Artificial Neural Networks Search for Gravitational Waves
Board Location: #42
Discipline: Physics
Subcategory: Astronomy and Astrophysics
Session: 2
Joe Popp - Saint Joseph's University
Co-Author(s): Marco Serra, National Institute for Nuclear Physics, Rome, IT
The sensitivity of the current gravitational wave detectors is only enough to detect signals from the strongest sources: binary black hole and binary neutron star collisions. The detection of other types of gravitational waves from other sources is important to continue the development of this field of physics. Using machine learning techniques to find gravitational waves from weaker sources has the potential to give physicists the ability to see through the high levels of noise in the data. In this study, we train a neural network to look for long-transient wave signals produced by newly born magnetars in the time-frequency domain. This type of signal is important because they can teach us more about the structure and behavior of neutron stars as a whole. To accomplish this, we used PyTorch to train a convolutional neural network using simulated data. The simulation is made to project randomly generated long-transient wave signals onto a square matrix of noise that resembles a time-frequency map that could be generated using real data. This is used to train and test our convolutional neural network. Training is repeated with variable model sizes and training loop inputs until the model can be successfully trained using data with a low signal strength relative to the background noise. With noise simulated from an exponential distribution as a function of the frequency, we were able to find a model and a proper set of training loop inputs to get an accuracy of 86.9% at a signal strength of 1.5e-23 1/√Hz above the noise. To train our model to detect signals in real data, future work would include much more of this training to refine the parameters of the training loop. Changes to the simulation are also likely required to more closely resemble the noise and signals being received by the detectors. Once our model can be trained to detect signals under the noise with a more accurate simulation, it will likely be able to find these signals and determine their characteristics in real data.
Funder Acknowledgement(s): This project was funded by the National Science Foundation fellowship number PHY-1950830.
Faculty Advisor: Marco Serra, marco.serra@roma1.infn.it
Role: I created the simulation, built all the structures required for the machine learning, and all of the training. This was an individual research project with my advisor.