Modelling SEDs with Artificial Neural Networks

1
Modelling SEDs with Artificial Neural Networks
Laura Silva (INAF/Trieste) GianLuigi Granato
(INAF/Trieste), Andrew Schurer (INAF/Padova),
Cesario Almeida, Carlton Baugh, Cedric Lacey,
Carlos Frenk (ICC/Durham)

• Outline
• SED modelling- approaches vs aims
• GRASIL and application to SAM
• Modelling SED with ANN

2
Multi-? SED modelling approaches vs aims
Stellar pop. synthesis
SFR(t)Mgas(t),Z(t) exponentials, chemical evol.
or galaxy formation models
UV/optical attenuation and IR emission
Semi-empirical attenuation curve for LIR IR
shape Pros non time consuming analysis of large
data sets . Cons not great predictive power
Theoretical Explicit computation of radiative
transfer and dust emission Pros broader
interpretative/predictive power Cons time
consuming
3
Modelling UV to radio SEDs with
GRA(phite)SIL(icate)
1) Realistic and flexible SED modelling

• Stars and dust in a bulge (King profile) disk
  (double exponential)

• 3 dusty environments dense (star forming
  Molecular Clouds), diffuse (cirrus) (?
  clumping of stars and dust), dusty envelopes of
  AGB stars

• Stars are born within MCs and gradually escape as
  a function of their age ? age-dependent
  extinction

Star- forming MCs

• Dust big grains, very small grains and PAHs.
  Emission is appropriately computed for each
  component

Extincted stars

• UV-to radio SEDs (continuum nebular lines)

Diffuse dust
2) Reasonable computing time

• Presence of symmetries

• Radiative transfer exactly solved for opt thick
  MCs, with approximation in the cirrus (real
  bottle-neck)

• No Monte Carlo

4
Computing SEDs in Semi-Analytical galaxy
formation Models

• SAM DM with gravity-only N-body or MC, baryons
  with analytical recipes compare with widest
  range of observed galaxy properties
• Outputs simulated catalogues of galaxies at
  different redshift slices SFR(t), Mgas(t), Z(t),
  morphology, scale radii for stars dust

• Associate to each mock galaxy its real SED
  butcomplexities in treating radiative effects -
  unknown dust properties - computing time
  fundamental issue for cosmological volumes
• Semi-empirical treatment fix ?v (?L or ?f(Mgas,
  Z)) ? dependence uniform distrib. of stars
  and dust in a 1D slab

• SAMs with theoretical SED GALFORMGRASIL
  (Granato00, Silva01, Baugh05,
  Lacey08)MORGANAGRASIL (Monaco07, Fontanot07,
  08)
• Anti-hierar.BarionicCollapseGRASIL(Granato04,
  Silva05,Lapi06)
  

5
Different treatments predict different SED for
the same SFR(t)
SAM GRASIL
SAM CF00 slab
SF histories from the Semi-Analytical Model for
galaxy formation MORGANA SED by GRASIL
(colored) Empirical attenuation curve with CF
slab (hatched)
Fontanot, Somerville, Silva08
6
Net attenuation A(90)-A(0) vs Mstar
SAM GRASIL
SAM CF00 slab
SF histories from the Semi-Analytical Model for
galaxy formation MORGANA SED by GRASIL
(colored) Empirical attenuation curve with CF
slab (hatched)
7
Different treatments predict different SED for
the same SFR(t)
SAMGRASIL
SAM templates CharyElbaz01, Lagache04,
Devriendt99
SF histories from SAM MORGANA SED by
GRASIL (black) Templates (color)
Fontanot, Somerville, Silva08
8
GALFORM GRASIL
(Granato00, Silva01, Baugh05, Lacey08)
Local universe
K-band
0.2 ?m
B-band
60 ?m
Revised model reproduce multi-? LFs and counts/
z-distr with top-heavy IMF in starbursts
But high-z universe
850?m Old model
850?m New model
9
Improving the computing time Modelling SEDs with
Artificial Neural Neworks
Spectral variance for a GALFORM GRASIL catalogue

• SEDs complex, non-linear, high dimensionality
  and large variance functions of some galaxy
  properties

gtANN Mathematical algorithms for data analysis,
introduced to replicate the brain behavior -
learn from examples
ANN is a black box that is trained to predict the
SED from controlling parameters using a suitable
precomputed training set (many couples
input-output)
10
Output layer SED, one unit for each L(?)
Input layer parameters determining the SED
wjk
nj?wjkik ojf(nj)
Hidden layers black box!
Propagation rule the output from each unit is
weighted and summed to form the input for the
upper layer units nj?wjkik The new output is
ojf(nj) , fnon linear function
Learning the ANN is trained with a given target-
weights are adjusted to best approximate a given
set of inputs/outputs
11
ANN SED 2 methods

• Universal and very fast (Silva08) input
  physical quantities determining the SED of MCs
  and Cirrus one single trained net

• Less universal and super-fast (Almeida08)
  input galaxy properties re-train the net for
  different realizations

• MCs Optical depth , R/Rsubl.
• Cirrus Ldust, Mdust, Polar Equatorial opt
  depth, R/Rdust, z/R, zdust/Rdust, Hardness
  of the rad. field
• ANN mode implemented in GRASIL compute
  extinction and predict IR SEDs with separately
  trained ANN for MCs and Cirrus - 1 sec. -gt
  large cosmological volumes

• Mstar, Zstar, Zgas, Lbol, vcirc, R1/2 (bulge
  disc), ?V, Mburst, ?tlast burst
• Each simulated catalogue from a SAM requires a
  trained net
• ltlt 1 sec -gt exploit the whole Millennium
  Simulation

12
Examples single objects
M82
ARP220
13
Examples models extracted from ABC SAM (G04, S05)
FULL/ANN Tot black/red MC dark
/light green Cirr blue/cyan
m12.00zv2.50 _at_ 0.25 Gyr
m13.00zv3.50 _at_ 0.25 Gry
m13.20zv3.50 _at_ 0.1 Gyr
m13.20zv3.50 _at_ 0.5 Gyr
14
Examples models extracted from a GALFORMGRASIL
catalogue
Almeida et al.
15
and one catastrophe .work in
progress..
m12.00zv2.50 _at_ 0.1 Gyr
Improving the reconstructed SED by splitting the
output neurons
16
100(1-Lpredicted/Loriginal) vs Loriginal
24 ?m
B-band
GALFORMGRASIL catalogue gt70 with error lt 10
- MIR and submm have larger variance
850 ?m
Almeida et al.
17
Colours for z0 GALFORMGRASIL catalogue
Almeida et al.
18
ABC SAM Galaxy counts
FULL red ANN blue
24 ?m
850 ?m
19
24 ?m z0.5
0.17 ?m z3
850 ?m z2
GALFORMGRASIL Luminosity Functions ______
original - - - - - - recontructed
Almeida et al.
20
Summary

• Multi-wavelength modelling as a tool to
  quantitatively deconvolve/ interpret observations
  make predictions/ constrain galaxy formation
  models
• Different treatments predict different SEDs for
  the same SFR(t)-gt necessity of a reliable
  computation of the SED for proper interpretations
  of observations and predictions of galaxy
  formation models
• The treatment of dust reprocessing of UV/optical
  in the IR requires a proper computation time
  cosuming for some applications
• For large cosmological applications promising
  solution with ANN