Stephen Portillo’s research focuses on using advances in statistics and machine learning to allow more science to be done with existing astronomical data sets. On the statistics front, he has been developing probabilistic cataloging, a Bayesian Markov chain Monte Carlo method that improves source detection and measurement in crowded images. On the machine learning front, he has been applying autoencoders, a type of deep neural network, to enable astronomers to more easily find patterns and outliers in large datasets.
Dr. Portillo is a DiRAC Postdoctoral Fellow and UW Data Science Postdoctoral Fellow. He joined the DiRAC Institute in September 2018 after finishing his PhD in Astronomy and Astrophysics at Harvard University. Before graduate studies, he completed a BSc in Astrophysics at the University of Alberta in Edmonton, Canada in 2012.
Crowded images are difficult to analyze because sources can appear blended with their close neighbors. This problem will worsen with more sensitive observatories like Rubin Observatory, James Webb Space Telescope, and Roman Space Telescope that will see unprecedented numbers of objects in the same area of sky. Unlike traditional methods that first identify sources before measuring them, probabilistic cataloging treats source identification probabilistically. Dr. Portillo has shown that this method can find stars four times fainter than state-of-the-art methods in extremely crowded images with 1 star per 10 pixels.
At DiRAC, Dr. Portillo has joined the KBMOD team, who are developing GPU-accelerated software to find Kuiper belt objects. Recently, he has been developing a method to correct for the Earth’s motion in the solar system, allowing KBMOD to better track objects over longer periods of time. He is also extending probabilistic cataloging to search for binaries among the objects found by KBMOD, because these binaries are powerful probes of the dynamical history of the outer Solar System.
Working with Prof. Connolly, Dr. Portillo has implemented a variational autoencoder to galaxy spectra from the Sloan Digital Sky Survey. He showed that the autoencoder can summarize spectra with thousands of pixels with only six numbers and easily separates known classes of galaxies. Currently, he is working with students to use this autoencoder to find massive black hole binaries, rare objects identified by unusual spectra.
Dr. Portillo is also passionate about public outreach and teaching. While at DiRAC, he has given an Astronomy at Home talk and given virtual presentations to school groups. He was also a lecturer at AstroHackWeek 2020 and is currently co-instructing an undergraduate course on astrostatistics with Prof. Juric.
Dr. Portillo is excited to be at the DiRAC Institute because it brings together researchers interested in all aspects of data-intensive science from software engineering to statistics and machine learning. He is also happy to be a part of the eScience Institute that encourages researchers across scientific fields to find commonalities in the ways they analyze data.