Research
Fast Times at UC Berkeley
Classifying Fast Radio Bursts and Radio Frequency Interference with Convolutional Neural Networks
During my time at Berkeley, I had the pleasure of doing research in the Department of Astronomy under Dr. Vishal Gajjar, a postdoc with the Breakthrough Listen team. Breakthrough Listen's main objective is to search for signals of intelligent life—not too shabby of a project to partake in as an undergrad.
Fast Radio Bursts
Fast radio bursts (FRBs) are one of the most mysterious objects in astrophysics. Powerful radio pulses lasting a few milliseconds at max, no one knows their origins, and their sources can't be tracked and studied further because they usually come once and never again from the same place. Current theories for progenitors include neutron star collisions, ridiculously energetic supernovae (very cool), and the collapse of the magnetospheres of supermassive black holes.
Some people think it could be aliens. If I were to put money on it, it's probably not.
[EDIT 1/23/2020] Recently, astronomers have been building observational evidence converging on magnetars as the source of these fast radio bursts. Magnetars are stars that have completed the main portion of their life fusing hydrogen into helium and have collapsed into neutron stars with extreme magnetic field environments. Although we are closing in on a source, it's still a mystery as to how these terrific bursts of electromagnetic radiation are generated.
To study these strange phenomena, Vishal designed a pipeline to search for FRB pulses from from the Green Bank and Parkes, two of the largest radio telescopes in the world. Unfortunately, a major problem for current detection software is radio frequency interference (RFI). Emissions like WiFi and microwave ovens would very likely be mistaken for an FRB, so the pipeline accepts many false candidates.
Differentiating FRBs from RFI
My project focused on developing a binary classification model that would be able to distinguish RFI from FRBs and whittle down the number of candidates that a human had to check. I tackled this problem with convolutional neural networks, inspired by the approach in Connor & van Leeuwen (2018).
Training data for FRBs is sparse, with currently only a few hundred real examples at most. Because neural networks require huge amounts of data to be effective, we decided to simulate them in the same way that Connor & van Leeuwen did. Here are a few of the results:
After training and debugging, testing the model's performance on known FRBs delivered good results, and the model was able to pick out most pulses. Below are samples of predicted FRBs that the model was highly confident in (> 95% probability).
Because we cared more about finding every single FRB, I weighted the network such that it penalized false negatives, leading to a higher recall but greater chance of false positives. You can see a false positive in the second prediction, which has no signal in it, even though the model thought it did.
I created multiple models, each trained on different simulated parameters depending on the telescope and what characteristic data from that telescope would look like. As of now, my models are integrated in Vishal's pipeline, happily sifting out that pesky RFI.Other Research
Stellar shock waves
My previous research involved working with Dr. Stephen Ro to study how shock waves propagated through stars in order to determine the origins of false supernovae. The premier code to simulate stellar conditions, MESA, had recently included hydrodynamics in its physics repertoire. By simulating explosions in stars—"setting off bombs," as Stephen would call it—and watching how the shocks traveled, I verified that MESA's new hydrodynamics calculations complied with theoretical predictions.
Stellar nucleosynthesis in tidal disruption events
My first research experience at Berkeley also dealt with MESA and was also very cool. Dr. Ken Shen and I were exploring stellar nucleosynthesis, the energy-production process by which stars fuse lighter elements to make heavier ones. Specifically, we were looking at what would happen to nucleosynthesis if we threw a star deep towards a black hole, causing it to be squeezed a ton and wildly increasing its density and temperature. This is known as a tidal disruption event, when the star is ripped apart and spaghettified by the tidal forces close to a black hole.
After getting rough numbers for the pressure, temperature, and density a star would encounter upon tidal disruption, I conducted simulations in MESA to investigate the fusion reactions taking place within the star.