Campus, Varian 355
In this talk, I will discuss the forward model approach to reconstruct cosmological fields in a Bayesian framework with focus on two examples - neutral hydrogen intensity mapping and galaxy clustering. For neutral hydrogen surveys, we loose over 50% of the modes at high redshifts due to foregrounds and it severely hampers their feasibility for cosmological analysis. With a novel bias framework for the forward model, I will show that we are able to reconstruct over 90% of these modes and this recovers cross-correlations with photo-z surveys like LSST and tracers like CMB lensing. In galaxy clustering example, we are interested in reconstructing the initial Lagrangian field since it has power spectrum as the singular summary statistic, and enhanced signal for baryon acoustic oscillations. Here, we develop a novel framework with neural networks to forward model halo masses and positions and demonstrate that our method outperforms standard reconstruction in both real and redshift space. Lastly, I will briefly touch upon assumptions made in this reconstruction framework regarding noise model and likelihood and preliminary ways to improve upon them using deep learning.