EAS Doctoral Proposal Defense by Anushka Doke
TXT 105
Dr. Sarah Caudill
scaudill@umassd.edu
Abstract:
Gravitational-wave detectors, such as the Laser Interferometer Gravitational-Wave Observatory (LIGO), are highly sensitive instruments designed to detect weak astrophysical signals, requiring sophisticated data analysis techniques to distinguish them from noise. This proposal aims to develop machine learning and data-driven methods to address challenges in gravitational-wave astronomy, with applications to signal detection, detector characterization, and low-latency analyses.
We first address the detection of precessing binary black hole (BBH) systems. Because a fully precessing search is computationally prohibitive, current searches filter with aligned-spin template banks and are not optimally sensitive to precessing BBHs. We have performed aligned-spin searches using the SGNL pipeline on data containing simulated aligned-spin and precessing BBH signals, producing matched-filter triggers for both classes. We propose to train a feed-forward neural network on the recovered trigger parameters, such as the signal-to-noise ratio, consistency between the data and the best-fitting template, and recovered masses and spins, to identify triggers likely to arise from precessing signals. This classifier will serve as a preprocessing step to rapidly identify the small subset of triggers that require further follow-up. For this subset, we propose an optimization method that combines the signal-to-noise ratio with the autocorrelation-based least-squares signal consistency test to improve the ranking of precessing candidates.
We next address the characterization of transient bursts of detector noise, or “glitches.” The LIGO detectors record data from more than 200,000 auxiliary channels that monitor the instrument and its environment. We have developed a GPU-accelerated framework to transform time-series data from these auxiliary channels into the fractal dimension, a compact representation of signal complexity. These representations will be used to train an unsupervised autoencoder to learn nominal detector behavior and flag anomalous activity. We will use the detected anomalies to identify auxiliary channels that witness distinct glitch populations, pointing to their environmental or instrumental origin.
Finally, we address the need for low-latency gravitational-wave alerts for multimessenger astronomy. Public alerts are currently issued approximately 30 seconds after a candidate is identified using GWCelery, a Python framework built on the Celery distributed task queue, in which each stage records its result in the GraceDB event database before the next proceeds. As part of the LIGO Scientific Collaboration, we are developing SGN-LLAI, a streaming infrastructure built on the Stream Graph Navigator (SGN) library, in which results pass directly between stages via Kafka topics and are asynchronously recorded in GraceDB, removing database operations from the critical path. SGN-LLAI currently produces preliminary alerts in the standard public format, with latencies as low as 5 seconds in initial tests. The proposed next phase will complete feature parity with GWCelery by adding source classification and properties, data quality information, and the full progression of public alerts from preliminary to initial and update notices, including human vetting and retraction. Alerts will then be disseminated through the GCN and SCiMMA networks.
Advisor(s): Dr. Sarah Caudill, Department of Physics (scaudill@umassd.edu)
Committee members:
- Dr. Scott Field, Department of Mathematics
- Dr. Vijay Varma, Department of Mathematics
- Dr. Derek Davis, Department of Physics – University of Rhode Island
Note: All EAS Students are encouraged to attend.