Direct Collapse Black Holes
High-Resolution Halo Resimulations
My current research focuses on the formation of direct collapse black holes (DCBHs), a particular pathway to form massive black holes in the early universe. Specifically, I'm performing high-resolution resimulations of halos expected to form DCBHs to inspect the small-scale features of their internal structures, allowing us to determine if these halos will form DCBHs or some other structure such as a dense star cluster.
DCBH Classification
This project was an application of a machine learning architecture called a support vector machine (SVM) to cosmological simulations. I developed an SVM that predicts the probability of a halo forming a DCBH given the halo's properties (mass, temperature, metallicity, stellar mass, etc.). I improved the SVM by optimizing its hyperparameters and measuring the relative importances of different features in our data. I also investigated the performance of SVMs with different feature subsets and inspected why some halos were misclassified by the SVMs. This work has been published in ApJ and is also available on arXiv.
Neutrinos
WIMP Annihilation
This project is a statistical search for neutrinos from dark matter (DM), where we assume DM consists of Weakly-Interacting Massive Particles (WIMPs). Given the location of observed neutrinos, the predicted spectra of neutrinos from WIMP annihilation into a particular particle-antiparticle pair, and the properties of the dwarf galaxies we assume are sources of WIMP annihilation, we calculate the best-fit number of events we expect to see from WIMP annihilation and compare this to a background-only hypothesis. If we do not find a statistically-significant deviation from the background-only hypothesis, we instead set upper limits on the velocity-averaged WIMP annihilation cross section $\left<\sigma v\right>$. The publication covering this work is under review at PRD (see arXiv preprint).
RNN Reconstructions for IceCube-Upgrade
This project was attempting to improve upon existing reconstruction methods for a future detector geometry. By treating each event as a sequence of hits, we used recurrent neural networks (RNNs) RNNs to attempt to reconstruct the energy and direction of incoming neutrinos from simulated neutrino data. This work showed promise for improving both energy and direction reconstructions at low energies $\left(\lesssim 10 \, \mathrm{GeV}\right)$; the data required additional processing to remove events that had large estimated uncertainities to truly see improvements in performance, which was done by another undergraduate student after I moved to the WIMP Annihilation project.
CNN Reconstructions for FLERCNN
This project was attempting to speed up existing event reconstruction methods by using neural networks, which can be several orders of magnitude faster than likelihood-based reconstructions. By summarizing the event information for each Digitial Optical Module (DOM) in the detector, we treated the data of an event like an image and used convolutional neural networks (CNNs) to attempt to reconstruct the neutrino's energy and zenith angle. My work on this project focused on testing different network archtiectures and loss functions. This led to the development of the Fast Low-Energy Reconstruction using Convolutional Neural Networks (FLERCNN) model (see PRL publication).