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

1. Problem Statement

The launch of the James Webb Space Telescope (JWST) and the upcoming ESA ARIEL mission have generated exoplanet atmospheric data with unprecedented spectral resolution and wavelength coverage. While this allows for the detection of complex aerosols (clouds and hazes) ranging from silicates to Titan tholins, analyzing these datasets via atmospheric retrieval remains computationally prohibitive.

• The Bottleneck: Modern retrieval frameworks (e.g., TauREx, petitRADTRANS) model aerosols by solving the Lorenz-Mie theory equations on-the-fly for every sampled particle radius and wavelength.
• Scalability Issues: The computational cost of Mie calculations scales non-linearly with particle size ($a$). In Nested Sampling retrievals, the algorithm's speed is dictated by the slowest process. As the number of aerosol species increases (e.g., modeling multiple condensates simultaneously), the computation time explodes, often becoming the dominant bottleneck.
• Consequence: This limits the ability to perform population-level studies, multi-species retrievals, or analyze high-information-content datasets (such as direct imaging of brown dwarfs) within practical timeframes.

2. Methodology

The authors propose replacing on-the-fly Mie calculations with pre-computed grids of optical properties, utilizing linear interpolation during the retrieval process.

• Pre-computation Strategy:
  • Instead of calculating extinction ($Q_{ext}$), scattering ($Q_{scat}$), and asymmetry ($g$) coefficients for every model iteration, the authors pre-compute these values for a dense grid of particle radii ($a$) and wavelengths ($\lambda$).
  • Species Covered: Seven condensate species relevant to exoplanets: $Mg_2SiO_4$ (amorphous sol-gel), $MgSiO_3$ (amorphous glass), $MgSiO_3$ (amorphous sol-gel), $SiO_2$ (alpha), $SiO_2$ (amorphous), $SiO$, and Titan tholins.
  • Wavelength Range: $0.3$ to $50$ $\mu$m, matching the EXOMOL cross-sections.
  • Radius Range: $1$ nm to $30$ $\mu$m.
• Grid Construction & Optimization:
  • Non-uniform Radius Grids: To ensure numerical accuracy while minimizing memory usage, the authors developed an adaptive grid spacing algorithm. They determined the maximum allowable step size ($\delta a$) such that the relative error between the linear interpolation and the exact Mie calculation remains below a tolerance of $\epsilon = 10^{-4}$.
  • Adaptive Wavelength Resolution: For species with sharp spectral features (e.g., $SiO_2$ amorphous), a standard resolution of 100 points was insufficient. The authors implemented a hybrid grid: a global resolution of 100 combined with local high-resolution grids (500–1000 points) in key feature regions to prevent systematic biases in direct imaging retrievals.
• Implementation:
  • The grids are integrated into a new open-source plugin for the TauREx retrieval framework called TauREx-PCQ.
  • During retrieval, the code performs linear interpolation between grid points to obtain optical properties, drastically reducing the computational load per iteration.

3. Key Contributions

1. Public Data Release: The authors released pre-computed grids for $Q_{ext}$, $Q_{scat}$, and $g$ for seven species on Zenodo (approx. 200–400 MB per species).
2. TauREx-PCQ Plugin: A publicly available, open-source plugin that allows users to utilize these grids within the TauREx framework, removing the need for on-the-fly Mie calculations.
3. Rigorous Validation: The paper provides a comprehensive sensitivity study demonstrating that the linearization of Mie theory via these grids introduces negligible errors in retrieved parameters, even for high-information-content datasets.
4. Scalability Demonstration: The work proves that retrieval speed becomes independent of the number of aerosol species when using pre-computed grids, solving a major scaling issue in current retrieval practices.

4. Results

The authors validated their approach using four synthetic test cases inspired by real exoplanets and brown dwarfs, comparing TauREx-PCQ (grid-based) against TauREx-PyMieScatt (on-the-fly Mie).

• Accuracy:
  • The retrieved posterior distributions for atmospheric parameters (e.g., temperature, metallicity, cloud properties) were statistically indistinguishable between the two methods.
  • The relative error in $Q_{ext}$ was kept below $0.1\%$ for all species, ensuring no bias in transit, emission, or direct imaging spectra.
• Performance Gains (Speed-ups):
  • Single Species: For a single cloud species (e.g., $MgSiO_3$ glass in WASP-107 b), the grid method was 1.4 times faster.
  • Multi-Species Scaling: The advantage grows exponentially with complexity. When retrieving four simultaneous cloud species, the grid method was 17 times faster than the on-the-fly method.
  • Specific Cases:
    • WASP-107 b (JWST Transit): 1.4x speed-up (single species); 17x speed-up (4 species).
    • HD 189733 b (ARIEL Transit): 2.3x speed-up.
    • GJ 436 b (ARIEL Transit): 2.75x speed-up.
    • 2MASS J2236+4751 b (JWST Direct Imaging): 2.1x speed-up.
• Convergence: The grid method allows Nested Sampling algorithms to converge significantly faster, as all MPI processes evaluate models in roughly the same time, avoiding the "straggler" problem caused by variable Mie calculation times.

5. Significance

This work represents a critical step forward for the exoplanet community in the JWST and ARIEL eras:

• Enabling Population Studies: By removing the computational bottleneck of aerosol modeling, it becomes feasible to analyze hundreds or thousands of exoplanet atmospheres uniformly, a requirement for the ARIEL mission.
• Complex Physics: It enables the inclusion of multiple condensate species and complex cloud physics in retrievals without prohibitive time costs, allowing for more realistic atmospheric models.
• High-Information Data: It makes the analysis of high-resolution, high-SNR direct imaging spectra (e.g., for L/T transition dwarfs) computationally tractable, which was previously limited to months of HPC usage per object.
• Open Science: By releasing the grids and the plugin, the authors democratize access to scalable, high-precision aerosol modeling, ensuring that the full scientific potential of current and future telescope data can be realized.