2019 Abstract from Hailey

Seeing the Sky Through a New Lens:

Using semi-supervised machine learning to find strong lenses

Student Researcher:  Hailey Hurd, (The Latin School, Chicago)

Scientist Mentor:  Brian Nord

Strong gravitational lenses occur when two astronomical objects line up with Earth such that the light from the furthest object appears to be distorted by the gravity from the middle object. Cosmologists study strong gravitational lenses because they offer important clues on the rate of expansion of the universe (The Hubble Constant), among other unknowns in astronomy and cosmology. However, the current number of known gravitational lenses is in the order of thousands, and the current methods of identifying more are slow and inefficient. Especially with all the data that will be arriving in the coming decades from newer, larger telescopes, it is necessary to develop a quicker, more accurate, and less labor-intensive way to locate new gravitational lenses.

This summer, our team (Brian Nord, Marwah Roussi, Qian Gong, and Hailey Hurd) explored semi-supervised machine learning as an efficient approach to searching for strong gravitational lenses. Semi-supervised learning is able to take advantage of both labeled and unlabeled image data, and our aim is to find out whether semi-supervised learning is more accurate than other machine learning methods for this application, in which labeled data is limited. After studying scientific papers, articles, and open-source code on Siamese networks and triplet networks, we were able to begin developing our semi-supervised network on simulated strong gravitational lens data, which can process multiple telescope filters corresponding to the same image. We also implemented some other machine learning methods, such as a simple convolutional neural network, as examples to which we can compare our semi-supervised machine learning model. Next steps will be to improve and optimize our model, and to continue exploring comparisons and metrics with which to assess the success of our model.