Skip to content

postprocessing.md

Latest commit

 

History

History
154 lines (126 loc) · 5.59 KB

postprocessing.md

File metadata and controls

154 lines (126 loc) · 5.59 KB
Raw

Post-processing and plotting

ROBERT separates inference and forward-model products from visualization. Run these commands from the simulation directory outside the ROBERT checkout. All relative outputs/ paths below refer to that simulation directory. Configured forward runs write forward_model.npz. Retrieval phases write result.json and result_arrays.npz. They also save the best-fit spectrum and observation for each supported dataset in best_fit_prediction.json and best_fit_prediction.npz. These retain masks, bin edges, units, parameter values, and provenance. Prepared HRS cubes and sampler-only device problems record why a portable spectrum is unavailable; they are not flattened.

Best-fit and parameter plots can use saved products without opacity files or a forward model. Posterior spectral intervals and atmospheric plots still require forward evaluations. To plot only the saved products:

from robert_exoplanets.postprocessing import postprocess_saved_best_fit_output

postprocess_saved_best_fit_output(
    "outputs/multinest", plot_dir="outputs/plots/saved_best_fit"
)

YAML settings

plotting:
  enabled: true
  retrieval: true
  forward: false
  style: default
  image_format: png
  dpi: 180
  max_posterior_samples: 20000
  posterior_predictive_samples: 100
  posterior_predictive_seed: 0
  corner_max_parameters: 20
  dataset_colors:
    dataset_a: mediumpurple
    dataset_b: mediumpurple
  parameter_labels:
    metallicity: "[M/H]"
    CtoO: "C/O"
  leave_one_out:
    enabled: false
    max_posterior_draws: 2000
    seed: 0
    pareto_k_threshold: null

max_posterior_samples limits samples rendered in marginal and corner plots; summary statistics still use the complete weighted posterior. posterior_predictive_samples controls reproducibly resampled forward evaluations for spectral, temperature, VMR, and cloud intervals.

style: default and style: robert select the ROBERT science style used for the WASP-178 b and WASP-15 b figures. This style uses the purple ROBERT palette, black observations, filled posterior contours, and clean axes. A Matplotlib style name or .mplstyle path remains a supported override.

Retrieval products

Run:

python postprocess_retrieval.py --config configuration.yaml

ROBERT discovers completed multinest, optimal_estimation, and hybrid phase directories. Each receives its own folder beneath outputs/plots/ containing:

  • fit_statistics.json;
  • posterior_summary.json;
  • fit_spectrum_residuals.png;
  • posterior_marginals.png or optimal_estimation_parameters.png;
  • parameter_correlation.png;
  • posterior_corner.png when the state dimension permits;
  • temperature_profiles.png when the model exposes an atmosphere builder;
  • vmr_profiles.png for its evaluated composition profiles;
  • cloud_profiles.png for supported cloud models;
  • posterior_predictive_quantiles.npz; and
  • plot_manifest.json.

After retrieval, the plotting step selects 100 posterior draws with replacement, using the sampler weights. It evaluates the same draws for each instrument and atmospheric region. It plots the median and central 68.27% and 95.45% intervals for spectra, temperature, VMR, and cloud profiles. The median is mediumpurple; the 1-sigma band is darker than the 2-sigma band. Dataset colour overrides apply to forward and best-fit-only curves.

The predictive NPZ stores the selected parameter vectors, quantile probabilities, all five curves, coordinates, and units. The shared keys are posterior_draw_vectors, parameter_names, quantile_probabilities, and quantile_labels. Each product has lower_2sigma, lower_1sigma, median, upper_1sigma, and upper_2sigma fields. These are pointwise intervals; the median curves need not describe one joint physical model. Cloud profiles use the actual model's condensate mass fraction, particle radius, or labelled reference-wavelength layer optical depth. Unsupported cloud diagnostics have an explicit unavailable record.

Spectrum panels distinguish observations, best-fit residuals, and posterior intervals. Each instrument keeps its own wavelength grid and bin edges. Parameter names and order come from the serialized result. ROBERT does not add truth or reference markers unless they are explicitly available.

Fit diagnostics include total and per-dataset chi-squared, reduced chi-squared, degrees of freedom, survival probability, RMSE, standardized residuals, recomputed log likelihood, AIC, AICc, and BIC. Nested results retain evidence, timing, weighted quantiles, and effective sample size. Optimal-estimation results use their state and covariance Gaussian approximation.

Process a particular phase or override appearance:

python postprocess_retrieval.py \
  --config configuration.yaml \
  --result-dir outputs/multinest \
  --style paper.mplstyle \
  --format pdf \
  --color dataset_a=mediumpurple \
  --label metallicity='[M/H]'

For a headless machine:

export MPLBACKEND=Agg
export MPLCONFIGDIR="$PWD/scratch/matplotlib"
mkdir -p "$MPLCONFIGDIR"
python postprocess_retrieval.py --config configuration.yaml

The generated Slurm and Glamdring launchers set these automatically.

Forward-model products

Run or regenerate forward diagnostics:

python run_forward.py --config configuration.yaml
python postprocess_forward.py --config configuration.yaml

ROBERT writes beneath outputs/plots/forward/:

  • fit_statistics.json;
  • forward_parameters.json;
  • forward_spectrum_residuals.png; and
  • plot_manifest.json.

Information criteria are included for a consistent schema but are descriptive when the parameters were prescribed rather than inferred.