search

HajimeKawahara / exojax

🐈 Automatic differentiable spectrum modeling of exoplanets/brown dwarfs using JAX, compatible with NumPyro and JAXopt
http://secondearths.sakura.ne.jp/exojax/
MIT License
45 stars 14 forks source link

DIT with no sigma_D sampling in a evenly-spaced log grid (MODIT) #97

Closed HajimeKawahara closed 3 years ago

HajimeKawahara commented 3 years ago

DIT for a evenly-spaced logarithm wavenumber grid (ESLOG) is straightforward for the framework of exojax, considering the instrument response, RV, and a rotation broadening.

Based on the snippet and with the great help by @dcmvdbekerom, I made a modified version of DIT in addit prior to the implementation in exojax. This modified version (herafter tagged by MODIT) assumes

  1. Input wavenumber grid evenly spaced in logarithm
  2. Common Doppler width normalized by line center (i.e. common T for a single isotope)

The second assumption can remove the dimension of Doppler width from a DIT grid, which reduces the computational complexity.

I will implement these features into exojax prior to checking auto-diff and/or HMC works.

HajimeKawahara commented 3 years ago

The branch is features/logdit.

HajimeKawahara commented 3 years ago

MODIT looks better than DIT+interpolation to the log grid. It is likely the interpolation introduces errors, or a common T in MODIT improves the error (not sure though).

dit2 dit3

HajimeKawahara commented 3 years ago

Emission spectrum and the response matrix worked for MODIT. (1:raw spectrum,2:difference btw MODIT and direct, 3: response applied, 4: difference)

dit_emis

HajimeKawahara commented 3 years ago

The raw spectrum and the response-applied one for CH4 using MODIT (N~140,000).

moditCH4

HajimeKawahara commented 3 years ago

I should check the optimal zero padding (note).

HajimeKawahara commented 3 years ago

closed w/ merge of #100

HajimeKawahara commented 2 years ago

156

Comparison for the Luhman16A model

compml