skunkworks-ra/vroomsbi

VROOM-SBI Trained Models

Trained neural posterior estimators for Rotation Measure (RM) synthesis.

Repository: github.com/arpan-52/vroom-sbi

Model Information

• Model Types: burn, external, faraday, internal
• Max Components: 5
• Classifier: Yes
• Upload Date: 2026-07-12

Files

File	Description
classifier.safetensors	Model selection classifier
classifier_training.png	Model selection classifier
posterior_burn_slab_n1.safetensors	Posterior model (.safetensors)
posterior_burn_slab_n2.safetensors	Posterior model (.safetensors)
posterior_burn_slab_n3.safetensors	Posterior model (.safetensors)
posterior_burn_slab_n4.safetensors	Posterior model (.safetensors)
posterior_burn_slab_n5.safetensors	Posterior model (.safetensors)
posterior_external_dispersion_n1.safetensors	Posterior model (.safetensors)
posterior_external_dispersion_n2.safetensors	Posterior model (.safetensors)
posterior_external_dispersion_n3.safetensors	Posterior model (.safetensors)
posterior_external_dispersion_n4.safetensors	Posterior model (.safetensors)
posterior_external_dispersion_n5.safetensors	Posterior model (.safetensors)
posterior_faraday_thin_n1.safetensors	Posterior model (.safetensors)
posterior_faraday_thin_n2.safetensors	Posterior model (.safetensors)
posterior_faraday_thin_n3.safetensors	Posterior model (.safetensors)
posterior_faraday_thin_n4.safetensors	Posterior model (.safetensors)
posterior_faraday_thin_n5.safetensors	Posterior model (.safetensors)
posterior_internal_dispersion_n1.safetensors	Posterior model (.safetensors)
posterior_internal_dispersion_n2.safetensors	Posterior model (.safetensors)
posterior_internal_dispersion_n3.safetensors	Posterior model (.safetensors)
posterior_internal_dispersion_n4.safetensors	Posterior model (.safetensors)
posterior_internal_dispersion_n5.safetensors	Posterior model (.safetensors)
spectral_shape_posterior.safetensors
training_burn_slab_n1.png	Training plot
training_burn_slab_n2.png	Training plot
training_burn_slab_n3.png	Training plot
training_burn_slab_n4.png	Training plot
training_burn_slab_n5.png	Training plot
training_external_dispersion_n1.png	Training plot
training_external_dispersion_n2.png	Training plot
training_external_dispersion_n3.png	Training plot
training_external_dispersion_n4.png	Training plot
training_external_dispersion_n5.png	Training plot
training_faraday_thin_n1.png	Training plot
training_faraday_thin_n2.png	Training plot
training_faraday_thin_n3.png	Training plot
training_faraday_thin_n4.png	Training plot
training_faraday_thin_n5.png	Training plot
training_internal_dispersion_n1.png	Training plot
training_internal_dispersion_n2.png	Training plot
training_internal_dispersion_n3.png	Training plot
training_internal_dispersion_n4.png	Training plot
training_internal_dispersion_n5.png	Training plot
training_spectral_shape.png	Training plot
training_summary.txt	Summary file

TARP Calibration

Joint posterior coverage was verified with the TARP (Test of Accuracy with Random Points) diagnostic (2000 cases, 1000 posterior samples each). calibrated means the empirical coverage curve stayed within the null band around the diagonal; under-confident / over-confident / mixed describe the direction of any deviation.

Model Type	N	Verdict	Unsigned Area	Calibrated
burn_slab	1	mixed	0.0414	no
burn_slab	2	mixed	0.0190	no
burn_slab	3	calibrated	0.0032	yes
burn_slab	4	under-confident	0.0141	no
burn_slab	5	under-confident	0.0171	no
external_dispersion	1	calibrated	0.0058	yes
external_dispersion	2	calibrated	0.0048	yes
external_dispersion	3	calibrated	0.0069	yes
external_dispersion	4	calibrated	0.0098	yes
external_dispersion	5	calibrated	0.0051	yes
faraday_thin	1	under-confident	0.0396	no
faraday_thin	2	mixed	0.0372	no
faraday_thin	3	under-confident	0.0392	no
faraday_thin	4	under-confident	0.0464	no
faraday_thin	5	under-confident	0.0146	no
internal_dispersion	1	calibrated	0.0077	yes
internal_dispersion	2	calibrated	0.0123	yes
internal_dispersion	3	calibrated	0.0053	yes
internal_dispersion	4	calibrated	0.0097	yes
internal_dispersion	5	calibrated	0.0100	yes

Usage

from src.inference import InferenceEngine

engine = InferenceEngine(model_dir="path/to/downloaded/models", device="cuda")  # falls back to CPU
engine.load_models()

# qu_obs: np.ndarray, shape (2*n_freq,) = [Q_0..Q_{n-1}, U_0..U_{n-1}]
result, all_results = engine.infer(qu_obs, n_samples=5000)
print(f"Best model: {result.n_components} components")

Citation

If you use these models, please cite:

Pal & Jagannathan, submitted to AJ (The Astronomical Journal).