Machine Learning for Nanohertz Gravitational Wave Detection and Parameter Estimation with Pulsar Timing Array

Chen MengNi,Zhong YuanHong,University of Chinese Academy of Sciences,University of KwaZulu-Natal,Li Jin
DOI: https://doi.org/10.1007/s11433-020-1609-y
2020-01-01
Science China Physics Mechanics and Astronomy
Abstract:Studies have shown that the use of pulsar timing arrays (PTAs) is among theapproaches with the highest potential to detect very low-frequencygravitational waves in the near future. Although the capture of gravitationalwaves (GWs) by PTAs has not been reported yet, many related theoretical studiesand some meaningful detection limits have been reported. In this study, wefocused on the nanohertz GWs from individual supermassive binary black holes.Given specific pulsars (PSR J1909-3744, PSR J1713+0747, PSR J0437-4715),the corresponding GW-induced timing residuals in PTAs with Gaussian whitenoise can be simulated. Further, we report the classification of the simulatedPTA data and parameter estimation for potential GW sources using machinelearning based on neural networks. As a classifier, the convolutional neuralnetwork shows high accuracy when the combined signal to noise ratio ≥1.33for our simulated data. Further, we applied a recurrent neural network toestimate the chirp mass (ℳ) of the source and luminosity distance(D_p) of the pulsars and Bayesian neural networks (BNNs) to obtainthe uncertainties of chirp mass estimation. Knowledge of the uncertainties iscrucial to astrophysical observation. In our case, the mean relative error ofchirp mass estimation is less than 13.6%. Although these results areachieved for simulated PTA data, we believe that they will be important forrealizing intelligent processing in PTA data analysis.
What problem does this paper attempt to address?