GravNet: A novel deep learning model with nonlinear filter for gravitational wave detection

Abstract

Gravitational waves (GW) detected by LIGO, VIRGO, and upcoming facilities have ushered in a transformative era for astronomy and physics. However, these cosmic ripples present unique challenges. Most GW signals are not only weak but also fleeting, lasting mere seconds. This poses a significant hurdle to current search strategies. The prevalent matched filtering technique, while effective, demands an exhaustive search through a template bank, slowing down data processing.To overcome these limitations, machine learning, particularly Convolutional Neural Networks (CNNs), has emerged as a solution. Recent studies demonstrate that CNNs surpass traditional matched-filtering methods in detecting weak GW signals, extending beyond the training set parameters. Nonetheless, optimizing these deep learning models and assessing their robustness in GW signal detection remains essential. In this study, we explore various methods to enhance CNN models effectiveness using simulated data from three gravitational wave interferometers. Our investigation spans denoising techniques, CNN architectures, and pretrained AI models. Notably, the Constant-Q transform (CQT) outperforms the Fast Fourier transform in denoising raw gravitational signals. Furthermore, employing the pretrained model EfficientNet enhances GW detection efficiency. Our proposed CNN model, GravNet, combines CQT, EfficientNet, and an optimized CNN structure. GravNet achieves an impressive 76.5% accuracy and 0.85 AUC. This innovative approach offers valuable insights into harnessing deep learning models for more efficient and accurate gravitational wave detection and analysis.