Automated spectroscopic pipeline to determine radial velocities and stellar atmospheric parameters
Author: Sergey Koposov skoposov AT ed DOT ac _DOT uk, University of Edinburgh
Citation: If you use it, please cite https://ui.adsabs.harvard.edu/abs/2019ascl.soft07013K/abstract and https://ui.adsabs.harvard.edu/abs/2011ApJ...736..146K/abstract
If something doesn't quite work, I'm happy to help
Dependencies: numpy, scipy, astropy, pyyaml, matplotlib, numdifftools, pandas
Optional: torch, scikit-learn
The code here can perform radial velocity fitting and general spectral fitting with interpolated templates to any spectroscopic data.
The library is primarily used with the PHOENIX library, but you can also use your own library of synthetic spectra (see below)
To get started you will need to
- Install rvspecfit
- Download PHOENIX library
- Create PHOENIX processed files for your spectral configuration. This involves convolving spectra to resolution of your instrument and producing interpolators.
- Create a configuration file
- Run the code
To install you can just do can just do
pip install https://github.com/segasai/rvspecfit/archive/master.zip
which will install the latest version
The library is avialable here https://phoenix.astro.physik.uni-goettingen.de/v2.0/HiResFITS/
(Make sure you get the v2.0 one!). If you want to download the whole thing you will likely need a command like
wget -r -np -l 10 https://phoenix.astro.physik.uni-goettingen.de/data/v2.0/HiResFITS/
The preparation requires several steps (you can find several examples of these steps in the surveys folder, i.e surveys/sdss/make_sdss.sh surveys/desi/make_desi.sh which prepare rvspecfit for processing SDSS or DESI spectra).
The commands below all support --help option, so you can always run that to get more help on the meaning of different options.
- Creating a PHOENIX file sqlite database that will be used in the processing (you only need to do this step once)
This is done with
$ rvs_read_grid --prefix $PATH/PHOENIX/v2.0/HiResFITS/PHOENIX-ACES-AGSS-COND-2011/ --templdb files.db
Some templates may be marked as bad there if you wish.
- Making interpolated spectra for your spectral configuration. If your instrument has multiple arms (i.e. blur/red), you will need to run this once for each arm.
$ rvs_make_interpol --setup myconf --lambda0 4000 --lambda1 5000 \
--resol_func '1000+2*x' --step 0.5 --templdb ${PREFIX}/files.db \
--oprefix ${PREFIX}/ --templprefix $TEMPLPREF --wavefile $PATH/HiResFITS/WAVE_PHOENIX-ACES-AGSS-COND-2011.fits \
--air --revision=v2020x
That will create the spectral configuration called myconf for spectra with wavelength range of 4000 to 5000, step 0.5 Angstrom and resolution being 1000+2*x (where x is wavelength in Angstroms). It also requires paths to the files.db database created at previous step as well as the file with the wavelength grid of PHOENIX library which is distributed with PHOENIX. You can also choose to do things in air or vacuum. TEMPLPREF is the path to the PHOENIX library.
This step can take up to an hour, depending on your machine.
- Making the n-d interpolator. That requires the path to the files created at previous step.
$ rvs_make_nd --prefix ${PREFIX}/ --setup myconf --revision=v2020x
This will create either a polylinear interpolator or Linear interpolator within Voronoi cells.
- Neural network interpolator An alternative to polylinear interpolator is a neural network based interpolator. This requires running
$ rvs_train_nn_interpolator --dir ${PREFIX}/ --setup myconf
instead of rvs_make_nd. When doing the fitting, you may also set this environmental variable RVS_NN_DEVICE=cpu to force the code to run on cpu.
- Making the cross-correlation files
$ rvs_make_ccf --setup myconf --lambda0 4000 --lambda1 5000 \
--every 30 --vsinis 0,10,300 --prefix ${PREFIX}/ --oprefix=${PREFIX} \
--step 0.5 --revision=v2020x
That creates a list of Fourier transformed templates for the CCF. IN this case this list will have every 30-th spectra from the database. It also uses a list of Vsini's of 0,10,300 when creating CCF templates.
- The very last step is creating of the configuration file config.yaml with contents like this
template_lib : '/home/username/template_data/'
min_vel: -1500
max_vel: 1500
min_vel_step: 0.2
vel_step0: 5
min_vsini: 0.01
max_vsini: 500
second_minimizer: 1where the only important parameter is the template_lib path which points at the location where all the processed template files are.
After that you should be able to use rvspecfit.
You need just a few commands
from rvspecfit import fitter_ccf, vel_fit, spec_fit, utils
config=utils.read_config('config.yaml') # optional
# you can create a configuration file with various options
# let's assume we have data stored in 1d arrays in a table
# with wavelength being in Angstrom
# espec being a vector of standard deviations
tab=atpy.Table().read('spec.fits')
wavelength = tab['wavelength']
spec = tab['spec']
espec = tab['espec']
# This constructs the specData object from wavelength, spectrum and error
# spectrum arrays. The rvspecfit works on arrays of SpecData's which may
# represent multiple exposures or multiple spectral configurations
# Here we just have one spectrum from the spectral configuration "myconf"
specdata = [spec_fit.SpecData('myconf',
wavelength,
spec,
espec)
]
# this tries to get a sensible guess for the stellar parameters/velocity
res = fitter_ccf.fit(specdata, config)
paramDict0 = res['best_par']
fixParam = []
if res['best_vsini'] is not None:
paramDict0['vsini'] = res['best_vsini']
options = {'npoly':10}
# this does the actual fitting performing the maximum likelihood fitting of the data
res1 = vel_fit.process(specdata,
paramDict0,
fixParam=fixParam,
config=config,
options=options)
print(res1)If your have multiple spectra, from multiple instrument arms, from multiple exposures, fitting them is easy.
If you have multiple instrument arms, you will need to prepare the spectral interpolators for all the arms and then combine the specdata objects from multiple arms in the list
sd_blue = spec_fit.SpecData('blue',
wavelength1,
spec1,
espec1,
badmask=badmask1)
sd_red = spec_fit.SpecData('red',
wavelength2,
spec2,
espec2,
badmask=badmask2)
res1 = vel_fit.process([sd_blue, sd_red],
paramDict0,
fixParam=fixParam,
config=config,
options=options)If you just have multiple exposures, you can obviously use the same spectral configuration for them and then combine the specdata objects for those
One advantage of rvspecfit is that you can use it as part of larger inference framework. I.e. you can add the spectral likelihood to other likelihood terms.
To do that you just need this call the get_chisq() function that will return the -2*log(L) (or chi-square) of your spectral dataset given the model parameters
loglike = 0
vel=300
atm_params = [4100,4, -2.5, 0.6] # 'teff', 'logg', 'feh', 'alpha'
spec_chisq = spec_fit.get_chisq(specdata,
vel,
atm_params,
None,
config=config, options = dict(npoly=10))
loglike += -0.5 * spec_chisqThat will compute the chi-square of the spectrum for a given template parameters and radial velocity. That will also use the 10 order polynomial as multiplicative continuum . The resulting chi-square can be then multiplied by (-0.5) and added to your log-likelihood if needed.
To run on DESI data use rvs_desi_fit or rvs_weave_fit
By default rvspecfit uses the multi-dimensional linear interpolation with Delaunay triangulation. It also supports the multilinear interpolation if the input grid is organized as a true n-D grid without gaps. The alternative interpolation method is Neural Network based.
You can substitute PHOENIX by your own library of synthetic spectra. The only thing needed there is a wavelength FITS file and a collection of FITS files with spectra and stellar parameter information in the headers (i.e. LOGG, TEFF keywords etc). You can then use rvs_read_grid to create a database of spectra and then follow the same steps as with PHOENIX