PRObabilistic Value-Added Bright Galaxy Survey (PROVABGS)

The PROVABGS catalog will provide measurements of galaxy properties, such as stellar mass, star formation rate, stellar metallicity, and stellar age for >10 million galaxies of the DESI Bright Galaxy Survey. Full posterior distributions of these galaxy properties will be inferred using state-of-the-art Bayesian spectral energy distribution (SED) modeling of DESI spectroscopy and photometry.

The provabgs Python package provides

• a state-of-the-art stellar population synthesis (SPS) model based on non-parametric prescription for star formation history, a metallicity history that varies over the age of the galaxy, and a flexible dust prescription.
• a neural network emulator (Kwon et al. in prep) for the SPS model that is >100x faster than the original SPS model and enables accelerated inference. Full posteriors of the 12 SPS parameters can be derived in ~10 minutes. The emulator is currently designed for galaxies from 0 < z < 0.6.
• a Bayesian inference pipeline based on the zeus ensemble slice Markov Chain Monte Carlo (MCMC) sample.

For additional details see documentation and Hahn et al (2022)

Installation

Clone the repo first:

git clone https://github.com/changhoonhahn/provabgs.git
cd provabgs

Option A: just the neural network emulators (recommended for most users)

This is all you need to evaluate SEDs and run inference with the emulator -- no compiled dependencies (no fsps, no torch), and it's what the example notebook uses.

pip install -e .

To also use the Bayesian MCMC fitting pipeline (provabgs.infer), add the mcmc extra:

pip install -e ".[mcmc]"

Option B: everything, including the original (non-emulator) FSPS model

python-fsps requires compiling FSPS's Fortran code, which is where most install trouble comes from. Installing via conda/mamba pulls prebuilt binaries and avoids that entirely:

conda env create -f environment.yml
conda activate provabgs

If you'd rather use plain pip, see the python-fsps documentation for build instructions, then run pip install -e ".[full]".

NERSC / cluster notes

If you're using provabgs on NERSC, you may run into an import error with libgfortran.so.5 when using fsps. The following has resolved it before:

module unload PrgEnv-intel
module load PrgEnv-gnu

Example

Checkout the nb/example.ipybn notebook for an example on conducting Bayesian SED modeling on galaxy spectra using provabgs. It requires less than 10 lines of code and about 10 minutes!

If you're interested in conducting Bayesian SED modeling on DESI spectra in particular, check out the nb/tutorial_desispec.ipynb notebook.

Development

If you're contributing notebooks or code, please install pre-commit so exploratory/debugging notebooks get their outputs stripped before they're committed:

pip install pre-commit
pre-commit install

This keeps large embedded images/data out of git history (the two tutorial notebooks under nb/ are exempt so their rendered outputs stay visible). Please also avoid committing large binary files (trained model weights, data files) directly -- attach them to a GitHub release or a Zenodo record and download them instead.

Team

• ChangHoon Hahn (Princeton)
• Rita Tojeiro (St Andrews)
• Justin Alsing (Stockholm)
• James Kyubin Kwon (Berkeley)

Contact

If you have any questions or need help using the package, please raise a github issue or contact me at changhoon.hahn@princeton.edu