Exo Skryer: A JAX-accelerated sub-stellar atmospheric retrieval framework

TL;DR

A new sub-stellar atmosphere retrieval modelling framework, Exo Skryer, that utilises the JAX library for Python to enable scalable, computationally efficient forward modelling as well as posterior sampling, and a new method to directly retrieve the real and imaginary optical constants of suspected aerosol infrared absorption features.

Abstract

Contemporary exoplanet and brown dwarf atmospheric research relies heavily on retrieval frameworks to recover thermal and chemical properties and perform model comparison in an observational data-driven approach. However, the computational effort required for retrieval modelling has rapidly increased, driven by JWST data that covers large spectral intervals at moderate spectral resolutions, and ground-based, high-resolution spectroscopy. To help tackle the computational burden faced by contemporary retrieval requirements, I present a new sub-stellar atmosphere retrieval modelling framework, Exo Skryer, that utilises the JAX library for Python to enable scalable, computationally efficient forward modelling as well as posterior sampling. I present example retrievals for pre- and current JWST era observations for both transmission and emission spectra, finding consistent results to previous retrieval modelling efforts, apart from a WASP-107b test case. In addition, I present a new method to directly retrieve the real and imaginary optical constants (n, k) of suspected aerosol infrared absorption features. Due to its computational expediency, Exo Skryer will be highly suited for future demanding retrieval efforts that incorporate more spatial dimensionality, complex forward models and high-dimensional parameter sets.

Figures (10)

• Figure 1: Example of the modified Milne T-p profile scheme for brown dwarf retrievals in Sect. \ref{['sec:Milne']} (purple, solid), compared to the original Milne profile (orange, dotted), classic Hopf solution (green, solid) and a chemical equilibrium ATMO profile (grey, dashed) from Phillips_2020. The bulk parameters are $T_{\rm int}$ = 1000 K, $\log_{10}$ g = 4.5, with the modified Milne profile parameters printed inside the figure.
• Figure 2: Example of the modified Guillot T-p profile scheme for hot Jupiter retrievals in Sect. \ref{['sec:Guillot']} (purple, solid), compared to the original Guillot_2010 profile (blue, dotted). The bulk parameters are $T_{\rm int}$ = 100 K, $T_{\rm eq}$ = 1000 K, $\log_{10}$ g = 3.0, with the modified Guillot profile parameters printed inside the figure.
• Figure 3: Equal weight posterior samples and kernel density estimation for the HD 209458b pre-JWST transmission spectra. The median and 1$\sigma$ confidence intervals are shown as the dashed and dotted lines respectively. The plot in the top right shows the best fit median spectra (orange) with 1$\sigma$ (shaded) region to the observed HD 209458b pre-JWST transmission spectra data (blue). The high-resolution median spectrum is in purple.
• Figure 4: Equal weight posterior samples and kernel density estimation for the WASP-107b JWST transmission spectra. The median and 1$\sigma$ confidence intervals are shown as the dashed and dotted lines respectively. The top left plots show the observational data (blue points), median model with 1$\sigma$ ranges (orange) and high-resolution median model (purple).
• Figure 5: Equal weight posterior samples and kernel density estimation for the WASP-17b HST and JWST combined transmission spectra retrieval from Louie_2025. The median and 1$\sigma$ confidence intervals are shown as the dashed and dotted lines respectively. The plots in the top right show the best fit median spectra (orange) with 1$\sigma$ (shaded) region to the observed transmission spectra data (blue), taken from Louie_2025, for the full spectral range (top right) and MIRI LRS wavelength range (middle right). The high-resolution median spectrum is in purple.