Exoplanet spectra · reproducible benchmark

Can a model recover what we deliberately injected?

CosmicML-Biodetect asks a narrow inverse-problem question before making any observational claim: can regularized linear regression recover five declared molecular-template amplitudes, with leakage-safe family splits and calibrated uncertainty?

Read the resultsReproduce the analysis
Synthetic molecular-amplitude recovery, interval coverage and noise-sensitivity benchmark plots

Prespecified synthetic benchmark

Held-out performance with controls.

0.01375Macro mean absolute error in injected amplitude units.
0.20846Training-mean baseline error; the model must materially beat it.
94.07%Empirical coverage of nominal 90% split-conformal intervals.

Scope boundary: these values measure recovery from the same analytic template family used to create the data. They are not evidence of biosignature detection, life, or observed-spectrum validity.

Published WASP-39 b JWST transmission spectrum with full and carbon-dioxide-removed model residuals

Published spectrum audit

One real-data check, kept separate.

The project also compares the published WASP-39 b JWST/NIRSpec PRISM spectrum with the publication-supplied full and remove-CO₂ ScCHIMERA curves from Zenodo DOI 10.5281/zenodo.6959427.

Across 93 common wavelength bins, the offset-fitted curves give χ² = 179.19 and 953.59 respectively. The descriptive difference is 774.40; 680.87 comes from 4.0–4.6 μm. This is a reproduction audit, not a new retrieval or detection significance.

Experimental design

Why the benchmark is auditable.

Family-level splits

360 atmosphere families are divided into disjoint train, tune, calibration and test partitions before final evaluation.

Negative control

Permuted labels produce 0.22352 macro MAE, exposing performance that could arise without the intended input–target relation.

Machine-readable outputs

Metrics, per-species results, noise sensitivity, block deletions and provenance hashes are committed with the code.