Orateur
Description
A primary target for LISA will be massive black hole binaries (MBHBs), the inspiral of which will be detectable in a few days to months before coalescence. If they occur in gas-rich environments, they may produce electromagnetic (EM) counterparts, allowing for multimessenger follow-up.
To provide triggers to the EM community rapid parameter estimation is needed, specifically of merger time, sky localization, distance and masses, ideally already during the inspiral phase.
However, the Global Fit's detection and parameter estimation of all sources at the same time will not be suitable for near real-time requirements, motivating the development of a low-latency alert pipeline.
We present a simulation-based inference (SBI) approach, that avoids likelihood evaluations entirely and shifts the computational cost to the training phase.
We employ conditional flow matching with optimal transport, in which the trajectory from a base distribution to the parameters’ posteriors is learnt conditioning on data.
SBI is suitable not only because of the short inference time thanks to amortisation, but also because it can easily allow for realistic data conditions. Future developments will include real noise, instrumental artifacts and data gaps into the training process.
We focus on a sub-population of MBHBs, with masses between 10^4 and 10^5 solar masses, astrophysically motivated by the possibility of observing an EM signature for such systems .
We train and evaluate our network on two data representations: time series and time-frequency.
We compare various time-frequency representations and investigate the use of autoencoders to project the data into a space where the galactic binary foreground is ignored.
We present preliminary parameter estimation results and discuss the future development of a full, low-latency pipeline that can handle realistic LISA data.