A fast, differentiable neural-network surrogate for precessing binary black-hole waveforms

Abstract

Gravitational-wave parameter estimation requires millions of waveform evaluations per event, a cost that constrains real-time inference and population studies. We present a fast, fully differentiable neural-network surrogate for the precessing numerical-relativity model \NRSur{}, spanning its full intrinsic parameter space $\lambda=(q,\,\boldsymbol{\chi}_1,\,\boldsymbol{\chi}_2)$ together with the reference orbital frequency $\omega_0$. Rather than a single polarization at a fixed orientation, the surrogate predicts the \emph{inertial-frame spherical-harmonic modes} $h_{\ell m}$ ($\ell\leq 4$), so that both polarizations $h_+,h_\times$ at an arbitrary orientation $(\iota,\varphi,\psi)$ are reconstructed from one network evaluation through an analytic, differentiable mode-to-strain projection. Trained on $6\times10^5$ waveforms, it attains a fixed-orientation match of mean $0.975$ (median $0.988$) and an orientation-averaged match of mean $0.940$ (median $0.975$) for $q\in[1,4]$, $|\boldsymbol{\chi}_{1,2}|\leq 0.8$, while keeping the overall strain amplitude physical (median ratio $0.98$). It generates a waveform in $\SI{12}{ms}$ (single) and ${\sim}3.5\times10^4$ per second in batches on a single GPU. Because the mode-to-strain projection is analytic, the surrogate is differentiable in both intrinsic and extrinsic parameters, yielding a full 13-dimensional Fisher matrix (validated against finite differences) and gradient-based (HMC/NUTS) parameter estimation.