Microlensing Detection and Inference via Learned Bayes Factors

概要

We present a unified framework for gravitational microlensing event detection and parameter inference. Traditional pipelines use deterministic hard cuts on photometric statistics, systematically missing low-magnification events in the finite-source regime. We instead frame detection as Bayesian model comparison using Evidence Networks, which learn calibrated Bayes factors from binary-labeled simulations, and combine this with Neural Posterior Estimation (NPE) for amortized parameter inference. B…