Speaker
Description
Next generation neutrino telescopes such as IceCube Gen2 call for inference methods that carry realistic detector response, atmospheric backgrounds, and selection effects through to population constraints. In practice, the ingredients that matter most for neutrino sky analyses, including effective area, energy dispersion, exposure, and angular resolution, make analytic likelihoods hard to specify and expensive to evaluate across many hypotheses. We are developing an amortized simulation based inference pipeline that ingests multi energy sky maps and returns posteriors on parametric source population and nuisance parameters directly from forward simulations. The method learns spherical representations of the maps and couples them to a conditional normalizing flow density estimator. We will present validation on realistic simulated neutrino skies, with coverage based calibration checks and stress tests under detector systematics and atmospheric background modeling choices, aimed at discovery potential forecasts for Gen2 era analyses.