Auto-encoder model for faster generation of gravitational waveform approximations

The LIGO-Virgo-KAGRA international collaboration have observed more than 250 gravitational wave events using their global network of observatories. Estimation of source parameters for each of these events requires about a million likelihood computations to properly constrain the posteriors. Each such computation entails solving the general relativity equations to obtain a theoretical waveform, which is then matched against the detected signal. This operation is computationally heavy, especially in the case of complex waveforms. The upcoming gravitational wave observatories, with an estimated $10^4-10^6$ detections per year, make it imperative to have solutions for the evident bottleneck for rapid parameter estimation. Towards this end, we present an auto-encoder model for generation of effective one-body \texttt{SEOBNRv4} binary black hole waveforms. We train our model with $\sim27,300$ samples. Our parameter space is made of the two binary component masses: $m_1,m_2\in[5,75]\,M_{\odot}$ with a hard mass ratio limit of $q=m_1/m_2<10$. Our model is able to generate $10^4$ samples in O(1) second, with a median polarization mismatch value of order $10^{-3}$. Our work provides the first step towards having a production ready framework for real-time rapid generation of highly-accurate gravitational waveform approximations. This will enable orders-of-magnitude faster online parameter estimation, while basically providing the same scientific potential.