Geological mapping in Mercury's southern latitudes

点击翻译按钮获取中文摘要

Description:The aim of this project is to use a relatively new machine learning technology, GenerativeAdversarial Networks, to predict the unresolved structures in submillimetre-wavelengthimages and then use these predictions to find candidate ultra-high redshift galaxies.Extreme high-redshift galaxies can be detected in submm-wavelength surveys made by theHerschel Space Observatory and others, but because of the redshifting, these galaxies aretypically only detected at the longest wavelengths, e.g. 500 microns. At these wavelengths,the angular resolution is at its coarsest, so distant galaxies are seen as blended withforeground interlopers. Several groups have attempted to get around this by modelling thedata at all wavelengths. This approach deconvolves the longest wavelength data by usingthe locations of galaxies detected at other wavelengths as a prior, but this iscomputationally difficult and solutions are not unique. Generative Adversarial Networks area relatively new machine learning technology that, once trained, can reconstruct missinginformation in images. This has been shown to work in blurred images of galaxies,accurately reconstructing galaxy morphologies in many cases. This project will deploy thistechnology on real and simulated submm-wave images, find candidate ultra-high redshiftgalaxies, follow them up with ground-based and space observatories, place statisticalconstraints on ultraluminous star forming galaxies at the earliest accessible epochs andconstrain models of high redshift star formation