New machine learning framework for including selection effects in population models

Understanding the underlying population of binary black holes requires accurate modeling of selection effects, which is the probability that a binary black hole with given source parameters will be detected, is essential for inferring the underlying population. Standard injection-based approaches perform well for common sources but become inefficient and potentially biased for rare complex systems, such as those with large mass ratios, strong spin precession, or measurable eccentricity, which may encode signatures of dynamical formation or other non-standard formation channels. We present a machine-learning framework that employs conditional normalizing flows to model selection effects directly from simulated data, learning the mapping between source parameters and detectability in a high-dimensional space. Once trained, the model enables fast, flexible forward modeling without importance re-weighting or population assumptions, and provides unbiased estimates of detection fractions even in sparsely sampled or high-variance regions of parameter space.