Research

Dark matter searches, computer vision, and detector performance testing

I've spent my time in grad school working on the South Pole Telescope (SPT) experiment. Located at the actual, geographic South Pole, SPT studies the oldest light in the universe: the cosmic microwave background, or CMB. The CMB is relic radiation from the Big Bang. By studying it, we can learn about the universe's origin and history; and we can tie our observations on these cosmic scales to particle physics happening on the smallest scales.

Image: Me in front of the SPT primary mirror during my first deployment to the South Pole in early 2020.

Axion search

Axions are a theoretical particle species and a possible dark matter candidate, particularly interesting due to their cosmology-free motivation. Axions were originally proposed to solve a particle physics problem, the Strong CP Problem, related to certain symmetries that are not broken in the "strong" sector of particle physics. It was only later realized that they are a strong candidate to answer one of the hottest questions in modern physics: "What is dark matter?"

Axions with ultralight masses (roughly 10^-23 - 10^-18 eV), were they to exist, would cause the polarization of the CMB to appear to oscillate back and forth on human-observable timescales from hours to years. I co-led a search for this effect in data from SPT's 2019 observing season. While we did not discover evidence of axions, our results allowed us to place strong constraints on the strength of the coupling between axions and photons; see the figure above, the main result from our published paper.

Detector computer vision

The camera currently installed on SPT comprises over 16,000 cryogenic detectors. The next SPT camera (currently under development) will more than double this number, and the upcoming CMB-S4 experiment will utilize over 500,000 detectors. These increasingly-large arrays are necessary to reach the ultra-low noise levels required by current and future CMB science goals. However, detector testing and quality assurance is a time-consuming process, raising an interesting challenge: "How do you test the performance of this many detectors in an effective manner?"

I spent the second half of my grad school tenure trying to answer this question, using a method whereby detectors are imaged at high resolution using an optical microscope; the visual features are then used to predict the cryogenic performance. Using computer vision techniques, I developed an effective method of finding fabrication defects on individual detectors. I also used a random forest machine learning algorithm to correlate certain visual features with detector performance properties; you can read about it here!

You can also find a public version of the code used in that paper (as well as an upcoming one on a related project using new detectors) here.