Supervised classification

Machine learning classification has been shown to be a powerful tool for the selection of rare astrophysical sources. I have developed a machine learning classification algorithm based on random forests to identify the most luminous quasars at z=3-5. Random forests is a state-of-the-art ensemble learning method that combines multiple decision trees to improve classification accuracy and reduce overfitting. The algorithm was trained on a large sample of known quasars and stars, and applied to panchromatic photometric data. Lukas Wenzl, a Bachelor student under my supervision, extended this work towards higher redshifts.

For our Euclid z>6.5 quasar selection, my postdoc Francesco Guarneri is using optimized gradient-boosted decision trees implemented in eXtreme Gradient Boosting (XGBoost). The algorithm is trained on synthetic quasar photometry (see QUEST below) and empirical contaminant photometry. Our current discoveries shows that this selection strategy is highly effective.

Relevant publications

• The Extremely Luminous Quasar Survey in the SDSS Footprint. I. Infrared-based Candidate Selection (Schindler et al. 2017)
• Random Forests as a Viable Method to Select and Discover High-redshift Quasars (Wenzl, Schindler et al. 2021)

Generative models

At high redshift, z>6, there are not sufficient quasars known to train a supervised machine learning classifier. As quasar spectra do not seem to evolve with redshifts, with the exception of the absorption due to the intergalactic medium (IGM), we can train a generative model on lower redshift quasars and use it to generate synthetic quasar spectra.

We have trained an information maximizing Variational Autoencoder (VAE) on a large sample of SDSS quasars and use is to generate synthetic quasar spectra. This generative model, Quasar unsupervised encoder and synthesis tool (QUEST), has been published and the architecture and the trained models are publicly available. We calculate synthethic photometry from the generated spectra, which have been redshifted and modified to account for IGM absorption.

Relevant publications

• QUEST: Quasar unsupervised encoder and synthesis tool: A machine-learning framework for generating quasar spectra (Guarneri et al. 2026)

Un-/Self-supervised classfication, representation learning, and anomaly detection

My new PhD student Tatsuyuki Sekine and my postdoc Laura Martínez are now exploring un-/self-supervised discovery strategies for rare sources in the Euclid Wide Survey. We are interested in sources that would be rejected by supervised classification, but are still astrophysically interesting. These include lensed quasars, quasars with unusual spectral energy distributions, and other exotic sources.

In particular, we are developing strategies based on probabilistic auto-encoders and contrastive learning (simCLR, MoCo).