This repository is the implementation of the paper Extraction of Binary Black Hole Gravitational Wave Signals in Detector Data Using Deep Learning by Chayan Chatterjee, Linqing Wen, Foivos Diakogiannis and Kevin Vinsen.
The aim of this work is to implement a neural network model, called denoising autoencoder, to extract binary black hole gravitational wave signals from stationary Gaussian and real detector noise. A denoising autoencoder is a deep learning architecture that consists of two neural networks, called an encoder and a decoder. Together, these two networks produce clean, noise-free versions of corrupted or noisy input data. The encoder compresses noisy data into a low-dimensional feature vector and the decoder reconstructs the pure input data, with noise removed from the compressed feature vector.
We have used a Convolutional Neural Network as our encoder and a Bi-directional Long Short Term Memory Network as our decoder. The architecture of the model is shown below: 
This is the first successful application of deep learning to denoise all ten binary black hole gravitational wave signals from real detector strain data at > 97% overlaps with pure waveform templates. We have also demonstrated that our model is robust against modelled glitches and non-injection samples, or pure noise.
This code is written in Python and uses the TensorFlow 2 package.
The training and test set samples used for this work were generated using the repository damonbeveridge/samplegen, which adds additional features to the sample generation code written by Timothy Gebhard and Niki Kilbertus described in the paper Convolutional neural networks: A magic bullet for gravitational-wave detection?
To train the model, simply run the main.py file:
python main.py
The hyperparameters of the network architecture and training parameters can be set in the file: configs/config.py. In this work, we have used a custom loss function called the FracTAL Tanimoto loss, first implemented in Diakogiannis at. al. (2021). This loss function has a hyper-parameter called depth, which steepens the loss curve towards the ground truth and helps escape local minimas during training. The depth should be set to 0 during the start of training and should be allowed to reach convergence. The trained model is saved in the folder called model. We can investigate possible improvements in the model's performance with a larger value of depth (and reduced learning rate). This can be done by restoring the model checkpoint corresponding to the minimum validation loss during training. The saved checkpoints can be found in the folder checkpoints.