Simulation-based Inference of Massive Black Hole Binaries using Sequential Neural Likelihood

We propose a machine learning-based approach for parameter estimation of Massive Black Hole Binaries (MBHBs), leveraging normalizing flows to approximate the likelihood function. By training these flows on simulated data, we can generate posterior samples via Markov Chain Monte Carlo with a relatively reduced computational cost. Our method enables iterative refinement of smaller models targeting specific MBHB events, with significantly fewer waveform template evaluations. However, dimensionality reduction is crucial to make the method computationally feasible: it dictates both the quality and time efficiency of the method. We present initial results for a single MBHB with Gaussian noise and aim to extend our work to increasingly realistic scenarios, including waveforms with higher modes, non-stationary noise, glitches, and data gaps.