PriorGen

PriorGen is a Python 3 package designed to use the Hayes et al. (2019) method of using machine learning classification to generate informed priors which can be used in Bayesian retrieval. PriorGen users can either use the ClassifiedRetriever to run Baysian retrieval on their data, or create a Classifier and incorporate it into their own retrieval algorithms.

In order to use PriorGen effectively, you will require a trusted forward model which takes continuous parameter values as inputs and produces some single dimensioned output F(x).

Citing Priorgen

If you find PriorGen useful in your work, please cite the paper which outlines the methodology: https://ui.adsabs.harvard.edu/abs/2020MNRAS.494.4492H/abstract. If you are using BibTeX, you can copy the following into your .bib file:

@ARTICLE{2020MNRAS.494.4492H,
       author = {{Hayes}, J.~J.~C. and {Kerins}, E. and {Awiphan}, S. and {McDonald}, I. and {Morgan}, J.~S. and {Chuanraksasat}, P. and {Komonjinda}, S. and {Sanguansak}, N. and {Kittara}, P. and {SPEARNet Collaboration}},
        title = "{Optimizing exoplanet atmosphere retrieval using unsupervised machine-learning classification}",
      journal = {\mnras},
     keywords = {methods: data analysis, methods: statistical, planets and satellites: atmospheres, Astrophysics - Earth and Planetary Astrophysics, Astrophysics - Instrumentation and Methods for Astrophysics},
         year = 2020,
        month = may,
       volume = {494},
       number = {3},
        pages = {4492-4508},
          doi = {10.1093/mnras/staa978},
archivePrefix = {arXiv},
       eprint = {1909.00718},
 primaryClass = {astro-ph.EP},
       adsurl = {https://ui.adsabs.harvard.edu/abs/2020MNRAS.494.4492H},
      adsnote = {Provided by the SAO/NASA Astrophysics Data System}
}

Contents

• Requirements

• The Classifier

• The ClassifiedRetriever

Requirements

Along with an installation of Python 3 (with the standard Conda distribution packages), priorgen requires the following packages to be installed:

• scikit-learn - For handling the machine learning classification.

• dynesty - For nested sampling retrieval.

The Classifer

In PriorGen, the Classifer is the object which deals with classification of a set of training data into different ObservablesClass objects, and handles classification of a previously unseen set of data.

In order to create a Classifer, you must provide it with two arrays - the parameters and the observables.

The observables are what the ML classifier is trained on, and are a single-dimensioned function F(x). The parameters should be provided as M-dimensional sets of values which were used to generate each observable using a trusted forward model.

The set of observables are then run through a principal component analysis (PCA) routine to perform a dimensionality reduction, before a k-means clustering routine is run to divide the observables into classes. The physical parameters for each of these classes are binned into histograms, and these distributions then act as a marginalised prior for each of the parameters in Bayesian retrieval.

The ObservablesClass is a Python object containing information on the parameter distributions of each of the classes found by the k-means clustering. It also contains a function which can convert a unitless value in range [0, 1] to a physical value, which is of use in nested sampling routines.

The ClassifiedRetriever

Whilst a user can construct a bespoke retrieval algorithm which uses a Classifier to give informed priors, PriorGen provides the ClassifiedRetriever object which can run basic retrieval on a set of data.

The ClassifiedRetriever must be initialised in the same way as the Classifier, and generates a Classifier itself. The user must then create a function lnprob which generates a log probability of a set of parameters fitting the data which is being fitted. This can then be passed to the retrieval functions within the ClassifiedRetriever and retrieval will then be run using the member class parameter distributions as informed priors.

Currently only nested sampling retrieval is implemented using dynesty, which can be accessed through ClassifiedRetriever.run_dynesty(). MCMC fitting may be provided in later versions but is not currently available.