1. WFIRST Photo-z Study Peter L. Capak, Andreas Faisst,
Shooby Hemmati, Dan Masters, Nathaniel Stickley
Caltech Infrared Processing and Analysis Center

2. Part I – Calibration Overview
P. Capak

3. Why Do Photometric Redshifts Work?
Rahman et al. 2015
Galaxies are very similar in many ways
Same physics
Cluster in color space
Photo-z is a map of the color space manifold to redshift
High dimensionality
Complex manifold
Lots of ways to do this mapping
Error distributions and systematics are important

4. Mapping the Color-Redshift relation
Can visualize the 7-dimentional color data
Test if the analytic model fits
Test where the data driven model is valid
Target grey areas with spectroscopy
Analytic (Template) Model
Data (spectra) Driven Model
Now we can see where and how well our models work
Need much more data to fill in the data driven model
Will be applying for NASA keck time to do that
Essential for the success of Euclid and WFIRST

5. Mapping the Color-Redshift relation
The SOM map also encodes P(color) and cosmic variance
P(color) is important because it allows you to use the whole sample to estimate the likelihood of degeneracy
Cosmic variance encoded in density at level of photometry
Typical cell has dz~0.02
Mean occupation by cell

6. Why Do Photometric Redshifts Work?
30 Band
Weight photometry by likelihood of data point
Weight models by likelihood they are represented in the data
Using only u,g,r,I,z,Y,J,H photometry
Significantly improve over raw template fitting
Outlier fraction > 1.5%
Sigma_NMAD=0.03 -> 0.02
Bias = 0.004
Photometric Redshift
8 Band 5
Photometric Redshift 4 3 2 1 1 2 3 4 5
Spectroscopic Redshift
Spectroscopic Redshift

7. Application to WFIRST/LSST
COSMOS Measured
GOODS-N interpolated
Used CANDELS data to simulate a real WFIRST lensing sample
Interpolated to LSST/WFIRST bands based on existing CANDELS photometry
Matches actual CFHT-LS + VISTA data
Collected ~50k high-quality redshifts across CANDELS + COSMOS + SXDS + VVDS fields
Single CANDELS field samples only a small fraction of CFHT-LS + VISTA color space
Z-H
G-Z

8. Application to WFIRST/LSST
Euclid samples ~90% of LSST/WFIRST Color space
WFIRST will require more z>3 spectroscopy than Euclid
G-H H
G-z I

9. Part II – WFIRST Specific Tests
D. Masters

10. WFIRST Specific Tests Need an analog to WIFIRST
CANDELS is closest data set
It is small (0.2 sq deg) and heterogeneous
Need to match it to LSST + WFIRST

11. Why Do Photometric Redshifts Work?
Galaxies are very similar in many ways
Same physics
Cluster in color space
Photo-z is a map of the color space manifold to redshift
High dimensionality
Complex manifold
Lots of ways to do this mapping
Error distributions and systematics are important
Rahman et al. 2015

12. Color distribution to Euclid depth
SDSS + COSMOS g-r vs. g-i
SDSS g-r vs. g-i
Galaxy distribution in multicolor space is (still) limited and measurable to Euclid depth!
11/14/2019

13. Moving CANDELS catalogs to LSST+WFIRST filter set
after applying WFIRST lensing criteria
~30,000 galaxies
LSST+WFIRST
11/14/2019

14. SOM on WFIRST+LSST catalog
11/14/2019

15. SOM on WFIRST+LSST catalog
 SOM is trained well.
11/14/2019

16. Redshifts on WFIRST SOM
11/14/2019

17. EUCLID, C3R2, the rest…
riz>25
11/14/2019

18. EUCLID, C3R2, the rest…
11/14/2019

19. EUCLID, C3R2, the rest…
While WFIRST galaxies that will not be in the EUCLID sample (riz>25) fill ~50% of the WFIRST color-space
The fraction of the WFIRST color-space not filled with EUCLID sources with similar colors is less than 5%, which will need further spectroscopy.
11/14/2019

20. Part III – Simulating Spectra
P. Capak

21. Simulating Spectra
An essential part of calibration is the spectroscopy
This is hard to simulate correctly due to the wide variety of galaxies and conditions
We developed a data driven model to do this
Based on SPHEREx mission simulations
Designed to re-produce emission line strengths and evolution as well as colors
Stickley et al. 2016

22. Simulating Spectra
Using SED fits from CANDELS -> WFIRST conversion to estimate spectra
Based on SPHEREx interpolative simulations
Using Brown et al templates
Predict actual emission line properties to 0.2 dex
Continuum SNR to 20% for Keck

23. Simulating spectra Working for Keck instruments
Adapted for WFIRST IFU and grism
Being used to determine how much and what spectroscopy is required

24. IFU Redshifts Weak lines at >0.8um Can be done with IFC
85% of WFIRST sample (17% of total)
Also “Hard” with GRISM
Need many 10’s of hour exposures

25. Part IV – Beyond Redshifts
P. Capak

26. Advanced Computational Techniques – Beyond Redshifts
The SOM map also encodes P(color)
Can be used to construct hierarchical priors
If redshift is well constrained P(color) = cosmic variance
Can be used to correct cosmic variance in small area surveys
Redshifts
Density Of Sources
Masters, Capak et al. 2015

27. Advanced Computational Techniques – Beyond Redshifts
Generated a 15-dimensional SOM on the SPLASH data
Verify sample selection and photo-z for high-z sample
Davizon (Capak) et al. 2017

28. Application Of Big Data Tools
Analytic Models
Unsupervised Learning
Supervised Learning
Analytic Models
Fundamental understanding of underlying problem
Supervised Learning
-Create an empirical model of complex systems based on data
-Reproduce human vision/intuition on large data sets\
-Complex interpolation
Unsupervised Learning
-Discover correlations you didn’t know existed
-Visually explore high-dimensional data sets
-Compare analytic models and complex data to look for discrepancies
-Understand data properties

29. Advanced Computational Techniques – Model Testing
χ2 of Best Fit Model
Best fit analytic models with respect to SOM grid
Clear regions where the fit is much worse
Our analytic model is not good in these regions
Why?
Capak et all in prep

30. Advanced Computational Techniques – Model Testing
The fit is bad because we do not correctly model red galaxies
Goodness of fit
Model
Dust Obscuration
Good
Bad
Red
Blue
Blue
Red
Capak et all in prep

31. Application Of Big Data Tools – Next Generation Surveys
Need a “Standard Model” of galaxies
Represents what we know about galaxies
Combines information from multiple surveys
Allows for optimized analysis
Need a “Standard Model” of galaxies
Needed to interpret large statistical data sets
Represents what we know about galaxies
Largely data driven statistics at this point
Allows for interpretation of data from telescope
Without modeling individual galaxies
Allows for clean combination of information from multiple surveys
Allows for optimized cosmological analysis
Choose what you marginalize over

32. Next Generation Surveys – Standard Model
Physical Models
Red
Blue
Data Model
Age
Using SOM as illustrative tool,
but there are better methods for this
Need to taylor to the problem
Mass

33. Next Generation Surveys – Standard Model
Physical model of galaxies
Red
Blue
Mass Function
Log10(Space Density of Galaxies)
Redshift Model
Sky Density
Log10(Mass of Galaxy)

34. Part V – Model Fitting with SOM
A. Faisst

35. SED fitting Challenges
Empirical libraries suffer from completeness
Theoretical libraries are not constrained
Can we combine the two?
Future missions such as LSST, EUCLID, WFIRST, will provide photometric data for millions of galaxies where fast and accurate measurement of their physical properties is helpful.
Can we measure physical properties faster using machine learning techniques?
11/14/2019

36. SOM to visualize galaxy model libraries
1) Make a library of model SEDs using different physical parameters:
2) Convolve model SEDs with filter sets dictated by observations.
lets say COSMOS
3) Multi-dimensional space defined by colors of the model SEDs in the filter set.
4) SOM will reduce multi-dimension to two dimensions  visualize
11/14/2019

37. SOM to visualize galaxy model libraries
11/14/2019

38. SOM to visualize galaxy model libraries
Each SOM cell now represents different physical properties of galaxies
11/14/2019

39. SOM to Optimize galaxy model libraries
Observations can be used to optimize the parameter space used to build the model library
So lets map cosmos galaxies to the trained SOM:
11/14/2019

40. SOM to Measure galaxy physical properties
11/14/2019

41. Uncertainty of parameters measured with SOM
11/14/2019

42. More applications, galaxy selections such as the UVJ
11/14/2019

43. More applications, resolved stellar populations
11/14/2019

44. Part VI – Measuring Emission Lines With Photometry
A. Faisst

45. Optical Emission Lines at z > 4 with SPLASH (COSMOS)
Statistically estimate emission lines (Hα and [OIII]) at z > 4 from Spitzer [3.6μm]-[4.5μm] color
Faisst et al. (2016a)
also: Rasappu et al. (2016), Marmol-Queralto et al. (2015); Shim et al. (2011); Stark et al. (2013)

46. Optical Emission Lines at z > 4 with SPLASH (COSMOS)
Success on COSMOS (Faisst et al. 2016a)
SPLASH: deep (25.5AB) Spitzer over 2 deg2 of COSMOS
~500 spectroscopic redshifts (COSMOS/DEIMOS & VUDS)
Scoville+07, LeFevre+14
Hα in 3.6μm
[OIII] in 3.6μm
Hα in 4.5μm
no emission lines
Faisst et al. (2016a)

47. Optical Emission Lines at z > 4 with SPLASH (COSMOS)
Observed color vs. redshift relation =
Intrinsic color (dust, age, metallicity, SFH) +
Redshift dependent emission line strengths and ratios (Hα, Hβ, [OII], [OIII])
Faisst et al. (2016a)

48. Optical Emission Lines at z > 4 with SPLASH (COSMOS)
Observed color vs. redshift relation =
Intrinsic color (dust, age, metallicity, SFH) +
color not sensitive to these for young galaxies (< 1 Gyr)!
Redshift dependent emission line strengths and ratios (Hα, Hβ, [OII], [OIII])
Faisst et al. 2016a

49. Optical Emission Lines at z > 4 with SPLASH (COSMOS)
Consistent results at z < 3 where spectroscopy is available
Strong increase of Hα EW beyond z = 3
Hα Equivalent-Width
Faisst et al. (2016a)

50. Optical Emission Lines at z > 4 with SPLASH (COSMOS)
Consistent results at z < 3 where spectroscopy is available
Strong increase of Hα EW beyond z = 3
Convert EW(Hα) into sSFR (e.g., Cowie et al. 2011)
Hα Equivalent-Width
Specific Star Formation Rate
Hα derived sSFR
out to z = 6 from
Spitzer colors
5 times faster mass growth at z = 6 compared to z = 2
Faisst et al. (2016a)

51. Optical Emission Lines at z > 4 with SPLASH (COSMOS)
Use Spitzer colors to estimate [OIII]/Hβ ratios
sensitive to metallicity
probe metallicity of z > 4 galaxies statistically
Faisst et al. 2016a
Faisst et al. 2016b
[OIII] emitters!
see also Masters, Faisst, Capak 2016

52. Part VII – Combining Heterogeneous Data
P. Capak

53. Information from Heterogeneous data
WFIRST will cover a large area of the sky
Including deep multi-wavelength fields
We can use this information to calibrate photo-z
Looking at the galaxy population as a density field provides a way of doing this
SOM a convenient tool for this

54. Next Generation Surveys – Standard Model
Generate model with one data set
Expand it to higher dimensions with another
LSST
WFIRST
Spitzer

55. Next Generation Surveys – Standard Model
Generate model with one data set
Expand it to higher dimensions with another
Uniform Cell
Bifurcated population
Non-Uniform Cell

56. How to do this in practice
Deeper data contains more information
Better localization on SOM/density field
Target previous deep fields in survey with deep data
CDFS, SXDS, VVDS-2h, COSMOS, EGS, GOODS-N, ect
These can then be mapped onto the wider survey
Using SOM including Spitzer for high-z studies
Reduce but don’t completely remove caustics
Need to explore even higher dimensional data
Galex, HST UV