Enhancing Core-Collapse Supernova Gravitational Wave Searches with Machine Learning

Andy H.Y. Chen^1*, Chia-Jui Chou², Kuo-Chuan Pan³, Yi Yang², Shih-Chieh Hsu⁴, Albert K.H. Kong³

¹Institute of Physics, National Yang-Ming Chiao Tung University, Hsinchu, Taiwan
²School of Physical Science and Technology, ShanghaiTech University, Shanghai, China
³Institute of Astronomy, National Tsing Hua University, Hsinchu, Taiwan
⁴Department of Physics, University of Washington, Hsinchu, USA

* Presenter:Andy H.Y. Chen, email:andy.c.80297@gmail.com

Detecting gravitational waves (GWs) from core-collapse supernovae (CCSNe) remains a major challenge due to their weak strain amplitudes and poorly modeled waveforms. Conventional searches typically employ coherence-based, signal-agnostic methods that depend on multiple detectors and yield limited insight into the underlying source dynamics. To address these limitations, we present CCSNet, a machine learning–based binary classifier designed to distinguish CCSN signals from transient noise artifacts. By selectively identifying CCSN-like features while suppressing nonastrophysical glitches, CCSNet enhances candidate event significance and improves the interpretability of potential detections. The framework supports both dual- and single-detector inputs, enabling CCSN searches over a broader operational duty cycle. We evaluate CCSNet using 31 CCSN wave form templates and four weeks of background data across multiple training configurations to assess robustness and generalization. In dual-detector mode, CCSNet successfully recovers 50% of injected CCSN signals (at a 5% false positive rate) when the signal-to-noise ratio (SNR) exceeds 12.57 in O3b data, while maintaining meaningful sensitivity up to an SNR of 15.83 in single-detector operation. These results demonstrate the potential of data-driven approaches to complement traditional gravitational-wave pipelines and enhance future CCSN detection prospects.

Keywords: Gravitational Wave, Machine Learning, Multi-messenger Astronomy