Waveform AnomaLy DetectOr (WALDO)

WALDO is a deep learning data quality tool developed to flag possible anomalous Gravitational Waves (GW) from Numerical Relativity (NR) catalogs. We use a U-Net architecture to learn the waveform features of a dataset. These waveforms are timeseries $h_{lm}(t)$ of modes $(l,m)$ from the spin-weighted spherical harmonics decomposition of the GW strain $h(t,\vec x)$,

$$h_{lm}(t) = \int d\Omega h(t, \vec x)_{-2}Y_{lm}^*(\theta, \phi) .$$

WALDO computes the mismatch between $h_{lm}(t)$ and its prediction $\bar h_{lm}(t)$ to compose a histogram. We can identify anomalous waveforms by isolating 1% of the highest measurement values. Below, the anomaly search associated with the radiation field $\psi_{32} = \ddot h_{32}$ from the dataset.

Installation

To install WALDO, we can use the pip command:

pip install grav-waldo

Content

The project is composed of three main codes:

• wfdset.py: for pre-processing NR dataset;
• unet.py: the neural network;
• waldo.py: for mismatch evaluation and anomaly search.

Check the tutorials in docs.

Reference

WALDO's paper: Deep learning waveform anomaly detector for numerical relativity catalogs.