Gravitational-wave source inference for LISA is a computationally intensive task due to costly likelihood evaluations.
In this talk, I will introduce a surrogate-based approach using Gaussian processes (GPs) to accelerate parameter estimation of deterministic LISA signals. Such method actively learns the posterior surface with minimal likelihood evaluations. I will present benchmarks against established samplers, on simulated LISA data containing injections of double white-dwarf systems, stellar-mass black hole binaries, and supermassive black hole binaries.
We find that GPry is able to yield accurate parameter estimates at a small fraction of the cost of traditional samplers, opening the door to more accurate instrument and signal descriptions at inference time.