Convergence of numerical relativity, deep learning at scale, and large scale computing for Multi-Messenger Astrophysics

Eliu Huerta Escudero, University of Illinois at Urbana-Champaign

Usage Details

Roland Haas, Philipp Moesta, Cunwei Fan, Eliu Huerta Escudero, Sibo Wang, Aaron Saxton, Shawn Rosofsky, Anushri Gupta, Arnav M Das, Sarah Habib, Bing-Jyun Tsao, Asad Khan, Debopam Sanyal, Rohan Baskar Prabhakar, Navid Tajkhorshid, Brockton Brendal, Joseph D Adamo, Bridgette Davey, Xiaolan Ke, Kaiwen Zhang

Our research products from previous Blue Waters allocations include the construction of the most comprehensive catalog of numerical relativity waveforms that describe non-spinning black hole binaries on eccentric orbits. That catalog enabled us to construct the state-of-the-art waveform model for eccentric binary coalescence; quantify the importance of including higher-order waveform multipoles for the detection of eccentric binary mergers; and the development of an algorithm to translate numerical relativity eccentricity into a quantifiable parameter in the context of post-Newtonian theory.

Numerical relativity catalogs provide the key materials to continue advancing our research program at the interface of numerical relativity and deep learning. Over the last year, we demonstrated that deep learning outperforms conventional machine learning techniques, achieves similar performance compared to matched filtering, while being several orders of magnitude faster, allowing real-time signal processing with minimal resources. Furthermore, we established that deep learning can detect and characterize waveform signals emitted from new classes of eccentric or spin-precessing binary black holes, even when trained with data sets of only non-spinning, quasi-circular binary black hole waveforms. The use of eccentric waveforms produced with our Blue Waters allocations were critical to realize these studies.

Through this allocation we will design and validate new neural network models to detect and characterize spin-aligned, quasi-circular binary black hole mergers, i.e., we will construct neural network models to detect and estimate in real-time the parameters of these waveforms signals in real LIGO noise. While we will use a large catalog of simulated waveforms to train these models, we will use our numerical relativity catalogs to test and validate the robustness of deep learning to identify new classes of gravitational wave signals, in particular those that describe eccentric binary black hole mergers.