine learning comes in! In this project, we will use contrastive learning to extract the age information from stellar spectra using similar techniques as used to, for example, provide captions for images (see, e.g., OpenAI’s clip). We will use this to obtain ages for large numbers of stars in the APOGEE and SDSS-V surveys and determine the age distribution of stars across the Milky Way’s disk. The student will be responsible for implementing the contrastive learning process in Pytorch using data that we will provide and for evaluating the model’s performance using a test set and by comparing to the results from other, previous techniques.