# WaveFormer: transformer-based denoising method for gravitational-wave data

**Authors:** He Wang, Yue Zhou, Zhoujian Cao, Zong-Kuan Guo, Zhixiang Ren  
**arXiv:** [2212.14283](https://arxiv.org/abs/2212.14283)  
**Date:** 2022-12-29  
**Summary page:** https://www.summarizepaper.com/en/arxiv-id/2212.14283v2/

## Key points

- WaveFormer is a deep learning model designed to denoise gravitational wave observational data from LIGO.
- It combines convolutional neural networks with transformers to capture generic and hierarchical features of GW signals.
- The architecture allows for extraction of low-level, mid-level, and high-level features through convolutional embedding, residual modules, and encoders respectively.
- Extensive testing on LIGO data shows remarkable noise reduction capabilities, reducing overall noise and glitch levels by more than one order of magnitude.
- Signal recovery accuracy improves significantly with error rates for recovering signal phase at approximately 1% and amplitude at 7%.
- WaveFormer achieves a decrease in false alarm rates, enhancing data analysis within the International Gravitational-Wave Observatories Network (IGWN).
- The study demonstrates the effectiveness of large neural networks in processing GW data and highlights WaveFormer's adaptability for broader applications in future observational runs.

## Summary

, , , , 

The study presents WaveFormer, a novel deep learning model designed to denoise gravitational wave (GW) observational data from the Laser Interferometer Gravitational-Wave Observatory (LIGO). By combining convolutional neural networks with transformers, WaveFormer effectively captures both generic and hierarchical features of GW signals. Its architecture allows for the extraction of low-level local features through convolutional embedding, mid-level features via residual modules, and high-level global features by the encoders. Extensive testing on LIGO data containing 75 reported binary black hole events demonstrates remarkable noise reduction capabilities, with overall noise and glitch levels reduced by more than one order of magnitude. This leads to a significant improvement in signal recovery accuracy, with error rates for recovering signal phase and amplitude at approximately 1% and 7%, respectively. Additionally, WaveFormer achieves a substantial decrease in false alarm rates, highlighting its potential for enhancing data analysis within the International Gravitational-Wave Observatories Network (IGWN). The study showcases the effectiveness of large neural networks in processing GW data and highlights WaveFormer's adaptability for broader applications in future observational runs. By leveraging advanced machine learning techniques like WaveFormer, researchers can improve data quality and extract valuable insights from complex GW observational data with unprecedented accuracy and efficiency.