Robust $\mathbf{M_h}$ using HaloFlow with Domain Adaptation

Precise halo mass ($M_h$) measurements are crucial for cosmology and galaxy formation. HaloFlow provides a new approach using simulation-based inference and state-of-the-art simulated galaxy images that can accurately measure $M_h$ with significantly higher precision. However, HaloFlow requires a simulated training dataset, and is thus limited by domain shifts. In this work, we extend HaloFlow with unsupervised domain adaptation (DA) methods to improve the robustness of inferred $M_h$. We implement two DA methods: a domain-adversarial network (DANN) and Maximum Mean Discrepancy (MMD) alignment. We test the performance of our DA methods on a suite of three different cosmological hydrodynamic simulations. Our results show that DA significantly improves robustness: in one test case, MMD reduces the normalized bias metric $\beta$ by 9-50\%. These gains represent a key step toward applying HaloFlow for reliable $M_h$ inference on galaxy survey observations.