ROBERT

ROBERT is an early-stage JWST exoplanet atmospheric retrieval code, initially focused on emission. It provides typed domain objects, modular atmosphere and chemistry components, opacity preparation, cloud-free and cloudy emission, absorption-dominated transmission, optimal estimation, and optional UltraNest and MultiNest adapters. Its physical forward models are ready for controlled validation analyses within the regimes described below. ROBERT is not yet a general-purpose production science model: calibrated pipeline-product ingestion, broader atmospheric parameterizations, independent science-opacity cloud validation, and long-run posterior validation remain planned.

The Python distribution name is robert-exoplanets to avoid colliding with the existing robert package on PyPI.

Python Support

ROBERT supports Python 3.10 through 3.14. The reproducible full environment uses Python 3.12, while CI tests every supported version.

Architecture

The project architecture is governed by RFC-0001: ROBERT Architectural Specification, with companion documents listed in docs/architecture. Future substantial contributions should follow that document suite.

The latest repository-wide capability and risk assessment is the 2026-07-21 detailed audit.

Quick Start

conda env create --file environment.yml
conda activate robert-exoplanets
# Point STScI stsynphot at a reference-data root containing grid/phoenix.
export PYSYN_CDBS=/path/to/grp/redcat/trds
pytest
python examples/plot_blackbody_reference.py
python examples/plot_synthetic_tau_weighting.py
python examples/plot_cloud_scattering_reference.py
python examples/benchmark_cloud_scattering_picaso_virga.py

For an already-created environment, refresh the editable install with:

python -m pip install -e ".[dev]"

Install pip-provided optional runtime integrations without Conda with:

python -m pip install -e ".[complete]"

MultiNest is supplied by the full Conda environment because its compiled library is installed from conda-forge rather than built locally.

For the runnable Jupyter examples, include the notebook environment:

python -m pip install -e ".[dev,notebooks,opacity,retrieval]"

Science runs use a strict, versioned YAML file rather than an edited Python workflow. Validate the supplied WASP-69b NIRCam+MIRI example without loading its external data or opacity:

python run_retrieval.py \
  --config configurations/wasp69b_cloud_free_R1000.yaml \
  --validate-only

See Configuring and running ROBERT for portable paths, directory initialization, opacity preparation, forward modelling, retrievals, 128-rank nested-sampler Slurm submission, and automatic or manual plotting. Scientific inputs and benchmark products are external by policy; see Scientific data and benchmark artifacts.

Completed configured runs can be post-processed without rerunning inference:

python postprocess_retrieval.py --config configuration.yaml
python postprocess_forward.py --config configuration.yaml
python postprocess_wasp69b_sampler_benchmark.py --project-dir /path/to/my_project

See Post-processing and plotting for fit statistics, plot products, YAML automation, and colour/style overrides.

The maintained forward-model benchmark suite uses PICASO and petitRADTRANS as independent gold-standard comparisons. Superseded benchmark artifacts are removed from Git history and will be archived with future benchmark releases on Zenodo.

The plotting example writes local figures under examples/outputs/, which is ignored by git.

What Exists Today

• Strict, versioned YAML configuration for reproducible forward-model and retrieval runs, backed by typed planet, star, atmosphere, observation, and instrument data models.
• STScI PHOENIX stellar-atmosphere spectra selected by effective temperature, log surface gravity, and metallicity, prepared once on each model grid and used by default for emission contrast; a blackbody fallback remains explicit.
• Correlated-k opacity preparation and evaluation for ExoMolOP/exo_k and ROBERT-native archives, plus target-bin empirical correlated-k preparation from real ExoMolOP cross-section HDF files, including provenance tracking.
• Configurable pressure-temperature profiles and equilibrium or prescribed chemistry, evaluated consistently within each likelihood call.
• Cloud-free and cloudy thermal-emission radiative transfer, including disk integration, hydrostatic geometry, Mie cloud optics, Toon two-stream, and SH4 multiple-scattering solvers.
• A typed parameterized transmission foundation with exact spherical shell chords, explicit reference radius and pressure, constant or self-consistent inverse-square gravity, correlated-k/CIA/Rayleigh extinction, and annulus diagnostics, strict YAML retrieval integration, and synthetic injection-recovery coverage. Scattering-return physics remains future work.
• Shared-atmosphere, multi-instrument forward modelling with instrument-aware binning and Gaussian likelihood support for masks, offsets, and jitter.
• Optimal-estimation, UltraNest, MultiNest, and OE-to-nested inference through a common retrieval interface, with MPI-compatible execution for larger runs.
• Versioned run manifests and portable result products containing the inputs, opacity identifiers, code and runtime provenance, settings, seeds, spectra, and inference outputs.
• General retrieval and forward-model post-processing with fit statistics, residuals, posterior/OE diagnostics, sampler comparisons, and optional automatic plotting controlled by YAML.
• Optional ArviZ-backed PSIS leave-one-out diagnostics with pointwise ELPD, Pareto-k reliability checks, aligned multi-dataset metadata, and guarded treatment of retrieved uncertainty nuisance parameters.
• Unit, integration, regression, injection-recovery, and scientific benchmark tests, including independent forward-model comparisons with pRT and PICASO.

What Comes Later

• Calibrated JWST pipeline-product ingestion and multi-instrument covariance.
• Production atmospheric parameterizations and broader sampler support.
• End-to-end PICASO/Virga parity using independent science molecular-opacity databases, plus high-stream validation beyond the matched SH4 closure.
• Long-run posterior diagnostics and science-grade validation suites.