Accurate redshifts are needed for all three techniques being used for WFIRST's dark energy experiments: studying cosmological parameters with samples of type Ia supernovae, identifying the imprints of baryon acoustic oscillations in the large-scale structure of galaxies, and probing the inhomogeneous matter distribution revealed by weak gravitational lensing of distant galaxies. The third technique requires measurement of hundreds of millions of faint galaxies, but obtaining spectroscopic redshifts of the entire set is impractical. Instead, photometric redshifts are used. Here, we explore how best to select spectroscopic subsamples with which to calibrate the photometric redshifts so that only a limited subset of galaxies need be targeted with spectra. The results reported here are based on the work of Peter Capak's "Precision Photometric Redshifts for Cosmology" WFIRST Preparatory Science Investigation Team and Olivier Dore's "Cosmology with the WFIRST High Latitude Survey" WFIRST Weak Lensing and Galaxy Redshift Survey Science Investigation Team.
Specifically, a machine learning algorithm called the Self Organizing Map (SOM) is used to reduce an N-dimensional color space to 2-D to identify undersampled regions of galaxy parameter space on which to concentrate spectroscopic efforts (see Masters et al. 2015, 2017). To demonstrate, we apply a cut to a deep photometric catalog from the Cosmic Assembly Near-infrared Deep Extragalactic Legacy Survey (CANDELS; Grogin et al. 2011, Koekemoer et al. 2011) to exclude galaxies which will not benefit the lensing analysis. The resulting catalog of targets is similar to what is expected from combined LSST + WFIRST weak lensing observations. Detailed information can be found in Hemmati et al. (submitted to ApJ).
As the CANDELS based simulation has the deepest currently available fields with multiband photometric data and the catalogs are H-band selected, it produces the best data-driven simulation of the expected WFIRST lensing sample. To benefit the weak lensing community, the results from the simulations are made available here. These include sample files that went into the training of the SOM, the trained SOM, and its application to simulated WFIRST+LSST photometry.
A FITS table of photometry of ∼ 36,000 objects from CANDELS five fields (0.2 sq deg) transferred to LSST (u,g,r,i,z) and WFIRST (Y,J,H,F184W) filters. In addition to photometry, the catalog contains redshifts (photometric and spectroscopic where available), FWHM in F160W, stellar mass, star formation rate, S/N in HST F160W, and Spitzer IRAC ch1 and ch2 magnitudes, based on measurements from the CANDELS team. A README describing the columns of the table is available.
This catalog is used to create, via the SOM, a 2-D map of the WFIRST color space. The SOM reduces the dimensionality of the color space while preserving its topology. That is, similar objects in multi-dimensional color space will remain neighbors on the 2D map.
An ASCII table of the trained SOM with the LSST+WFIRST lensing sample. This SOM covers an 80x60 grid with colors, median redshift and its standard deviation, and density at each cell.
The results of the SOM can then be mapped back into the input catalog.
A FITS table of the LSST+WFIRST lensing sample, containing the synthesized photometry in LSST and WFIRST filters (SED shape), redshifts (photometric and spectroscopic), Euclid riz photometry (with Euclid cut being brighter than 25 ABmag), SOM cell i and j, median redshift and its standard deviation, and density. A README describing the columns of the table is available.