Gaussian Process Acceleration of Eccentricity Reduction in Binary Black Hole Simulations

Accurately reducing orbital eccentricity in numerical-relativity (NR) simulations of binary black holes is essential for producing astrophysically relevant gravitational waveforms. Standard eccentricity reduction currently relies on iterative refinement schemes of initial orbital parameters — often requiring four or more trial simulations to achieve target eccentricities, over multiple days or longer. We introduce a machine learning framework that reduces the number of iterations, often to just one. Using Gaussian Process Regression (GPR) trained on a large archive of past NR simulations, we model corrections to post-Newtonian initial guesses for the orbital frequency and radial expansion rate as functions of binary parameters. Our trained model predicts improved initial parameters that reliably output quasi-circular orbits. We demonstrate the power of data-driven methods in accelerating NR pipelines and exploring binary black hole parameter space.