Skip to main content
Back to timeline
arXivSource publication:

JWST exoplanet retrievals get a transformer posterior estimator: train once, then seconds per spectrum

Synopsis

JWST has pushed exoplanet atmospheric spectra to unprecedented precision, but running a nested-sampling Bayesian retrieval on every planet makes population-level analysis computationally hard. NEXONN, built inside the NEXOTRANS framework, uses a transformer-conditioned normalizing flow to learn the posterior distribution of atmospheric parameters directly, with PandExo-simulated wavelength-dependent JWST noise injected during training. On WASP-39b's MIRI/LRS spectrum and its combined NIRISS, NIRSpec and MIRI spectrum, its central C/O and metallicity values sit close to PyMultiNest's, while per-parameter posterior overlap varies and reference pressure differs noticeably.

Source-provided article image: NEXONN: Advancing the NEXOTRANS Exoplanet Atmospheric Retrieval Framework with Transformer-Based Neural Posterior Estimation for JWST Transmission Spectroscopy of WASP-39b
arXiv

Interpretation

Exoplanet atmospheric retrieval shifts from resampling for every spectrum to training once and then producing parameter posteriors directly for new spectra. The earlier machine-learning implementation in NEXOTRANS stacked random forests, gradient boosting and k-nearest neighbors into a regressor for deterministic parameter estimates, without uncertainty quantification. On a simulated MIRI/LRS spectrum with known true values, the eight parameters show high mean posterior overlap, with the cloud and haze parameters overlapping most strongly, while temperature and some abundance parameters still show visible differences.

The model encodes a spectrum and a set of atmospheric parameters as separate token sequences, then generates posterior samples for parameters one at a time conditioned on the spectrum. The mechanism is a transformer-conditioned autoregressive normalizing flow: wavelength, transit depth and uncertainty enter as unmasked spectral tokens that attend to each other, parameter tokens use a causal mask, and cross-attention lets them read the spectral representation. Each retrieval experiment used an independently generated training set of two million NEXOTRANS synthetic spectra, with six flow layers, four attention heads, batch size 128 and 100 training epochs.

Amortized inference moves cost from every retrieval to a one-time training phase, at the price of reliability only within the parameter and observing configurations the model was trained on. Conventional Bayesian retrieval resamples for each spectrum, whereas a trained model needs a single forward inference pass per spectrum. The 15-parameter full-JWST model trained in roughly hours on an NVIDIA L4 GPU and then inferred a single spectrum in seconds; the corresponding PyMultiNest retrieval took roughly hours on 60 CPU cores, about 2915 CPU-hours.

Perspective

The demonstrated scope is a simulated MIRI/LRS spectrum under the NEXOTRANS forward model, the observed WASP-39b MIRI/LRS spectrum, and a 15-parameter retrieval of the combined NIRISS/SOSS, NIRSpec/PRISM and MIRI/LRS spectrum; coverage tests use 500 simulated test spectra and are averaged over all retrieved parameters.

Posterior agreement varies by parameter: reference pressure, metallicity and several P–T parameters have close central values but noticeably different distributions; real observations may fall outside the training distribution, and training-set coverage, parameter degeneracies and noise-model approximations may all contribute, though these remain possible explanations rather than established causes.