browndwarf / jammer Goto Github PK

Here are my thoughts on priors:

uniform in logg, uniform in T_eff

These are the parameters of interest.

Fraction of a pixel for v_z.

This is effectively a wavelength calibration consideration, since the spectra are insufficiently low-spectral resolution to assess v_z for any plausible range.

vsini rebranded to sigma_R

The uncertainty in the spectral resolution is much larger than the vsini. They have slightly different kernels, so this requires a tweak to the update_theta function.

flat in logOmega (for now)

We have distances estimates and decent priors for radii for these objects, so we could assign an informative prior on Omega. For now, let's just allow the values to float and see what we get. We could repeat the experiment later and see if we get a consistent estimate for effective temperature.

maximum deviation for Cheb calibration polynomials

Already in place, this seems to make the most sense. I lowered the peak-to-valley to 2%, which is about equal to the maximum S/N per pixel.

+/-5% in sigmaAmp

We assume we have estimated the mean sigma correctly to within 5%, which is probably accurate. I did impute some sigma values (they had zero values before), but this level of uncertainty seems to be fine.

logAmp?

getting this wrong caused major errors before. We have put a cap of logAmp < -17 from trial and error, and by-eye checking.

GP length scale?

Ideally this should be comparable to the slit function. In practice the length scale likes to smooth out correlated residuals from errors in the broad band brown dwarf spectral shape. I assigned 5000 - 25000.

This source has better and more abundant comparison analysis

This experiment will help us get a handle of what's going on.

mimic the Gully-Santiago et al. 2017 style

Here are my thoughts on priors:

uniform in logg, uniform in T_eff

These are the parameters of interest, though brown dwarfs could have strong priors based on radius and distance arguments... worth looking into.

v_z and vsini set through "empirical Bayes"

Well, kinda...

flat in logOmega (for now)

IGRINS data aren't flux calibrated.

maximum deviation for Cheb calibration polynomials

This is a huge issue. On one hand, could be unlimited to basically filter out any low frequency structure. On the other hand, it could be strict (<1% peak to valley) to force a fit to the spectral shape. Finally, it could be somewhere in between. I recommend the inbetween, with high limit: peak-to-valley to 10%.

+/-5% in sigmaAmp

We assume we have estimated the mean sigma correctly to within 5%, which is probably accurate. The real problem is the bad pixels. These should just be thrown out!!

logAmp?

We need to get this right and it's a bit involved and bespoke for each order.

GP length scale?

We need to get this right and it should be close to the same for each order.

• what are the range in parameters
• how to read the files in
• will the grids be sufficient for analysis of T dwarfs?

We want to know the stats on different systems.

devel queue on NASA Pleiades with n_threads = n_cpus = 28 on the Broadwell system:

mgullysa at r633i2n8 in ~/GitHub/jammer/sf/2M0136/m112/output/marley_grid/run02
$ time $jammer/code/star_marley_beta.py --samples=200 --incremental_save=10
keeping grid as is

2017 Aug 17, 7:01 PM: 9/200 = 4.5%
...
2017 Aug 17, 7:07 PM: 199/200 = 99.5%
The end.

real	6m52.675s
user	112m18.393s
sys	1m0.155s

Only half the processes were used at one time...

awesome! :)

We eventually want to scale up the spectral inference to more sources with supercomputing resources. For now, let's focus on a few good demo sources, with good priors, good initial guesses, good noise models, and reproducible workflows.

Tasks:

• Make sure all order have the same user-defined priors
• Spot check the residuals for all initial guesses, iteratively refine
• git commit all config.yaml and s0_phi.json files
• Run token MCMC of ~3 samples to make sure everything compiles
• pull all changes to NAS HECC resources
• git commit the conda environment.yml file and .envrc file for reproducibility
• Pick system for deployment-- pleiades or other
• Deploy to Pleiades

We already tried with 4, let's do many more

It will merely create bias

Outline:

• Run Starfish with Sonora bobcat models on SpeX spectrum (prism library, etc)
• Run on 2M0136 and 2M0314+1603.
• For 2M0314+1603 use Teff in [1900 K to 2400 K]
• For 2M0136 use Teff in [900 K to 1600K ]
• Run on Pleiades (use Mark's account)

TBD:

• Decide whether to fix priors, tentatively fix priors on vsini and v_z when available, not Teff, logg

This change will facilitate modularity.

basically automate everything

The full-sampling ones are available in the SpeX Prism Library

need to do this

this will be very dope if possible

• Convert the units

Mark gave me all the new Sonora Bobcat model grid files. I should use those for the next Starfish runs.

show where we are looking

I will perform the same flux scaling, and we can probably get away with truncating the wavelength range to longward of 1 micron. The S/N is incredibly low below that, and no models predict measurable flux there anyways...