A Sulfur-Rich Atmosphere for the Young Jupiter Analog AF Lep b Reveals Significant Solid Accretion

AF Lep b is one of the closest analogs to Jupiter in terms of mass (3 − 4 MJup) and semi-major axis (9 AU) amenable to spectroscopic characterization. We present JWST/NIRSpec high-contrast spectroscopy of the planet from 2.85 − 5.3 µm at R ∼ 3000, which provide detections of CO2, H2 S, CH4, 12 CO (and 13 CO), and H2 O, as well as complementary JWST/NIRCam imaging that captures the planet’s continuum flux from 4.0 − 4.7 µm. Combining the JWST observations with spectra from VLTI/GRAVITY and VLT/SPHERE (1.0 − 2.5 µm), we carry out atmospheric retrievals that include the effects of clouds and disequilibrium chemistry while allowing the C, O, and S abundances to vary independently. AF Lep b exhibits metal enrichment across C, O, and S with C/H = 2.9 ± 0.5, O/H = 3.7 ± 0.6, and S/H = 4.7 ± 0.7 × solar (and stellar). The planet’s slightly sub-solar C/O and C/S are consistent with formation near its observed location, and disfavor formation beyond the CO snowline. The sulfur enrichment in AF Lep b implies significant accretion of disk solids during formation, and we estimate the planet contains 56 ± 7 M⊕ of solids. The C, O, and S enrichment levels of AF Lep b are similar to those of Jupiter, and other super-Jupiters like HR 8799 bcde. We also show that the degree of atmospheric metal enrichment of these imaged planets is similar to the bulk metal enrichment of transiting gas giants with masses greater than ∼ 1 MJup, suggesting that the process and efficiency of metal accretion for gas giants may not be strongly dependent on orbital distance or planet mass.

Improvements to AI Systems:

1. Self-Consistent Retrieval with Multi-Species Abundance Coupling

• Improvement: Train a neural network to jointly infer C/H, O/H, S/H, and C/O, C/S ratios while explicitly modeling cloud opacity and disequilibrium chemistry (e.g., vertical mixing) as coupled parameters, rather than treating them independently.

• Capability: The AI can automatically disentangle degeneracies between metallicity, cloud properties, and chemical kinetics, producing robust abundance estimates for exoplanet atmospheres from multi-instrument spectra.

2. Cross-Instrument Spectral Fusion with Wavelength-Dependent Noise Modeling

• Improvement: Develop a transformer-based architecture that ingests heterogeneous data (NIRSpec R 3000, NIRCam photometry, GRAVITY, SPHERE) and learns a shared latent representation of the planet’s emission, while modeling instrument-specific systematics (e.g., telluric absorption, slit losses) as learnable noise kernels.

• Capability: The AI can seamlessly combine low- and high-resolution data across 1–5.3 µm, improving S/N and enabling detection of weak species like 13CO and H2S that would be missed by single-instrument analysis.

3. Formation-Physics-Guided Abundance Predictor

• Improvement: Implement a physics-informed prior (e.g., pebble accretion, snowline locations) into a Bayesian neural network that predicts C/O and C/S from orbital parameters (semi-major axis, stellar metallicity) and observed metal enrichment.

• Capability: The AI can directly test formation scenarios (e.g., in-situ vs. beyond CO snowline) by outputting posterior probabilities for formation pathways, given a planet’s measured abundances—reducing the need for manual disk-evolution modeling.

4. Solid Mass Accretion Estimator

• Improvement: Train a regression model on hydrodynamical simulations of gas giant formation to map observed S/H enrichment to the total mass of accreted solids, with uncertainty propagation.

• Capability: The AI can instantly estimate the solid mass content (e.g., 56 ± 7 M⊕ for AF Lep b) from atmospheric sulfur abundance alone, enabling rapid comparative studies across imaged and transiting planets without running expensive simulations.

5. Universal Metal-Enrichment Scaling Law Discovery

• Improvement: Use a graph neural network to analyze a database of gas giants (imaged, transiting, and solar system) with features like mass, orbital distance, and measured C/O, C/S, and bulk metallicity, to identify hidden scaling relations.

• Capability: The AI can predict the bulk metal enrichment of a gas giant from its mass and orbital distance alone, and flag outliers (like AF Lep b) that deviate from the trend—revealing potential differences in formation efficiency across environments.

6. Disequilibrium Chemistry Emulator

• Improvement: Replace slow photochemical kinetics solvers with a deep emulator (e.g., a normalizing flow) trained on the outputs of the same chemistry models used in the retrieval, conditioned on temperature-pressure profiles and mixing coefficients.

• Capability: The AI can perform real-time retrieval iterations (100× faster than current methods), allowing for full Bayesian exploration of chemical networks and enabling the detection of subtle species like H2S in low-SNR data.

7. Automated Spectral Line Identification and Isotopologue Sensitivity

• Improvement: Fine-tune a masked autoencoder on synthetic spectra to isolate absorption features of CO2, CH4, H2O, and 12CO/13CO, with attention maps highlighting wavelength regions most sensitive to each species and isotopologue ratio.

• Capability: The AI can autonomously flag which spectral windows to prioritize in future observations, and provide uncertainty maps for isotopic ratios—critical for probing formation temperature and disk chemistry.

What the improved AI system can do:

• Given raw multi-instrument spectra of any directly imaged exoplanet, it can output joint posterior distributions for C, O, S abundances, C/O, C/S, cloud properties, and formation pathway probabilities in under an hour.

• It can automatically compare a new planet’s enrichment pattern to a growing database of gas giants, identifying whether it fits the mass-independent metal accretion trend or requires a distinct formation history.

• It can predict the solid mass accreted during formation from sulfur abundance alone, with quantified uncertainty, and suggest whether additional observations (e.g., NH3 or H2O lines) would break remaining degeneracies.

AF Lep b is one of the closest analogs to Jupiter in terms of mass (3-4 M) and semi-major axis (9 AU) amenable to spectroscopic characterization. We present JWST/NIRSpec high-contrast spectroscopy of the planet from 2.85-5.3 mu m at R about3000, which provide detections of CO 2, H 2 S, CH 4, 12 CO (and 13 CO), and H 2 O, as well as complementary JWST/NIRCam imaging that captures the planet's continuum flux from 4.0-4.7 mu m. Combining the JWST observations with spectra from VLTI/GRAVITY and VLT/SPHERE (1.0-2.5 mu m), we carry out atmospheric retrievals that include the effects of clouds and disequilibrium chemistry while allowing the C, O, and S abundances to vary independently. AF Lep b exhibits metal enrichment across C, O, and S with C/H=2.9 plus or minus0.5, O/H=3.7 plus or minus0.6, and S/H=4.7 plus or minus0.7 times solar (and stellar). The planet's slightly sub-solar C/O and C/S are consistent with formation near its observed location, and disfavor formation beyond the CO snowline. The sulfur enrichment in AF Lep b implies significant accretion of disk solids during formation, and we estimate the planet contains 56 plus or minus7 M of solids. The C, O, and S enrichment levels of AF Lep b are similar to those of Jupiter, and other super-Jupiters like HR 8799 bcde. We also show that the degree of atmospheric metal enrichment of these imaged planets is similar to the bulk metal enrichment of transiting gas giants with masses greater than about1 M, suggesting that the process and efficiency of metal accretion for gas giants may not be strongly dependent on orbital distance or planet mass.