BEAST HackDay Intro (17 Jan 2020)

Transcript

1. The BEAST Karl D. Gordon Space Telescope Sci. Inst. Baltimore, MD, USA BEAST hack day 17 Jan 2020 STScI
2. The BEAST Bayesian Extinction And Stellar Tool • SED fitter for individual stars • Include stellar and dust physics • Use Probabalistic/Bayesian techniques • Accurate noise model for crowded data • Must be fast – need to fit many stars – PHAT has >100 million stars
3. Model of single star in a galaxy • The physical description of a stellar SED – 7 parameters give stellar and interstellar dust properties • Stellar evolutionary tracks – Birth mass, age, & metallicity – Mapping to radius, T(eff), & log(g) – Stellar atmospheres (surface flux) → T(eff) & log(g) • Dust Extinction curves – A(V) = Dust column; R(V) ~ average grain size – f A = mixing between dust w/ 2175 Å bump (Fitzpatrick 1999 ) and SMC Bar 2175 Å bumpless dust (Gordon et al. 2003 ) • The distance
4. BEAST Strengths • Parameterized dust extinction model – Encompasses the known behavior – MW, LMC, SMC, & M31 • Noise model includes covariance – Correlation in noise between observed bands – From crowding/confusion and absolute calibration • Based on Bayesian techniques – Allows for the inclusion of prior information – Allows for hierarchical models (ensemble modeling)
5. Visual BEAST SED Construction
6. New Mixture Model encompassing all known Local Group Extinction Curves Gordon et al. (2015) and Gordon & Tchernyshyov (in prep.)
7. Covariance! Covariance Matrix & Bias in Fitting Calculating Covariance & Bias from ASTs
8. Priors = Other Knowledge = Assumptions
9. Impact of Crowding/Confusion
10. Visualization of Covariance Matrix & Bias from ASTs (from dedicated ASTs tests with 125 realizations per model)
11. Visualization of 6-band covariance matrices (~2000 models x 20 realizations per model)
12. BEAST Results PHAT Star in M31
13. Recovery of Simulated Star
17. MegaBEAST • Multi-level model • Hierarchical Bayesian model • Ensemble model • Works by modeling the outputs of the BEAST – Mathematically equivalent to modeling the observed fluxes! – But only needs to model the BEAST fit parameters • Age, mass, metallicity • A(V), R(V), f A • Account for completeness and star/dust geometry, star formation history, ...
19. BEAST Math • Bayes’ Theorem – P(A|B) = P(B|A)P(A)/P(B) • P(θ|d) = P(d|θ)P(θ)/P(d) – θ = BEAST model parameters • Age, mass, metallicity, A(V), R(V), f A , distance • Observation model parameters: source density – d = data (fluxes) – P(θ) = priors – P(d) = constant for all stars (grid solution)
20. MegaBEAST Math • P(θ|d) = P(d|θ)P(θ) = BEAST output • P(φ|d) = product[ int i { P(φ|θ)C(θ)(P(θ|d) i } ] – φ = MegaBEAST parameters • “meta” parameters • Star formation vs time, IMF, mass-metallicity, • log-normal parameters for A(V), R(V), f A • Distance function (Gaussian with width?) – P(φ|θ) = P(θ|φ)P(φ)/P(θ) = MegaBEAST computation • Adjustable priors • P(θ) = 1 (??) – C(θ) = completeness in θ space • From ASTs run from BEAST models – P(θ|d) = BEAST/P(θ) ∏∫ i P(ϕ∣θ)
21. Results from the BEAST
22. Results from the MegaBEAST
24. Decomposed A(V) maps for Brick 21 Foreground M 31
25. BEAST/MegaBEAST Effort • Technique paper – Gordon et al. (2016, ApJ, 826, 104) • Active team – Gordon, Boyer, Choi, Clark, Fouseneau, Goldman, Hagen, Johnson, Lindberg, Murray, Tollerud, Van De Putte, Williams, Wu • Regular BEAST Hack Days – Code development, testing code, documentation, etc. • BEAST is fairly mature – public on github – Distributed development of code, documentation, etc. • MegaBEAST in development
26. FatBEAST 29 Dec 2017 • 1st official release! • Captures what was run on the PHAT data – 45 million stars • See BEAST paper for details
27. DistantBEAST (v1.1) 28 Apr 2018 • Distance added as 7th variable • Optimized generation of ASTs inputs for toothpick noismodel • AST input locations based on surface brightness • Automated regression testing • Pip installable version • Docs, Bugfixes, etc.
28. UnleashtheBEAST (v1.2) 22 Jun 2018 • ASTs based on source density or surface brightness • Model grid can be split – Allow for parallel execution – Resulting pDFS can be reassembled • Systemic velocity of galaxy added • Symlog for flux pPDFs allows negative values • Docs, Bugfixes, etc.
29. HerdtheBEAST (v1.3.2) 5 Sep 2019 • AGB model atmospheres added • Generalized dust extinction curves • Simulated observations • run_beast.py split up (more pythonic) • Simple splinter noisemodel added • Python 2 support dropped • Docs, Bugfixes, etc.
30. DevBEAST (v2.0) • Lots of script development – Batch files, xsede • Adjustable stellar priors • Docs, Bugfixes, etc.
31. BEAST READY! • BEAST is ready for use • Priority is to write BEAST papers – Talk by Lea & github issue #333 • Improvements continuous, but major capabilities are there • Issues for work needed – Categorized by area and effort level
32. MegaBEAST – DEV • Currently only solves for A(V) parameters • Internal mechanics setup for full solution • Need to expand to allow for fitting other parameters – R(V), f_A – Star formation history, metallicity history, IMF, etc. • Testing, testing, testing – Gain confidence that MegaBEAST working – Test versus MATCH (high priority for Scylla)
33. Thanks!