How do we optimally sample model grids of exoplanet spectra?

TL;DR

This paper compares different sampling methods for atmospheric model grids in exoplanet spectra analysis, showing that random and Latin hypercube sampling are more efficient than linear sampling, and highlighting potential biases in traditional interpolation methods.

Contribution

It introduces and evaluates alternative sampling techniques for model grids, demonstrating their advantages over linear sampling in exoplanet atmospheric characterization.

Findings

Random and Latin hypercube sampling outperform linear sampling in high-dimensional models.

Linear interpolation can bias posterior distributions, especially for non-linear parameters.

Information content analysis identifies key spectral features for parameter detection.

Abstract

The construction and implementation of atmospheric model grids is a popular tool in exoplanet characterisation. These typically vary a number of parameters linearly, containing one model for every combination of parameter values. Here we investigate alternative methods of sampling parameters, including random sampling and Latin hypercube (LH) sampling, and how these compare to linearly sampled grids. We use a random forest to analyse the performance of these grids for two different models, as well as investigate the information content of the particular model grid from Goyal et al. 2019. We also use nested-sampling to implement mock atmospheric retrievals on simulated JWST transmission spectra by interpolating on linearly sampled model grids. Our results show that random or LH sampling out-performs linear sampling in parameter predictability for our higher dimensional models, requiring fewer models in the grid, and thus allowing for more computationally intensive forward models to be used. We also find that using a traditional retrieval with interpolation on a linear grid can produce biased posterior distributions, especially for parameters with non-linear effects on the spectrum. In particular, we advise caution when performing linear interpolation on the C/O ratio, cloud properties, and metallicity. Finally, we find that the information content analysis of the grid from Goyal et al. 2019 is able to highlight key areas of the spectra where the presence or absence of certain molecules can be detected, providing good indicators for parameters such as temperature and C/O ratio.