Simulation-based inference of profile evolution for high-precision pulsar timing
Pulsars are exceptional cosmic laboratories, the most stable of which rival atomic clocks. Nearly all of that precision rests on one observable, the average pulse profile unique to each pulsar and stable over decades. These profiles change shape systematically with observing frequency, reflecting the physics of the emitting pulsar beam, but if left unmodelled, it can bias the arrival-time measurements, undermining precision timing. Profile-domain timing has been proposed to address this effect, but current implementations use stochastic sampling, which is computationally prohibitive and has limited the application in large-scale analyses. Here, we present a simulation-based inference approach to profile-domain timing, using a normalising flow conditioned on a convolutional embedding of the frequency-phase portrait. From a single portrait, it returns joint posteriors on dispersion measures, arrival times, and the parameters governing profile evolution instantaneously, with uncertainties verified against known inputs following an initial training run. The value of a likelihood-free approach is that the simulator becomes the interface to pulsar properties. Fundamental properties like non-Gaussian noise, pulsar self-noise, jitter, scattering, and profile change can be inferred by simulating them rather than by deriving a new likelihood for each, opening a promising pathway to include machine learning methods into pulsar timing workflows.