# Classification of astrophysical events from gravitational wave signature

**Authors:** Shashwat Singh, Amitesh Singh, Ankul Prajapati, Kamlesh N Pathak  
**arXiv:** [2008.06550](https://arxiv.org/abs/2008.06550)  
**Date:** 2020-08-14  
**Summary page:** https://www.summarizepaper.com/en/paper/2008.06550v1/

## Key points

- Advancements in Deep Learning techniques have led to a rise in classification algorithm development for Gravitational Wave astronomy.
- Current algorithms lack success with significant time duration and further prediction of merger events parameters.
- Researchers aimed to advance deep learning techniques using convolutional neural networks (CNNs) to predict complicated features that possess physical significance.
- The deep learning model was trained on 12 seconds of data segment, aimed at predicting the 27 features of any LIGO time series data.
- Curriculum learning Bengio et al. (2009) was used to generate the dataset for training, involving first training the model with easier datasets before moving on to more difficult datasets.
- Synthetic preparation of data using numerical relativity simulations for black hole binaries (BH), neutron star binaries (NS), neutron star black hole binaries (NSBH) and core collapse supernovae (CCSNe) was introduced as a methodology.
- The maximum accuracy attained during training was 90.93%, with a validation accuracy of 89.97%.
- The model does face accuracy loss due to scattered labels, small dataset required for extraction of 27 features and slightly unbalanced data; however it could lead to more accurate predictions and classifications in future supported by National Science Foundation research grants paving way for further research in this area.

## Summary

In recent years, advancements in Deep Learning techniques towards Gravitational Wave astronomy have led to a significant rise in various classification algorithm development. While some of these algorithms have been successfully employed to search gravitational waves from binary blackhole merger events, they still lack success with significant time duration and further prediction of the merger events parameters. To address this issue, researchers aimed to advance the boundaries of deep learning techniques using convolutional neural networks (CNNs) to go beyond binary classification and predict complicated features that possess physical significance. The method presented is not a replacement for already established and thoroughly examined methods like matched filtering for the detection of gravitational waves but is an alternative method wherein human interference is minimal. The deep learning model was trained on 12 seconds of data segment, aimed at predicting the 27 features of any LIGO time series data. During training, the maximum accuracy attained was 90.93%, with a validation accuracy of 89.97%. To generate the dataset used for training, researchers used curriculum learning Bengio et al. (2009), which involves first training the model with easier datasets before moving on to more difficult datasets. Each data component previously discussed in Sec-2 has been synthetically injected primarily as an individual class to the background wave while maintaining physical reality and ensuring that the model works efficiently in real-time combination of individual trigger events' (BH binaries, NS binaries, CCSNe) combination of synthetic data injected randomly in sample background wave. Researchers also investigated GW Echoes and were able to synthetically inject two categories of signals proposed by Abedi et al. (2017) as well as Lo & Weinstein (2017). They carried out an in-depth analysis of state-of-art deep CNNs for multiclass classification to classify data during the running time of LIGO detector itself. While preparing the dataset, researchers generated several examples not containing any injections by simply skipping injection steps so that the model maintains physical reality as far as possible. They generated four times as many examples that contain an injection than a noise sample. In conclusion, researchers introduced a methodology involving synthetic preparation of data using numerical relativity simulations for black hole binaries (BH), neutron star binaries (NS), neutron star black hole binaries (NSBH) and core collapse supernovae (CCSNe). They critically examined existing classification methods opted and well tested before attempting to extend classification problems towards multi-label classification tasks such as predicting 27 features from LIGO time series data within 12 seconds duration with maximum accuracy attained being 90.93%. The model does face accuracy loss due to scattered labels, small dataset required for extraction of 27 features and slightly unbalanced data; however it could lead to more accurate predictions and classifications in future supported by National Science Foundation research grants paving way for further research in this area .