Mapping inspiral-merger-ringdown waveforms of binary black holes from black hole perturbation waveforms by machine learning

Identifying weak gravitational wave signals in noise and estimating the source properties require high-precision waveform templates. Numerical relativity (NR) simulations can provide the most accurate waveforms. However, it is challenging to compute waveform templates in high-dimensional parameter space using NR simulations due to high computational costs. In this work, we implement a novel waveform mapping method, which is an alternative approach to the existing analytical approximations, based on closed-form continuous-time neural networks. This machine-learning-based method greatly improves the efficiency of calculating waveform templates for arbitrary source parameters, such as the binary mass ratio and the spins of component black holes. Based on this method, we present \textit{BHP2NRMLSur}, a class of models (including nonspinning and spin-aligned ones) that maps point-particle black hole perturbation theory waveforms into NR and surrogate waveforms. The nonspinning model provides highly accurate waveforms that match the NR waveforms to the level of $\gtrsim 0.995$. The spin-aligned model reduces the required input parameters and hence improves the efficiency of the waveform generation -- it takes a factor of $\sim 50$ less time than existing NR surrogate models to generate $100,000$ waveforms, with a mismatch of $<0.01$ compared to the NR waveforms from the Simulating eXtreme Spacetimes collaboration.