Exploiting Deep Learning to Improve Roman Photometric Redshifts
Wide-Field Science – Regular
Brett Andrews / University of Pittsburgh, PI
Photometric redshifts (photo-zs) will be a critical ingredient for studies of both galaxy evolution and cosmology with Roman. The Roman Science Operations Center will estimate photo-zs, but these will only use integrated photometry in Roman bands, disregarding the key color information in Rubin bands and morphological information that recent deep learning methods have unlocked. At low redshift (z < 0.2), algorithms that exploit resolved imaging of galaxies via deep learning methods have delivered much better photo-z performance than those that rely on integrated photometry alone. However, at higher redshifts the potential constraining power of deep, multi-band resolved imaging has been inaccessible due to the comparatively small sizes of galaxies in comparison to ground-based seeing.
Roman will bypass this limitation by providing deep and well-resolved images of galaxies – even for those at z>1 – in multiple bands from the red-optical to the near-IR. We propose to apply bleeding-edge deep learning methods to existing Roman-like data sets with resolved imaging from Hubble to test how much improvement the incorporation of resolved imaging will yield for deep samples. This should be particularly effective for removing the degeneracies between low and high redshift that plague many photo-z algorithms.
Specifically, we propose to train a state-of-the-art type of deep neural network – a masked autoencoder – to distill color and morphology information from Roman-like images from CANDELS and 3D-DASH HST imaging into low-dimensional encodings. We will then test our ability to predict the high-precision multi-band photo-zs from COSMOS2020 and from spectroscopic redshifts using regression from the low-dimensional encodings and integrated Roman + Rubin-like broad-band photometry. We will compare the results to photo-zs estimated with methods that only use integrated Roman + Rubin-like broad-band photometry.
The strikingly distinct morphologies of low-z and high-z galaxies, as well as the radically different resolved morphologies observed in bands above versus below the 4000Å break, should be instrumental in breaking the photo-z degeneracies that beset photometry-only methods. Based upon the performance of deep learning-based photo-z methods with resolved SDSS imaging in past work, we expect that the more modern techniques we plan to provide should produce the most accurate Roman photo-zs across a broad range of magnitude-redshift-color space. We propose to develop and test the algorithms needed to realize that potential with existing HST optical and NIR imaging. Should our method prove valuable, additional time and effort (beyond the scope of this proposal) will be required to implement it at scale on Roman data (in coordination with the Roman Science Operations Center) early in the mission’s lifetime for maximum benefit, highlighting the urgency of our development effort.
We will share the resulting code, produce thorough documentation (including lessons learned), and release the new CANDELS photo-z catalogs with the community. The resulting photometric redshift catalogs could significantly enhance the legacy value of the CANDELS data sets as a free byproduct of our work to develop new methodologies for Roman.