Near-instantaneous Atmospheric Retrievals and Model Comparison with <tt>FASTER</tt>

Anna Lueber, Konstantin Karchev, Chloe Fisher, Matthew David Heim, Roberto Trotta, Kevin HengThe Astrophysical Journal Letters 2025 · DOI

TL;DR

This paper introduces FASTER, a novel neural network-based framework that dramatically accelerates exoplanet atmospheric retrieval and model comparison. By leveraging simulation-based inference and an "amortized" learning approach, FASTER can analyze thousands of exoplanet spectra almost instantaneously, a task that would be computationally prohibitive with traditional methods. This breakthrough promises to unlock new insights into exoplanetary atmospheres, especially with the wealth of data expected from JWST.

What Problem Does This Paper Address?

The era of the James Webb Space Telescope (JWST) is yielding an unprecedented volume of exoplanet atmospheric spectra. Analyzing these spectra to understand the physical and chemical properties of exoplanet atmospheres (a process called "retrieval") is crucial. However, current methods, like Bayesian inference using nested sampling, are computationally intensive. Performing retrievals on thousands of spectra, each requiring comparison against numerous complex atmospheric models (varying in chemistry, temperature profiles, and cloud properties), quickly becomes intractable. This bottleneck hinders our ability to conduct large-scale studies of exoplanet populations and to fully exploit the potential of JWST data.

Methods

The core of FASTER is Simulation-Based Inference (SBI), a machine learning technique that learns a probabilistic model from simulated data rather than directly evaluating a complex likelihood function. Specifically, FASTER employs:

  • Neural Ratio Estimation (NRE): For parameter inference, FASTER trains neural networks to approximate the ratio of the joint probability distribution of parameters and data to the marginal probability distribution of the data. This allows for near-instantaneous estimation of marginal posterior distributions for individual parameters.
  • Amortized Inference: The neural networks are trained on a large suite of simulations covering a wide range of possible parameter values and atmospheric models. Once trained, these networks can perform inference on new data (real or simulated) in milliseconds, without needing to be retrained. This "amortization" means the significant upfront computational cost of training is spread across many subsequent analyses.
  • Amortized Model Comparison: FASTER also trains a separate network to directly estimate the posterior probabilities of different atmospheric models. This bypasses the need to compute Bayesian evidences for each model individually, a notoriously difficult and computationally expensive task with traditional methods.

Datasets

The paper utilizes simulated exoplanet transmission spectra generated by their forward model. These simulations cover various atmospheric compositions, temperature-pressure profiles, and cloud models. For evaluation, they use:

  • Mock Spectra: Simulated noisy spectra generated from known parameter values and model configurations. These serve as a controlled environment to validate FASTER against traditional nested sampling results.
  • Real JWST Data: The NIRSpec PRISM spectrum of the hot Jupiter WASP-39b, previously analyzed by Lueber et al. (2024) using nested sampling. This allows for a direct comparison of FASTER's performance on real-world observational data.

Key Insights and Ideas

  • Speed is Paramount: The sheer volume of upcoming exoplanet data necessitates methods that can analyze spectra orders of magnitude faster than current techniques. FASTER delivers this speedup through its amortized neural network approach.
  • Model Comparison Revolution: Traditional Bayesian model comparison is a major computational bottleneck. FASTER's ability to directly estimate model posterior probabilities is a significant advancement, enabling efficient comparison of numerous complex models.
  • Amortization is Key: The "amortized" nature of FASTER is its superpower. Once trained, the network can perform analyses on new data almost instantaneously, making it ideal for large-scale surveys and population studies.
  • Physical Intuition Verified: FASTER's results, particularly in model comparison, align with physical expectations. For instance, it correctly identifies cloudy models when cloud signatures are present and favors simpler models (Occam's razor) when data are insufficient to distinguish between them.

Key Results

  • Near-Instantaneous Inference: FASTER performs parameter inference in milliseconds after training, compared to hours for traditional nested sampling per model.
  • Accurate Posterior Distributions: Marginal posterior distributions obtained with FASTER closely match those from nested sampling on both mock and real data.
  • Reliable Model Comparison: FASTER's model posterior probabilities are in excellent agreement with those derived from nested sampling, correctly identifying the preferred model for WASP-39b.
  • Computational Efficiency: Training FASTER for 12 models and 1.2 million spectra took approximately 9 hours of GPU time. Subsequent analysis of thousands of spectra is nearly instantaneous.
  • Insights into Model Dependence: FASTER's amortized nature allows for unprecedented analysis of how model selection results depend on underlying parameters, a task computationally infeasible with traditional methods.

Limitations and Future Work

  • Model Misspecification: The authors acknowledge that if the real data are not well-represented by the simulator's assumptions (model misspecification), discrepancies between FASTER and traditional methods may arise. This is an active area of research in SBI.
  • Systematic Uncertainties: The performance of FASTER can be influenced by choices in network architecture, training parameters, and the number of training examples. Quantifying these "systematic" uncertainties is challenging.
  • Future Directions: The authors suggest extending FASTER to handle two-dimensional joint marginal posteriors, calibrating posteriors for exact frequentist coverage, and applying it to large samples of exoplanet spectra for population studies and Bayesian model averaging. They also envision incorporating uncertainties in cross-section data, which is currently computationally prohibitive.

Summary updated

Abstract

Abstract In the era of the James Webb Space Telescope (JWST), the dramatic improvement in the spectra of exoplanetary atmospheres demands a corresponding leap forward in our ability to analyze them: atmospheric retrievals need to be performed on thousands of spectra, applying to each large ensembles of models (that explore atmospheric chemistry, thermal profiles, and cloud models) to identify the best one(s). In this limit, traditional Bayesian inference methods such as nested sampling become prohibitively expensive. We introduce Fast Amortized Simulation-based Transiting Exoplanet Retrieval ( FASTER ), a neural-network-based method for performing atmospheric retrieval and Bayesian model comparison at a fraction of the computational cost of classical techniques. We demonstrate that the marginal posterior distributions of all parameters within a model and the posterior probabilities of the models we consider match those computed using nested sampling both on mock spectra and for the real NIRSpec PRISM spectrum of WASP-39b. The true power of the FASTER framework comes from its amortized nature, which allows the trained networks to perform practically instantaneous Bayesian inference and model comparison over ensembles of spectra—real or simulated—at minimal additional computational cost. This offers valuable insight into the expected results of model comparison (e.g., distinguishing cloudy from cloud-free and isothermal from nonisothermal models), as well as their dependence on the underlying parameters, which is computationally unfeasible with nested sampling. This approach will constitute as large a leap in spectral analysis as the original retrieval methods based on Markov Chain Monte Carlo have proven to be.
BibTeX
@article{lueber2025near,
  author = {Anna Lueber and Konstantin Karchev and Chloe Fisher and Matthew David Heim and Roberto Trotta and Kevin Heng},
  title = {{Near-instantaneous Atmospheric Retrievals and Model Comparison with <tt>FASTER</tt>}},
  journal = {The Astrophysical Journal Letters},
  year = {2025},
  doi = {10.3847/2041-8213/adc7aa},
}
Figure