Machine Learning in Astronomy Physical Property Prediction
Faculty Mentor Information
Dr. Eric Friedlander, College of Idaho; Dr. Katie Devine, College of Idaho; Dr. Grace Wolf-Chase, Planetary Science Institute; and Dr. Charles Kerton, Iowa State University
Presentation Date
7-16-2026
Abstract
Yellowballs (YBs) are young star forming regions primarily consisting of dust and gas. They contain valuable information about the formation of intermediate-sized stars due to their unique structure. Citizen scientists identified approximately 6,000 YBs by their distinct yellow color at the 70 micron wavelength as part of the Milky Way Project, an NSF funded crowdsourcing project with the goal of identifying bubble star formations in the Milky Way. Researchers created a catalog containing flux measurements of four primary wavelengths (8, 12, 24, and 70 microns) and additional information for 4,000 of the YBs, including eight different physical properties such as diameter and luminosity along with longer wavelengths of light (up to 1100 microns). The most prominent means of estimating these physical properties is by first estimating the Spectral Energy Distribution (SED), which is computationally expensive and relatively inefficient. Machine learning can speed up this process by directly predicting the properties, skipping the SED fitting process. Our goal was to explore machine learning as an alternative to SED fitting to predict the properties of the rest of the YBs in the catalog and determine the most important wavelengths and colors used in this prediction. Using models such as histogram gradient-boosted trees, we were able to predict the physical properties with r-squared values ranging from 0.4 to over 0.9, allowing us to predict the physical properties of the remaining YBs left in the catalog. We also identify wavelengths of 160 and 350 microns as containing significantly important information regarding YB properties.
Machine Learning in Astronomy Physical Property Prediction
Yellowballs (YBs) are young star forming regions primarily consisting of dust and gas. They contain valuable information about the formation of intermediate-sized stars due to their unique structure. Citizen scientists identified approximately 6,000 YBs by their distinct yellow color at the 70 micron wavelength as part of the Milky Way Project, an NSF funded crowdsourcing project with the goal of identifying bubble star formations in the Milky Way. Researchers created a catalog containing flux measurements of four primary wavelengths (8, 12, 24, and 70 microns) and additional information for 4,000 of the YBs, including eight different physical properties such as diameter and luminosity along with longer wavelengths of light (up to 1100 microns). The most prominent means of estimating these physical properties is by first estimating the Spectral Energy Distribution (SED), which is computationally expensive and relatively inefficient. Machine learning can speed up this process by directly predicting the properties, skipping the SED fitting process. Our goal was to explore machine learning as an alternative to SED fitting to predict the properties of the rest of the YBs in the catalog and determine the most important wavelengths and colors used in this prediction. Using models such as histogram gradient-boosted trees, we were able to predict the physical properties with r-squared values ranging from 0.4 to over 0.9, allowing us to predict the physical properties of the remaining YBs left in the catalog. We also identify wavelengths of 160 and 350 microns as containing significantly important information regarding YB properties.