Recorded Talks

tedxsut2025

Magic and astronomy do not have a lot in common, or do they? Lukas Steinwender explains that these two very different things share the one thing in common, and may provide a new way to teach science. Driven by child-like wonder and awe, Lukas aspires to find mysteries within our universe and inspire others to do the same. This desire is rooted in his passion for sleight of hand magic, the art of sharing little wonders with everyone. Thinking like a magician sparked a wide variety of interests, bridging fields from 3d-art via artificial intelligence to astrophysics – always with the motive of learning more about the wonders of our universe. This talk was given at a TEDx event using the TED conference format but independently organized by a local community. Learn more at https://www.ted.com/tedx (taken from youtube video description). Reference: TEDxSUT2025_AstroThroughMagic

Selection of Slides

ozfink2025

The Vera C. Rubin LSST (Rubin) will revolutionize modern optical astronomy in terms of the sheer data-volume and data production rate. Within this rich survey will be millions of supernovae, which makes spectroscopic classification of all of them infeasible. An accurate classification of core-collapse supernovae (CC SNe) is of particular importance, as their numbers are low especially at high redshifts. In turn, rates and properties of CC SNe at these distances are unknown, and they are significant contaminants in type Ia SNe cosmology.
Community brokers such as Fink are an integral part in accessing the Rubin alert stream. As such, they rely heavily on classification algorithms. We present first results of a data-driven machine learning approach for the photometric classification of CC SNe into SN Ib/c and SN II, two of their subclasses. Our study uses LSST-like precursor datasets and simulations to uncover advantages and pitfalls of existing architectures and proposes a new architecture based on transformers. Our photometric classifier will enable us to increase the number of known CC SNe, probing up to high redshifts to get rates and properties, and remove CC SNe contaminants at scale.

britestars2024

Big Datasets are becoming increasingly important in the field of astrophysics due to new powerful telescopes and surveys continuously observing our night sky, producing giant datasets. A prominent example of an observatory that will produce a large data stream is Vera C. Rubin LSST, designed to simultaneously monitor the southern night sky in 6 colors down to a limiting magnitude of r~24.5 (in a single image). Unsupervised machine learning (ML), a method of learning from patterns within the data, is key for gaining insight into this expansive photometric data.
We present our prototype classifier dedicated to classifying RR Lyrae variable stars into their relevant subclasses (RRab, RRc, and RRd), optimized for the expected needs of Rubin LSST. Using the classification from the RR Lyrae Catalog of the ESA Gaia mission, we constructed training sets containing ~30000 lightcurves using data from the Zwicky Transient Facility Catalog of Periodic Variable Stars (ZTF CPVS) and lightcurves extracted from Full-Frame-Images of the NASA TESS space telescope. We further preprocessed and analyzed these lightcurves using a custom-built Python package. Subsequently, using a Variational Autoencoder (VAE), we created a deep generative model to project the lightcurves into a low-dimensional representation, which quantitatively describes the characteristic shape elements of the lightcurves. In preparation for the Rubin LSST data, we applied our pipeline to the Zwicky Transient Facility (ZTF) data, a precursor facility for Rubin LSST. We show how the ML results are improved by including physical features such as period and variational amplitude. Through unsupervised clustering of this data representation, we identified the RR Lyrae subclasses in the ZTF data, which we anticipate to be easily adaptable to the Rubin LSST data, once available. Reference: Steinwender2024_UnsupRRLyr_Talk_BRITE

aas2023

Big Datasets are becoming more and more prominent in the field of astrophysics with new powerful telescopes and surveys capturing parts of our night sky in an almost continuous fashion. Consequently, unsupervised Machine Learning is key to getting first insight into this abundance of data. We present our results on the performance of unsupervised Machine Learning on the task of classifying RR Lyrae stars into their respective subclasses (RRab, RRc, and RRd) using their lightcurve morphology. We crossmatched the Gaia DR3 RR Lyrae Catalog with the TESS Input Catalog and subsequently extracted a training dataset of over 30000 lightcurves from TESS full-frame images. We further preprocessed and analyzed these lightcurves using a custom-built python-package.
We applied a β-Variational Autoencoder (β-VAE) followed by an unsupervised clustering technique to the extracted lightcurves to infer the subclasses based on structure within the data. We additionally discuss the effect of enhancing the light curves with information from the Gaia DR3 Catalog such as effective temperature and metallicity before clustering. We present a β-VAE that is able to depict the distribution of lightcurves in a low-dimensional latent space and artificially generate new lightcurves from said latent space. Furthermore, we present our initial results for the unsupervised classification including a comparison with the Gaia DR3 RR Rylae Catalog. We further discuss our intentions to apply the method to ZTF lightcurves in preparation for the Vera C. Rubin LSST. Our pipeline will be an efficient way of obtaining initial insight into the Vera C. Rubin LSST data. Reference: Steinwender2023_UnsupRRLyr_Talk

Poster Gallery