Uploaded July 2022 | Updated September 2026, 8 minutes ago
During their first three observing runs, the LIGO, Virgo, and KAGRA collaborations have detected 90 gravitational wave signals. As detectors improve, particularly with the introduction of LISA in the coming decades, this number will grow to thousands of signals. Detecting and understanding these signals will require dense banks of simulated waveforms that show the expected gravitational radiation emitted from binary systems. For comparably massed systems near merger, these waveforms come from numerical relativity simulations which solve Einstein's equation numerically. However, these simulations are computationally expensive and time consuming, so we must choose their initial parameters carefully to ensure we gain as much benefit from each new simulation as possible. To optimize this, we train a neural network to predict how similar two waveforms will be (using the match) without performing the simulations. Using this predicted match as a measure of distance, we can place new simulations such that they are as different from existing simulations as possible. This will help us optimize the use of our time and computational resources as we prepare a bank of waveforms for use with LISA.
Author and Presenter: Deborah Ferguson
During their first three observing runs, the LIGO, Virgo, and KAGRA collaborations have detected 90 gravitational wave signals. As detectors improve, particularly with the introduction of LISA in the coming decades, this number will grow to thousands of signals. Detecting and understanding these signals will require dense banks of simulated waveforms that show the expected gravitational radiation emitted from binary systems. For comparably massed systems near merger, these waveforms come from numerical relativity simulations which solve Einstein's equation numerically. However, these simulations are computationally expensive and time consuming, so we must choose their initial parameters carefully to ensure we gain as much benefit from each new simulation as possible. To optimize this, we train a neural network to predict how similar two waveforms will be (using the match) without performing the simulations. Using this predicted match as a measure of distance, we can place new simulations such that they are as different from existing simulations as possible. This will help us optimize the use of our time and computational resources as we prepare a bank of waveforms for use with LISA.
Author and Presenter: Deborah Ferguson
