Speakers
Description
The detection of gravitational waves from binary black hole (BBH) systems by next-generation interferometers like the Einstein Telescope presents a complex parameter estimation challenge. This work explores a deep learning approach using Normalizing Flows, conditioned on feature vectors extracted by diverse machine learning models, to robustly estimate the chirp mass and luminosity distance, which are two out of 15 parameters describing a BBH system.
To determine the optimal feature extraction architecture, we evaluated and compared multiple preprocessing models on labeled, simulated waveform data. Our exploration includes 1D Convolutional Neural Networks (CNNs) with and without residual connections, 2D CNNs operating on Short-Time Fourier Transform (STFT) spectrograms, and time-series Transformer models. Furthermore, to reduce the impact of colored detector noise, we integrated power spectral density (PSD) estimation and data whitening into our preprocessing phase.
Our results demonstrate that both a 1D Residual CNN and a 2D CNN utilizing whitened spectrograms as preprocessing models achieve the highest predictive accuracy for parameter estimation compared to the other models. Specifically, comparing the 2D CNN variants highlighted that data whitening prior to convolution results in better model performance.
Finally, we investigated the viability of transfer learning to expand our estimations to a wider array of BBH parameters, such as right ascension, declination, and tilt. By freezing the base layers of a pre-trained 1D CNN and re-training only the final linear layer, we found that shallow re-training is insufficient for robust multiparameter transfer. This suggests that future multiparameter estimation will require deeper network fine-tuning and re-training.