Massive telescopes, both here on Earth and orbiting in space, have been a boon to humanity’s understanding of the cosmos. They reveal the nature of countless galaxies in the night sky, various types of stars, thousands of exoplanets and clues to unanswered questions about the Big Bang.
Older methods to decipher information coming from these instruments have often sufficed. But in recent years, astronomy has seen a strong growth in more complex and sophisticated analysis tools. These new techniques, advanced machine learning and modern statistics, are giving scientists a more accurate image of stars and galaxies, supplanting older, less precise methodologies.
Peter Melchior, an assistant professor jointly appointed to Princeton University’s Department of Astrophysical Sciences and the Center for Statistics and Machine Learning (CSML), has made it his research focus to develop these sophisticated machine learning algorithms for astronomical research.
Melchior is also the newest faculty member at CSML, having joined this past fall. He leads the Astronomy Data Group, his research lab, where they are developing techniques for night sky surveys that are poised to launch in the next several years.
“We are pleased to have Peter Melchior be part of our center,” said Peter Ramadge, the CSML director. “He is not only doing important research in astrophysics, but his work shows that advanced machine learning methods are vital to pushing forward growth in the sciences. His work also fits with the collaborative and interdisciplinary nature of the center.”
A Background in Physics
Melchior earned a bachelor’s and master’s degree in physics in 2006 from the University of Heidelberg in Heidelberg, Baden-Württemberg, Germany. He later earned his doctoral degree in physics in 2010 from the same university. Before assuming his professorship, he was a professional specialist on campus and served as a postdoctoral fellow at Ohio State University and as a research postdoc at Heidelberg. He also had a stint as a software developer at a startup company, Certon Systems, in Heidelberg. To date, Melchior has authored 183 research papers and is a builder of the Dark Energy Survey, a collaborative and international project dedicated to studying dark energy in the universe.
An Interdisciplinary Approach to Teaching
In spring 2020, Melchior is slated to teach SML 515 - Statistical Data Analysis, which is open to students across the sciences. This graduate-level class provides an introduction to modern data analysis and data science for students who have studied linear algebra, multivariate analysis, basic statistics and programming in Python. Melchior plans on taking an interdisciplinary direction with the course, which entails using datasets and problems from various science disciplines and having faculty members from different departments present their research to students. More information on the course can be found here.
This past fall at CSML, Melchior also led a series of informal, interdisciplinary seminars called Machine Learning for the Sciences, which will continue in the spring as well. Catering mostly to graduate students and postdocs, Melchior organizes the presentation of machine learning techniques for science and engineering research, along with tutorials and discussions on projects and research papers. Melchior said the seminars have attracted many people from different research areas such as astrophysics, physics, molecular biology, neuroscience, mechanical and aerospace engineering and others. More information on these seminars can be found here.
Using Machine Learning to Understand the Night Sky
Melchior said he has found the environment at CSML ideal for his research.
“It’s a great place because it allows researchers to connect with different faculty and groups on campus who are working on or who are interested in machine learning and advanced statistics. That linkage was missing in Princeton before the establishment of CSML,” said Melchior. “Being able to connect to other researchers and leverage their knowledge and find out if we can use these techniques in astronomy to address problems is a benefit.”
In his own research, specifically, Melchior is developing better machine learning methods to extract data from three upcoming space surveys: The Large Synoptic Survey Telescope (LSST) in Chile, and the Wide Field Infrared Survey Telescope (WFIRST) and Euclid, which will both be launched into space. The problem with these surveys is that the information that will be sent back to Earth will be complex and difficult to understand, Melchior said. These telescopic instruments are so sensitive that the data churned out will show galaxies overlapping, making data hard to measure.
To make sense of the data coming from these surveys, Melchior said researchers can either develop sophisticated algorithms to parse the information or merge together data from different telescopes. Melchior is focused on the latter, which involves, as an example, combining optical and infrared data in order to get a more detailed and accurate picture of the night sky. On top of solving measurement problems, joint data analysis also allows to discover astrophysical signals that were hard or impossible to find in a single data set.
“Astronomy, more than anything, is a great test bed for both modern machine learning and statistics,” he said. “Most of our data is free and there is lots of it, so we can ask many questions. There is a great interest in the community to extract as much information as possible from the data.”