Orateur
Description
Accurate localisation of continuous gravitational waves (CGWs) from supermassive black hole binaries (SMBHBs) remains one of the key challenges in Pulsar Timing Array (PTA) data analysis. Traditional searches based on the $\mathcal{F}_e$ statistic provide a robust analytic framework, but the resulting sky maps are strongly affected by the PTA antenna pattern, which redistributes signal power across the sky and generates secondary peaks that complicate the identification of the true source position. This degeneracy motivates the development of alternative approaches capable of disentangling instrumental artefacts from true localisation information.
In this work, we investigate whether Sequential Simulation-Based Inference (SBI) can improve CGW sky localisation by learning the mapping between $\mathcal{F}_e$ maps and true source positions. Using simulated PTA datasets spanning multiple configurations, isotropic and non-isotropic 25-pulsar arrays, with and without the pulsar term, and under non-uniform sensitivity, we demonstrate that our SBI pipeline effectively marginalises over antenna pattern search artefacts, providing reliable posterior distributions for the source coordinates. A key advantage of this framework is its extreme computational efficiency: our pipeline can generate hundreds of thousands of $\mathcal{F}_e$ statistic maps in approximately 3 minutes, with the subsequent network training completed in under 2 hours on a single GPU. This enables both the rapid generation of large training sets and near-instantaneous source characterisation. Our results show that the angular resolution ($\Delta\Omega$) achieved via SBI is consistent with the theoretical lower bounds predicted by the Fisher Information Matrix.
We further extend this framework to the detection problem, training a binary neural classifier on the same $\mathcal{F}_e$-statistic maps to distinguish signal-containing observations from noise-only realisations. Independent classifiers are trained for each GW frequency bin, covering the sub-threshold SNR regime ($\mathrm{SNR} \in [3, 10]$). We characterise detection performance across all tested configurations, finding consistent and robust performance.