Pre-computed aerosol extinction, scattering and asymmetry grids for scalable atmospheric retrievals

TL;DR

This work improves the computational efficiency and scaling behavior of aerosol models in atmospheric retrievals, enabling in-depth studies including multiple condensate species within practical time scales and enabling more complex retrievals as well as broader population studies without increasing the overall error budget.

Abstract

The unprecedented wavelength coverage and sensitivity of the James Webb Space Telescope (JWST) permits to measure the absorption features of a wide range of condensate species from Silicates to Titan tholins. Atmospheric retrievals are uniquely suited to analyse these datasets and characterize the aerosols present in exoplanet atmospheres. However, including the optical properties of condensed particles within retrieval frameworks remains computationally expensive, limiting our ability to fully exploit JWST observations. In this work, we improve the computational efficiency and scaling behavior of aerosol models in atmospheric retrievals, enabling in-depth studies including multiple condensate species within practical time scales. Rather than computing the aerosol Mie coefficients for each sampled model, we pre-compute extinction efficiency (Qext), scattering efficiency (Qscat) and asymmetry parameter (g) grids for seven condensate species relevant in exoplanet atmospheres (Mg2SiO4 amorph sol - gel, MgSiO3 amorph glass, MgSiO3 amorph sol - gel, SiO2 alpha, SiO2 amorph, SiO and Titan tholins). The pre-computed Qext grids significantly reduce computation time between 1.4 and 17 times with negligible differences on the retrieved parameters. They also scale effortlessly with the number of aerosol species while maintaining the accuracy of cloud models. Thereby enabling more complex retrievals as well as broader population studies without increasing the overall error budget. The Qext, Qscat and g grids are freely available on Zenodo as well as a public TauREx plugin -TauREx-PCQ- that utilize them.

Figures (9)

• Figure 1: PyMieScatt computation time for the extinction coefficient of a Titan tholin spherical particle versus the particle's size. For each particle size, the $\rm{Q_{ext}}$ are computed on the full wavelength grid of the optical constants (see section \ref{['sec:optical_cste']}).The green and red lines represent the run time of a model with PyMieScatt with one and four clouds species respectively. The time displayed are the median times for 10 iterations to avoid outliers. The dashed blue line is the average model time for a cloud-free TauREx forward model with $100$ layers al-refaie_TauREx3FastDynamic_2021.
• Figure 2: Extinction coefficient for a $700$ nm SiO$_2$ amorph particle versus wavelength. The black, red and green lines respectively shows the $\mathrm{Q}_{\text{ext}}$ computed using optical constants at native resolution, at a resolution of $100$ and at a global resolution of 100 but with a local 500 resolution for key features.
• Figure 3: $\rm{Q_{ext}}$ relative error between PyMieScatt and linear interpolation for each radius interval in the Titan tholin grid. For each radius, the $\rm{Q_{ext}}$ is computed for 517 wavelengths from $0.3$ to $50$ microns. The $0.5$, $0.9$ and $0.99$ quantile shown in red, blue and green respectively are computed at each radius from the relative error versus wavelength array.
• Figure 4: Synthetic spectra created for the self-retrievals, inspired by WASP 107 b, HD 189733 b, GJ 436 b and 2MASS 2236 b. To generate the errorbars for WASP 107 b we use the PandExo framework batalha_PandExoCommunityToolTransiting_2017. The errorbars for 2MASS 2236 b come from real JWST observations (PID:1188). For ARIEL we make use of ARIELRAD to simulate tier 3 noise, stacking 2 transits for HD 189733 b and 10 for GJ 436 b. In each panel, the colored lines represent the best fit models using TauREx-PyMieScatt and our grid based methods.
• Figure 5: Posterior distributions for WASP 107 b inspired self-retrieval, the PyMieScatt and TauREx-PCQ are shown in blue and red respectively. The median value of each retrieved parameter is shown on top of the corresponding histogram. The truth values are displayed in black.