PACIS: Precision and Accuracy for Cosmological Imaging Surveys

Cosmological measurements from new-era surveys of galaxies are expected to rigorously test the standard model of the Universe dominated by mysterious dark matter and dark energy. One of the most promising avenues towards this goal is to carefully study various physical aspects of matter distribution on the largest cosmic scales. This distribution is observationally traced by luminous beacons – galaxies, and new and better approaches to efficiently cataloging them and estimating their distances (via redshift) are necessary. The aim of this PACIS project is to harvest the largest imaging datasets and construct state-of-the-art galaxy catalogs with reliable redshift estimates, to be employed for precise and accurate cosmological tests using jointly weak lensing and galaxy clustering. For that purpose my team will develop robust and self-contained methodologies based on recent advancements in machine learning and Bayesian inference. We will apply our novel framework to modern photometric data, first capitalizing upon my expertise gained in the recently completed Kilo-Degree Survey - a unique multi-wavelength imaging dataset of the present time. We will then use it as a stepping stone towards the next-generation overwhelmingly big data from the forthcoming Legacy Survey of Space and Time. Following careful validation performed by my team, the novel galaxy samples extracted from these deep wide-angle datasets will be used for cosmological multi-probe studies based on lensing and clustering. To achieve the ambitious goals of PACIS, we will employ state-of-the-art computer techniques such as machine learning to robustly select complete and pure galaxy samples with statistically precise and accurate photometric redshifts from the full extent of the imaging datasets. To maximally exploit the rich information available in the input photometric data, we will work with two general approaches at galaxy selection, calibrated on spectroscopic samples overlapping with imaging data. The first method uses luminous red galaxies, which are bright up to large cosmological distances, can be very robustly identified from photometry, and their redshifts reliably estimated. The other approach is to select all galaxies brighter than a certain minimum observed flux limit. Such a bright-end selection reaches to lower redshifts than possible with red galaxies, but is much more complete and maps the large-scale structure in much more detail. We will ensure the reliability of these galaxy selections thanks to using adequate machine learning techniques to maximize completeness and purity of the samples. Precise and accurate redshift estimates will be derived for all the galaxies, in particular by using artificial neural networks, and the most novel deep learning. Final post-processing steps will make the datasets immune to various systematics inherent to observational data. This will enable employing the new samples for cosmological analyses, in particular for the very powerful multi-probe approach with weak lensing and galaxy clustering. The eventual aim of these measurements will be not only to constrain, but also to understand the physical nature of the dark components of the Universe.