Gravitational-wave surrogate models powered by artificial neural networks

Inferring the properties of black holes and neutron stars is a key science goal gravitational-wave (GW) astronomy. To extract as much information possible from GW observations, we must develop methods to reduce cost Bayesian inference. In this paper, use artificial neural networks (ANNs) parallelization power graphics processing units (GPUs) improve surrogate modeling method, which can produce accelerated versions existing models. As first application our model (ANN-Sur), build time-domain spin-aligned binary hole (BBH) waveform SEOBNRv4. We achieve median mismatches approximately $2\mathrm{e}\ensuremath{-}5$ no worse than $2\mathrm{e}\ensuremath{-}3$. For typical BBH generated 12 Hz with total mass $60\text{ }\text{ }{M}_{\ensuremath{\bigodot}}$, original SEOBNRv4 takes 1794 ms. Existing custom-made code optimizations (SEOBNRv4opt) reduced 83.7 ms, interpolation-based, frequency-domain SEOBNRv4ROM generate in 3.5 Our ANN-Sur when run on CPU 1.2 ms unit (GPU) just 0.5 also large batches waveforms simultaneously. find that up ${10}^{3}$ be evaluated GPU 1.57 corresponding time per 0.0016 This method promising way utilize GPUs drastically increase computational efficiency parameter estimation.