Stellar Parameter Estimation for Gaia

1
Stellar Parameter Estimation for Gaia
Carola Tiede
2
GSP-phot
BP

• Estimates APs (log(Teff), log g, Av, Fe/H) for
  all objects with P(YSTAR) gt threshold
• Based on photometry and parallax
• Uses Machine Learning techniques (currently SVMs)
• Estimated APs ( uncertainties) also serve as
  inputs to other parts of processing pipeline

RP
3
Support Vector Machine
ltw,xgtb0
4

• Supervised method
• Define linear hyperplane with maximal margin
• Use kernel trick to transform into higher dim.
  feature space and compute linear hyperplane there
• Multiclass problem
• K models, each of type 1 class vs. combined K-1
  classes
• K(K-1)/2 models, each with 1 class vs. 1 class
  (pairwise coupling)
• Regression
• Same as classification but searches for a
  function. Epsilon has to be defined as noise
  tolerance

Support Vector Machine
ltw,xgtb0
5
Sensitivity Analysis

• Sensitivity Apportion variation in output
  variables to variation in input sources.
• Model is sensitive to an input if changing that
  input variable changes the model output
• Computation
• Input sources Each bin of RP,BP and parallax
• Output A certain AP
• Keep one AP fixed and vary for each value k of
  the AP all others, compute variation s(i)APk for
  all values k and for all bins (RP,BP, parallax) i
  separately
• Compute maximal variation in i
• Normalize over all bins (RP, BP and parallax) for
  each AP

6
Sensitivity Analysis
AV

• Data, either
• Normalize or
• Normalize and re-weight by variations
• Grid search over SVM parameters for each AP and
  for both settings

Teff
7
Who am I ???

• PostDoc in DLR Gaia group
• General Work
• Supervised Learning, Parameter Estimation
• Practical Experience
• Study of Geodesy, PhD in Physical Geodesy
• Physical/Astronomical Geodesy, Geoscience
• Data Analysis Optimization, Modeling, Data
  Mining