Auto-differentiable line-by-line spectral modeling of exoplanets/brown dwarfs using JAX. Read the docs ๐. In a nutshell, ExoJAX enables you to do a HMC-NUTS fitting using the latest database.
Voigt Profile โ๏ธ
from exojax.spec import voigt
nu=numpy.linspace(-10,10,100)
voigt(nu,1.0,2.0) #sigma_D=1.0, gamma_L=2.0
Cross Section using HITRAN/HITEMP/ExoMol โ๏ธ
from exojax.spec import AutoXS
nus=numpy.linspace(1900.0,2300.0,200000,dtype=numpy.float64) #wavenumber (cm-1)
autoxs=AutoXS(nus,"ExoMol","CO") #using ExoMol CO (12C-16O). HITRAN and HITEMP are also supported.
xsv=autoxs.xsection(1000.0,1.0) #cross section for 1000K, 1bar (cm2)
![](https://user-images.githubusercontent.com/15956904/111430765-2eedf180-873e-11eb-9740-9e1a313d590c.png)
Do you just want to plot the line strength?
ls=autoxs.linest(1000.0) #line strength for T=1000K
plt.plot(autoxs.mdb.nu_lines,ls,".")
autoxs.mdb is the moldb.MdbExomol class for molecular database. Here is a entrance to a deeper level. exojax is more flexible in the way it calculates the molecular lines. ๐ Go to the docs for the deeper level.
Emission Spectrum โ๏ธ
from exojax.spec.rtransfer import nugrid
from exojax.spec import AutoRT
nus,wav,res=nugrid(1900.0,2300.0,200000,"cm-1")
Parr=numpy.logspace(-8,2,100) #100 layers from 10^-8 bar to 10^2 bar
Tarr = 500.*(Parr/Parr[-1])**0.02
autort=AutoRT(nus,1.e5,2.33,Tarr,Parr) #g=1.e5 cm/s2, mmw=2.33
autort.addcia("H2-H2",0.74,0.74) #CIA, mmr(H)=0.74
autort.addcia("H2-He",0.74,0.25) #CIA, mmr(He)=0.25
autort.addmol("ExoMol","CO",0.01) #CO line, mmr(CO)=0.01
F=autort.rtrun()
![](https://user-images.githubusercontent.com/15956904/116488770-286ea000-a8ce-11eb-982d-7884b423592c.png)
Are you an observer?
nusobs=numpy.linspace(1900.0,2300.0,10000,dtype=numpy.float64) #observation wavenumber bin (cm-1)
F=autort.spectrum(nusobs,100000.0,20.0,0.0) #R=100000, vsini=10km/s, RV=0km/s
![](https://user-images.githubusercontent.com/15956904/116488769-273d7300-a8ce-11eb-8da1-661b23215c26.png)
If you want to customize the model, see here.
HMC-NUTS of Emission Spectra โ๏ธ
To fit a spectrum model to real data, you need to know a little more about exojax. See here.
![](https://github.com/HajimeKawahara/exojax/raw/master/documents/tutorials/results.png)
๐ฅฅ HMC-NUTS modeling of a brown dwarf, Luhman 16 A using exojax. See here for an example of the Bayes inference using the real spectrum.
Clouds โ
Only for brave users.HMC-NUTS of Transmission Spectra โ
Not supported yet.pip install exojax
or
python setup.py install
Note on installation w/ GPU support
๐ You need to install CUDA, NumPyro, JAX w/ NVIDIA GPU support, and cuDNN.
- NumPyro
exojax supports NumPyro >0.5.0, which enables the forward differentiation of HMC-NUTS. Please check the required JAX version by NumPyro. In May 2021, it seems the recent version of NumPyro requires jaxlib>=0.1.62 (see setup.py of NumPyro for instance).
- JAX
Check you cuda version:
nvcc -V
Install such as
pip install --upgrade "jax[cuda111]" -f https://storage.googleapis.com/jax-releases/jax_releases.html
This is the case for cuda11.1 to 11.4. Visit here for the details.
- cuDNN
For instance, get .deb from NVIDIA and install such as
sudo dpkg -i libcudnn8_8.2.0.53-1+cuda11.3_amd64.deb
cuDNN is used for to compute the astronomical/instrumental response for the large number of wave number grid (exojax.spec.response). Otherwise, we do not use it.
- Kawahara, Kawashima, Masuda, Crossfield, Pannier, van den Bekerom (2021) accepted by ApJS: arXiv:2105.14782
๐ Copyright 2020-2021 Hajime Kawahara and contributors. exojax is publicly available under the MIT license. Under development since Dec. 2020.