Big HI Data and Artificial Intelligence Team Up to Decode the

Big HI Data and Artificial
Intelligence Team Up to Decode
the Multiphase ISM in Galaxies
GALFA-HI
Claire Murray
University of Wisconsin - Madison
Robert Lindner (UW Madison), Snežana Stanimirović (UW Madison), W. M. Goss
(NRAO), Carl Heiles (UC Berkeley), John Dickey (UTas), Patrick Hennebelle (CEA)
+ the rest of the 21-SPONGE team
Outline
1. Open questions, observations of the multiphase ISM
2. Pros and cons of future “big data”
3. Autonomous Gaussian Decomposition: turning the
cons into pros!
4. Applications to observations and simulations
Important Open Questions
•
What are the properties (Ts, N(HI), etc…) of HI in all phases?
Important Open Questions
•
What are the properties (Ts, N(HI), etc…) of HI in all phases?
Unconstrained!
Important Open Questions
How much HI exists in each phase?
CNM, WNM, and unstable fractions depend on input physics
(e.g. MacLow et al. 2005, Audit & Hennebelle 2005, Hill et al. 2012)
Strong turbulence
Pressure
Weak turbulence
Pressure
•
Density
Density
Audit & Hennebelle 2005
Comparing simulations with
observations is essential, and difficult!
Simula:ons*
Observa:ons*
Physical*quan::es*
Observed*spectra*
Synthe:c*spectra*
AGD()
Gaussian*components*
v
⇥i , i , v i ,
AGD()
Gaussian*components*
v
⇥i , i , v i ,
Better observational constraints…
21-SPONGE
21-cm Spectral line Observations of Neutral Gas with the (E)VLA
public.nrao.edu, NRAO/AUI/NSF
VLA
21-SPONGE
21-cm Spectral line Observations of Neutral Gas with the (E)VLA
• 58 sources: S>3 Jy, |b|>10
• Matching HI emission from Arecibo Observatory
• High-sensitivity HI absorption: σ𝜏 ~7 x 10-4
Arecibo
VLA
21-SPONGE
21-cm Spectral line Observations of Neutral Gas with the (E)VLA
• 58 sources: S>3 Jy, |b|>10
• Matching HI emission from Arecibo Observatory
Arecibo
• High-sensitivity HI absorption: σ𝜏 ~7 x 10-4
Superior sensitivity allows us to detect
unstable/WNM lines in absorption!
Millennium Survey
(Heiles & Troland 2003) 21-SPONGE Max
Max Ts ~600 K
Ts ~ 1500 K
Murray et al. 2015, in press
VLA
21-SPONGE
21-cm Spectral line Observations of Neutral Gas with the (E)VLA
• 58 sources: S>3 Jy, |b|>10
• Matching HI emission from Arecibo Observatory
Arecibo
• High-sensitivity HI absorption: σ𝜏 ~7 x 10-4
Unstable fraction ~ 20%
Millennium Survey = 48% (upper limits)
Millennium Survey
(Heiles & Troland 2003) 21-SPONGE Max
Max Ts ~600 K
Ts ~ 1500 K
Murray et al. 2015, in press
21-SPONGE
21-cm Spectral line Observations of Neutral Gas with the (E)VLA
Stacked absorption
Murray et al. 2014
21-SPONGE
21-cm Spectral line Observations of Neutral Gas with the (E)VLA
Stacked emission
Probes Lyα scattering excitation!
Stacked absorption
Murray et al. 2014
21-SPONGE limited to 10s of sources
•
WNM
GASS; McClure-Griffiths et al. 2009
HI Absorption Sightlines Visible to SKA-1
•
WNM
~106 spectra, 5 minutes each… ~10 years to analyze!
McClure-Griffiths et al. 2015
To the rescue…
AUTONOMOUS GAUSSIAN
DECOMPOSITION
Lindner et al. 2015 in press (arXiv:1409.2840)
• Automatic, efficient decomposition of 1D spectral data
into Gaussian functions
• Good initial guesses are chosen without human interaction
• Generally applicable to any data set!
Lindner et al. 2015 in press (arXiv:1409.2840)
Autonomous Gaussian Decomposition (AGD)
Provides optimized initial guesses for multi-component Gaussian fit
1. Computer vision using
st th
1 -4 numerical derivatives
2. Regularization provides
smooth derivatives
3. Supervised machinelearning optimizes
regularization parameters
to maximize guess accuracy
!
Guess criteria:
• Python package “GaussPy” (available upon publication)
Email: gausspy.software@gmail.com
21-SPONGE
AGD Guesses
AGD Best-Fit
Human Best-fit
AGD vs. Human (me) on 21-SPONGE
Number of
components
Noise in
resulting
best fit
Lindner et al. 2015 in press (arXiv:1409.2840)
Future Applications
Simula:ons*
Observa:ons*
Physical*quan::es*
Observed*spectra*
Synthe:c*spectra*
AGD()
Gaussian*components*
v
⇥i , i , v i ,
AGD()
Gaussian*components*
v
⇥i , i , v i ,
30*
AGD Results
AGD*results*
Gaussian*components*
•  Input:*10k*spectra*
• Blue: 104 synthetically•  Output:*20k*Gaussian*
observed HI absorption
components*(blue)*
lines (Kim et al. 2014)
•  Component*space*has*two*
•phases*(CNM,*WNM)*
Black: 21-SPONGE
VLA HI absorption lines
•  Simulated*“WNM”*
(Murray et al. 2015)
components*not*seen*in*
observa:ons*
Stacking%
result*
21ASPONGE*data:**Murray*et*al.*(2014b,*in*prep)*
34*
Simula:on*data:*Kim*et*al.*(2014)*
??
Matching Gaussians to Clouds in Simulations
Density
0
100
200 300 400
LOS distance
Optical Depth
Velocity
Matching Gaussians to Clouds in Simulations
First statisticallyrobust quantification
of cloud-component
correspondence!
Conclusions
•
21-SPONGE will constrain the uncertain mass distribution of HI as a
function of Ts, as the largest high-sensitivity HI absorption survey
•
Autonomous Gaussian Decomposition (AGD) provides optimized
initial guesses (and final fits) for multi-component Gaussian models
-
Python implementation: “GaussPy” available soon
Generally applicable to any spectral data set
•
AGD results are comparable to human decompositions
•
AGD enables objective, unbiased comparisons between
observations and simulations (e.g., 21-SPONGE vs. 3D hydro sims)
-
Confirms correspondence between physical clouds and Gaussians
Future: Ts completeness function, constraints on Lyα photon field and
more!